isotree 0.2.2 → 0.3.0

data/vendor/isotree/src/dist.hpp

@@ -0,0 +1,1886 @@
+/*    Isolation forests and variations thereof, with adjustments for incorporation
+*     of categorical variables and missing values.
+*     Writen for C++11 standard and aimed at being used in R and Python.
+*     
+*     This library is based on the following works:
+*     [1] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
+*         "Isolation forest."
+*         2008 Eighth IEEE International Conference on Data Mining. IEEE, 2008.
+*     [2] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
+*         "Isolation-based anomaly detection."
+*         ACM Transactions on Knowledge Discovery from Data (TKDD) 6.1 (2012): 3.
+*     [3] Hariri, Sahand, Matias Carrasco Kind, and Robert J. Brunner.
+*         "Extended Isolation Forest."
+*         arXiv preprint arXiv:1811.02141 (2018).
+*     [4] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
+*         "On detecting clustered anomalies using SCiForest."
+*         Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Berlin, Heidelberg, 2010.
+*     [5] https://sourceforge.net/projects/iforest/
+*     [6] https://math.stackexchange.com/questions/3388518/expected-number-of-paths-required-to-separate-elements-in-a-binary-tree
+*     [7] Quinlan, J. Ross. C4. 5: programs for machine learning. Elsevier, 2014.
+*     [8] Cortes, David.
+*         "Distance approximation using Isolation Forests."
+*         arXiv preprint arXiv:1910.12362 (2019).
+*     [9] Cortes, David.
+*         "Imputing missing values with unsupervised random trees."
+*         arXiv preprint arXiv:1911.06646 (2019).
+*     [10] https://math.stackexchange.com/questions/3333220/expected-average-depth-in-random-binary-tree-constructed-top-to-bottom
+*     [11] Cortes, David.
+*          "Revisiting randomized choices in isolation forests."
+*          arXiv preprint arXiv:2110.13402 (2021).
+*     [12] Guha, Sudipto, et al.
+*          "Robust random cut forest based anomaly detection on streams."
+*          International conference on machine learning. PMLR, 2016.
+*     [13] Cortes, David.
+*          "Isolation forests: looking beyond tree depth."
+*          arXiv preprint arXiv:2111.11639 (2021).
+*     [14] Ting, Kai Ming, Yue Zhu, and Zhi-Hua Zhou.
+*          "Isolation kernel and its effect on SVM"
+*          Proceedings of the 24th ACM SIGKDD
+*          International Conference on Knowledge Discovery & Data Mining. 2018.
+* 
+*     BSD 2-Clause License
+*     Copyright (c) 2019-2022, David Cortes
+*     All rights reserved.
+*     Redistribution and use in source and binary forms, with or without
+*     modification, are permitted provided that the following conditions are met:
+*     * Redistributions of source code must retain the above copyright notice, this
+*       list of conditions and the following disclaimer.
+*     * Redistributions in binary form must reproduce the above copyright notice,
+*       this list of conditions and the following disclaimer in the documentation
+*       and/or other materials provided with the distribution.
+*     THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+*     AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+*     IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+*     DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+*     FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+*     DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+*     SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+*     CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+*     OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+*     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#include "isotree.hpp" 
+
+
+/* Calculate distance or similarity or kernel/proximity between data points
+* 
+* Parameters
+* ==========
+* - numeric_data[nrows * ncols_numeric]
+*       Pointer to numeric data for which to make calculations. If not using 'indexer', must be
+*       ordered by columns like Fortran, not ordered by rows like C (i.e. entries 1..n contain
+*       column 0, n+1..2n column 1, etc.), while if using 'indexer', may be passed in either
+*       row-major or column-major format (with row-major being faster).
+*       If categorical data is passed, must be in the same storage order (row-major / column-major)
+*       as numerical data (whether dense or sparse).
+*       The column order must be the same as in the data that was used to fit the model.
+*       If making calculations between two sets of observations/rows (see documentation for 'rmat'),
+*       the first group is assumed to be the earlier rows here.
+*       Pass NULL if there are no dense numeric columns.
+*       Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
+* - categ_data[nrows * ncols_categ]
+*       Pointer to categorical data for which to make calculations. If not using 'indexer', must be
+*       ordered by columns like Fortran, not ordered by rows like C (i.e. entries 1..n contain
+*       column 0, n+1..2n column 1, etc.), while if using 'indexer', may be passed in either
+*       row-major or column-major format (with row-major being faster).
+*       If numerical data is passed, must be in the same storage order (row-major / column-major)
+*       as categorical data (whether the numerical data is dense or sparse).
+*       Each category should be represented as an integer, and these integers must start at zero and
+*       be in consecutive order - i.e. if category '3' is present, category '2' must have also been
+*       present when the model was fit (note that they are not treated as being ordinal, this is just
+*       an encoding). Missing values should be encoded as negative numbers such as (-1). The encoding
+*       must be the same as was used in the data to which the model was fit.
+*       Pass NULL if there are no categorical columns.
+*       If making calculations between two sets of observations/rows (see documentation for 'rmat'),
+*       the first group is assumed to be the earlier rows here.
+* - Xc[nnz]
+*       Pointer to numeric data in sparse numeric matrix in CSC format (column-compressed),
+*       or optionally in CSR format (row-compressed) if using 'indexer' and passing 'is_col_major=false'
+*       (not recommended as the calculations will be slower if sparse data is passed as CSR).
+*       If categorical data is passed, must be in the same storage order (row-major or CSR / column-major or CSC)
+*       as numerical data (whether dense or sparse).
+*       Pass NULL if there are no sparse numeric columns.
+*       Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
+* - Xc_ind[nnz]
+*       Pointer to row indices to which each non-zero entry in 'Xc' corresponds
+*       (column indices if 'Xc' is in CSR format).
+*       Must be in sorted order, otherwise results will be incorrect.
+*       Pass NULL if there are no sparse numeric columns in CSC or CSR format.
+* - Xc_indptr[ncols_categ + 1]
+*       Pointer to column index pointers that tell at entry [col] where does column 'col'
+*       start and at entry [col + 1] where does column 'col' end
+*       (row index pointers if 'Xc' is passed in CSR format).
+*       Pass NULL if there are no sparse numeric columns in CSC or CSR format.
+*       If making calculations between two sets of observations/rows (see documentation for 'rmat'),
+*       the first group is assumed to be the earlier rows here.
+* - nrows
+*       Number of rows in 'numeric_data', 'Xc', 'categ_data'.
+* - use_long_double
+*       Whether to use 'long double' (extended precision) type for the calculations. This makes them
+*       more accurate (provided that the compiler used has wider long doubles than doubles), but
+*       slower - especially in platforms in which 'long double' is a software-emulated type (e.g.
+*       Power8 platforms).
+* - nthreads
+*       Number of parallel threads to use. Note that, the more threads, the more memory will be
+*       allocated, even if the thread does not end up being used (with one exception being kernel calculations
+*       with respect to reference points in an idexer). Ignored when not building with OpenMP support.
+* - assume_full_distr
+*       Whether to assume that the fitted model represents a full population distribution (will use a
+*       standardizing criterion assuming infinite sample, and the results of the similarity between two points
+*       at prediction time will not depend on the prescence of any third point that is similar to them, but will
+*       differ more compared to the pairwise distances between points from which the model was fit). If passing
+*       'false', will calculate pairwise distances as if the new observations at prediction time were added to
+*       the sample to which each tree was fit, which will make the distances between two points potentially vary
+*       according to other newly introduced points.
+*       This was added for experimentation purposes only and it's not recommended to pass 'false'.
+*       Note that when calculating distances using 'indexer', there
+*       might be slight discrepancies between the numbers produced with or without the indexer due to what
+*       are considered "additional" observations in this calculation.
+*       This is ignored when passing 'as_kernel=true'.
+* - standardize_dist
+*       Whether to standardize the resulting average separation depths between rows according
+*       to the expected average separation depth in a similar way as when predicting outlierness,
+*       in order to obtain a standardized distance. If passing 'false', will output the average
+*       separation depth instead.
+*       If passing 'as_kernel=true', this indicates whether to output a fraction (if 'true') or
+*       the raw number of matching trees (if 'false').
+* - as_kernel
+*       Whether to calculate the "similarities" as isolation kernel or proximity matrix, which counts
+*       the proportion of trees in which two observations end up in the same terminal node. This is
+*       typically much faster than separation-based distance, but is typically not as good quality.
+*       Note that, for kernel calculations, the indexer is only used if it has reference points stored on it.
+* - model_outputs
+*       Pointer to fitted single-variable model object from function 'fit_iforest'. Pass NULL
+*       if the calculations are to be made from an extended model. Can only pass one of
+*       'model_outputs' and 'model_outputs_ext'.
+* - model_outputs_ext
+*       Pointer to fitted extended model object from function 'fit_iforest'. Pass NULL
+*       if the calculations are to be made from a single-variable model. Can only pass one of
+*       'model_outputs' and 'model_outputs_ext'.
+* - tmat[nrows * (nrows - 1) / 2] (out)
+*       Pointer to array where the resulting pairwise distances or average separation depths or kernels will
+*       be written into. As the output is a symmetric matrix, this function will only fill in the
+*       upper-triangular part, in which entry 0 <= i < j < n will be located at position
+*           p(i,j) = (i * (n - (i+1)/2) + j - i - 1).
+*       Can be converted to a dense square matrix through function 'tmat_to_dense'.
+*       The array must already be initialized to zeros.
+*       If calculating distance/separation from a group of points to another group of points,
+*       pass NULL here and use 'rmat' instead.
+* - rmat[nrows1 * nrows2] (out)
+*       Pointer to array where to write the distances or separation depths or kernels between each row in
+*       one set of observations and each row in a different set of observations. If doing these
+*       calculations for all pairs of observations/rows, pass 'tmat' instead.
+*       Will take the first group of observations as the rows in this matrix, and the second
+*       group as the columns. The groups are assumed to be in the same data arrays, with the
+*       first group corresponding to the earlier rows there.
+*       This matrix will be used in row-major order (i.e. entries 1..nrows2 contain the first row from nrows1).
+*       Must be already initialized to zeros.
+*       If passing 'use_indexed_references=true' plus an indexer object with reference points, this
+*       array should have dimension [nrows, n_references].
+*       Ignored when 'tmat' is passed.
+* - n_from
+*       When calculating distances between two groups of points, this indicates the number of
+*       observations/rows belonging to the first group (the rows in 'rmat'), which will be
+*       assumed to be the first 'n_from' rows.
+*       Ignored when 'tmat' is passed or when 'use_indexed_references=true' plus an indexer with
+*       references are passed.
+* - use_indexed_references
+*       Whether to calculate distances with respect to reference points stored in the indexer
+*       object, if it has any. This is only supported with 'assume_full_distr=true' or with 'as_kernel=true'.
+*       If passing 'use_indexed_references=true', then 'tmat' must be NULL, and 'rmat' must
+*       be of dimension [nrows, n_references].
+* - indexer
+*       Pointer to associated tree indexer for the model being used, if it was constructed,
+*       which can be used to speed up distance calculations, assuming that it was built with
+*       option 'with_distances=true'. If it does not contain node distances, it will not be used.
+*       Pass NULL if the indexer has not been constructed or was constructed with 'with_distances=false'.
+*       If it contains reference points and passing 'use_indexed_references=true', distances will be
+*       calculated between between the input data passed here and the reference points stored in this object.
+*       If passing 'as_kernel=true', the indexer can only be used for calculating kernels with respect to
+*       reference points in the indexer, otherwise it will not be used (which also means that the data must be
+*       passed in column-major order for all kernel calculations that are not with respect to reference points
+*       from an indexer).
+* - is_col_major
+*       Whether the data comes in column-major order. If using 'indexer', predictions are also possible
+*       (and are even faster for the case of dense-only data) if passing the data in row-major format.
+*       Without 'indexer' (and with 'as_kernel=true' but without reference points in the idnexer), data
+*       may only be passed in column-major format.
+*       If there is sparse numeric data, it is highly suggested to pass it in CSC/column-major format.
+* - ld_numeric
+*       If passing 'is_col_major=false', this indicates the leading dimension of the array 'numeric_data'.
+*       Typically, this corresponds to the number of columns, but may be larger (the array will
+*       be accessed assuming that row 'n' starts at 'numeric_data + n*ld_numeric'). If passing
+*       'numeric_data' in column-major order, this is ignored and will be assumed that the
+*       leading dimension corresponds to the number of rows. This is ignored when passing numeric
+*       data in sparse format.
+*       Note that data in row-major order is only accepted when using 'indexer'.
+* - ld_categ
+*       If passing 'is_col_major=false', this indicates the leading dimension of the array 'categ_data'.
+*       Typically, this corresponds to the number of columns, but may be larger (the array will
+*       be accessed assuming that row 'n' starts at 'categ_data + n*ld_categ'). If passing
+*       'categ_data' in column-major order, this is ignored and will be assumed that the
+*       leading dimension corresponds to the number of rows.
+*       Note that data in row-major order is only accepted when using 'indexer'.
+*/
+template <class real_t, class sparse_ix>
+void calc_similarity(real_t numeric_data[], int categ_data[],
+                     real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
+                     size_t nrows, bool use_long_double, int nthreads,
+                     bool assume_full_distr, bool standardize_dist, bool as_kernel,
+                     IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+                     double tmat[], double rmat[], size_t n_from, bool use_indexed_references,
+                     TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ)
+{
+    if (use_long_double && !has_long_double()) {
+        use_long_double = false;
+        fprintf(stderr, "Passed 'use_long_double=true', but library was compiled without long double support.\n");
+    }
+    #ifndef NO_LONG_DOUBLE
+    if (likely(!use_long_double))
+    #endif
+        calc_similarity_internal<real_t, sparse_ix, double>(
+            numeric_data, categ_data,
+            Xc, Xc_ind, Xc_indptr,
+            nrows, nthreads,
+            assume_full_distr, standardize_dist, as_kernel,
+            model_outputs, model_outputs_ext,
+            tmat, rmat, n_from, use_indexed_references,
+            indexer, is_col_major, ld_numeric, ld_categ
+        );
+    #ifndef NO_LONG_DOUBLE
+    else
+        calc_similarity_internal<real_t, sparse_ix, long double>(
+            numeric_data, categ_data,
+            Xc, Xc_ind, Xc_indptr,
+            nrows, nthreads,
+            assume_full_distr, standardize_dist, as_kernel,
+            model_outputs, model_outputs_ext,
+            tmat, rmat, n_from, use_indexed_references,
+            indexer, is_col_major, ld_numeric, ld_categ
+        );
+    #endif
+}
+
+template <class real_t, class sparse_ix, class ldouble_safe>
+void calc_similarity_internal(
+                     real_t numeric_data[], int categ_data[],
+                     real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
+                     size_t nrows, int nthreads,
+                     bool assume_full_distr, bool standardize_dist, bool as_kernel,
+                     IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+                     double tmat[], double rmat[], size_t n_from, bool use_indexed_references,
+                     TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ)
+{
+    if (nrows < 2 && (!use_indexed_references || indexer == NULL || indexer->indices.empty() || indexer->indices.front().reference_points.empty()))
+        throw std::runtime_error("Cannot calculate distances from less than 2 rows.\n");
+    if (as_kernel && (tmat != NULL || !use_indexed_references || (indexer != NULL && !indexer->indices.empty() && indexer->indices.front().reference_points.empty())))
+        indexer = NULL;
+
+    if (indexer != NULL && model_outputs != NULL)
+    {
+        if (model_outputs->missing_action == Divide) {
+            indexer = NULL;
+            if (use_indexed_references) throw std::runtime_error("Invalid indexer - cannot use references from it.\n");
+        }
+        if (model_outputs->new_cat_action == Weighted && model_outputs->cat_split_type == SubSet && categ_data != NULL) {
+            indexer = NULL;
+            if (use_indexed_references) throw std::runtime_error("Invalid indexer - cannot use references from it.\n");
+        }
+    }
+    if (
+        !as_kernel &&
+        indexer != NULL &&
+        (indexer->indices.empty() || indexer->indices.front().node_distances.empty())
+    ) {
+        if (use_indexed_references && !indexer->indices.empty() && !indexer->indices.front().reference_points.empty())
+            throw std::runtime_error("Indexer was built without distances. Cannot use references from it.\n");
+        else {
+            indexer = NULL;
+            fprintf(stderr, "Indexer has no pre-computed distances, will not be used for distance calculations.\n");
+        }
+    }
+    if (
+        !is_col_major &&
+        indexer == NULL &&
+        (
+            Xc_indptr != NULL
+                ||
+            (nrows != 1 &&
+             ((numeric_data != NULL && ld_numeric > 1) || (categ_data != NULL && ld_categ > 1)))
+        )
+    )
+        throw std::runtime_error("Cannot calculate distances with row-major data without indexer.\n");
+    if (indexer != NULL)
+    {
+        if (use_indexed_references && tmat == NULL && !indexer->indices.empty() && !indexer->indices.front().reference_points.empty())
+        {
+            if (unlikely(!assume_full_distr))
+                throw std::runtime_error("Cannot calculate distances to reference points in indexer with 'assume_full_distr=false'.\n");
+            
+            if (!as_kernel)
+            {
+                calc_similarity_from_indexer_with_references(
+                    numeric_data, categ_data,
+                    Xc, Xc_ind, Xc_indptr,
+                    nrows, nthreads, standardize_dist,
+                    model_outputs, model_outputs_ext,
+                    rmat,
+                    indexer, is_col_major, ld_numeric, ld_categ
+                );
+            }
+            
+            else
+            {
+                kernel_to_references(*indexer,
+                                     model_outputs, model_outputs_ext,
+                                     numeric_data, categ_data,
+                                     Xc, Xc_ind, Xc_indptr,
+                                     is_col_major, ld_numeric, ld_categ,
+                                     nrows, nthreads,
+                                     rmat,
+                                     standardize_dist);
+            }
+        }
+
+        else
+        {
+            if (as_kernel) goto skip_indexer_if_kernel;
+            calc_similarity_from_indexer(
+                numeric_data, categ_data,
+                Xc, Xc_ind, Xc_indptr,
+                nrows, nthreads, assume_full_distr, standardize_dist,
+                model_outputs, model_outputs_ext,
+                tmat, rmat, n_from,
+                indexer, is_col_major, ld_numeric, ld_categ
+            );
+        }
+
+        return;
+    }
+    skip_indexer_if_kernel:
+
+    PredictionData<real_t, sparse_ix>
+                   prediction_data = {numeric_data, categ_data, nrows,
+                                      false, 0, 0,
+                                      Xc, Xc_ind, Xc_indptr,
+                                      NULL, NULL, NULL};
+
+    size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
+
+    if (tmat != NULL) n_from = 0;
+
+    if (n_from == 0) {
+        #if SIZE_MAX == UINT32_MAX
+        size_t lim_rows = (size_t)UINT16_MAX - (size_t)1;
+        #elif SIZE_MAX == UINT64_MAX
+        size_t lim_rows = (size_t)UINT32_MAX - (size_t)1;
+        #else
+        size_t lim_rows = (size_t)std::ceil(std::sqrt((ldouble_safe)SIZE_MAX));
+        #endif
+        if (nrows > lim_rows)
+            throw std::runtime_error("Number of rows implies too large distance matrix (integer overflow).");
+    }
+
+    if ((size_t)nthreads > ntrees)
+        nthreads = (int)ntrees;
+    #ifdef _OPENMP
+    std::vector<WorkerForSimilarity> worker_memory(nthreads);
+    #else
+    std::vector<WorkerForSimilarity> worker_memory(1);
+    nthreads = 1;
+    #endif
+
+    /* Global variable that determines if the procedure receives a stop signal */
+    SignalSwitcher ss = SignalSwitcher();
+    check_interrupt_switch(ss);
+    #if defined(DONT_THROW_ON_INTERRUPT)
+    if (interrupt_switch) return;
+    #endif
+    /* For handling exceptions */
+    bool threw_exception = false;
+    std::exception_ptr ex = NULL;
+
+    if (
+        tmat == NULL &&
+        use_indexed_references &&
+        indexer != NULL &&
+        !indexer->indices.empty() &&
+        !indexer->indices.front().reference_points.empty() &&
+        (as_kernel || !indexer->indices.front().node_distances.empty())
+    ) {
+        n_from = indexer->indices.front().reference_points.size();
+    }
+
+    if (model_outputs != NULL)
+    {
+        #pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
+                shared(ntrees, worker_memory, prediction_data, model_outputs, ex, threw_exception, n_from)
+        for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
+        {
+            if (threw_exception || interrupt_switch) continue;
+            try
+            {
+                initialize_worker_for_sim(worker_memory[omp_get_thread_num()], prediction_data,
+                                          model_outputs, NULL, n_from, assume_full_distr);
+                traverse_tree_sim<PredictionData<real_t, sparse_ix>, ldouble_safe>(
+                                  worker_memory[omp_get_thread_num()],
+                                  prediction_data,
+                                  *model_outputs,
+                                  model_outputs->trees[tree],
+                                  (size_t)0,
+                                  as_kernel);
+            }
+
+            catch (...)
+            {
+                #pragma omp critical
+                {
+                    if (!threw_exception)
+                    {
+                        threw_exception = true;
+                        ex = std::current_exception();
+                    }
+                }
+            }
+        }
+    }
+
+    else
+    {
+        #pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
+                shared(ntrees, worker_memory, prediction_data, model_outputs_ext, ex, threw_exception, n_from)
+        for (size_t_for hplane = 0; hplane < (decltype(hplane))ntrees; hplane++)
+        {
+            if (threw_exception || interrupt_switch) continue;
+            try
+            {
+                initialize_worker_for_sim(worker_memory[omp_get_thread_num()], prediction_data,
+                                          NULL, model_outputs_ext, n_from, assume_full_distr);
+                traverse_hplane_sim<PredictionData<real_t, sparse_ix>, ldouble_safe>(
+                                    worker_memory[omp_get_thread_num()],
+                                    prediction_data,
+                                    *model_outputs_ext,
+                                    model_outputs_ext->hplanes[hplane],
+                                    (size_t)0,
+                                    as_kernel);
+            }
+
+            catch (...)
+            {
+                #pragma omp critical
+                {
+                    if (!threw_exception)
+                    {
+                        threw_exception = true;
+                        ex = std::current_exception();
+                    }
+                }
+            }
+        }    
+    }
+
+    check_interrupt_switch(ss);
+    #if defined(DONT_THROW_ON_INTERRUPT)
+    if (interrupt_switch) return;
+    #endif
+
+    if (threw_exception)
+        std::rethrow_exception(ex);
+    
+    /* gather and transform the results */
+    gather_sim_result< PredictionData<real_t, sparse_ix>,
+                       InputData<real_t, sparse_ix>,
+                       WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t> >
+                     (&worker_memory, NULL,
+                      &prediction_data, NULL,
+                      model_outputs, model_outputs_ext,
+                      tmat, rmat, n_from,
+                      ntrees, assume_full_distr,
+                      standardize_dist, as_kernel, nthreads);
+
+    check_interrupt_switch(ss);
+    #if defined(DONT_THROW_ON_INTERRUPT)
+    if (interrupt_switch) return;
+    #endif
+}
+
+template <class PredictionData, class ldouble_safe>
+void traverse_tree_sim(WorkerForSimilarity   &workspace,
+                       PredictionData        &prediction_data,
+                       IsoForest             &model_outputs,
+                       std::vector<IsoTree>  &trees,
+                       size_t                curr_tree,
+                       const bool            as_kernel)
+{
+    if (interrupt_switch)
+        return;
+
+    if (workspace.st == workspace.end)
+        return;
+
+    if (workspace.tmat_sep.empty())
+    {
+        std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
+        if (workspace.ix_arr[workspace.st] >= workspace.n_from)
+            return;
+        if (workspace.ix_arr[workspace.end] < workspace.n_from)
+            return;
+    }
+
+    /* Note: the first separation step will not be added here, as it simply consists of adding +1
+       to every combination regardless. It has to be added at the end in 'gather_sim_result' to
+       obtain the average separation depth. */
+    if (trees[curr_tree].tree_left == 0)
+    {
+        ldouble_safe rem = (ldouble_safe) trees[curr_tree].remainder;
+        if (workspace.weights_arr.empty())
+        {
+            if (!as_kernel)
+            {
+                rem += (ldouble_safe)(workspace.end - workspace.st + 1);
+                if (!workspace.tmat_sep.empty())
+                    increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                          prediction_data.nrows, workspace.tmat_sep.data(),
+                                          workspace.assume_full_distr? 3. : expected_separation_depth(rem));
+                else if (!workspace.rmat.empty())
+                    increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                                    workspace.n_from, prediction_data.nrows, workspace.rmat.data(),
+                                                    workspace.assume_full_distr? 3. : expected_separation_depth(rem));
+            }
+
+            else
+            {
+                if (!workspace.tmat_sep.empty())
+                {
+                    size_t i_, j_;
+                    for (size_t i = workspace.st; i < workspace.end; i++)
+                    {
+                        i_ = workspace.ix_arr[i];
+                        for (size_t j = i + 1; j <= workspace.end; j++)
+                        {
+                            j_ = workspace.ix_arr[j];
+                            workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]++;
+                        }
+                    }
+                }
+
+                else if (!workspace.rmat.empty())
+                {
+                    size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
+                                                   std::lower_bound(workspace.ix_arr.begin() + workspace.st,
+                                                                    workspace.ix_arr.begin() + workspace.end + 1,
+                                                                    workspace.n_from));
+                    double *restrict rmat_this;
+                    for (size_t i = workspace.st; i < workspace.st + n_group; i++)
+                    {
+                        rmat_this = workspace.rmat.data() + workspace.ix_arr[i]*workspace.n_from;
+                        for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
+                        {
+                            rmat_this[workspace.ix_arr[j] - workspace.n_from]++;
+                        }
+                    }
+                }
+            }
+        }
+
+        else
+        {
+            if (!as_kernel)
+            {
+                if (!workspace.assume_full_distr)
+                {
+                    rem += std::accumulate(workspace.ix_arr.begin() + workspace.st,
+                                           workspace.ix_arr.begin() + workspace.end,
+                                           (ldouble_safe) 0.,
+                                           [&workspace](ldouble_safe curr, size_t ix)
+                                           {return curr + (ldouble_safe)workspace.weights_arr[ix];}
+                                          );
+                }
+
+                if (!workspace.tmat_sep.empty())
+                    increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                          prediction_data.nrows, workspace.tmat_sep.data(),
+                                          workspace.weights_arr.data(),
+                                          workspace.assume_full_distr? 3. : expected_separation_depth(rem));
+                else if (!workspace.rmat.empty())
+                    increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                                    workspace.n_from, prediction_data.nrows,
+                                                    workspace.rmat.data(), workspace.weights_arr.data(),
+                                                    workspace.assume_full_distr? 3. : expected_separation_depth(rem));
+            }
+
+            else
+            {
+                if (!workspace.tmat_sep.empty())
+                {
+                    size_t i_, j_;
+                    double w_this;
+                    for (size_t i = workspace.st; i < workspace.end; i++)
+                    {
+                        i_ = workspace.ix_arr[i];
+                        w_this = workspace.weights_arr[i_];
+                        for (size_t j = i + 1; j <= workspace.end; j++)
+                        {
+                            j_ = workspace.ix_arr[j];
+                            workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]
+                                +=
+                            w_this * workspace.weights_arr[j_];
+                        }
+                    }
+                }
+
+                else if (!workspace.rmat.empty())
+                {
+                    size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
+                                                   std::lower_bound(workspace.ix_arr.begin() + workspace.st,
+                                                                    workspace.ix_arr.begin() + workspace.end + 1,
+                                                                    workspace.n_from));
+                    double *restrict rmat_this;
+                    double w_this;
+                    size_t i_, j_;
+                    for (size_t i = workspace.st; i < workspace.st + n_group; i++)
+                    {
+                        i_ = workspace.ix_arr[i];
+                        rmat_this = workspace.rmat.data() + i_*workspace.n_from;
+                        w_this = workspace.weights_arr[i_];
+                        for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
+                        {
+                            j_ = workspace.ix_arr[j];
+                            rmat_this[j_ - workspace.n_from]
+                                +=
+                            w_this * workspace.weights_arr[j_];
+                        }
+                    }
+                }
+            }
+        }
+        return;
+    }
+
+    else if (curr_tree > 0 && !as_kernel)
+    {
+        if (!workspace.tmat_sep.empty())
+        {
+            if (workspace.weights_arr.empty())
+                increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                      prediction_data.nrows, workspace.tmat_sep.data(), -1.);
+            else
+                increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                      prediction_data.nrows, workspace.tmat_sep.data(),
+                                      workspace.weights_arr.data(), -1.);
+        }
+        else if (!workspace.rmat.empty())
+        {
+            if (workspace.weights_arr.empty())
+                increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                                workspace.n_from, prediction_data.nrows, workspace.rmat.data(), -1.);
+            else
+                increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                                workspace.n_from, prediction_data.nrows,
+                                                workspace.rmat.data(), workspace.weights_arr.data(), -1.);
+        }
+    }
+
+
+    /* divide according to tree */
+    if (prediction_data.Xc_indptr != NULL && !workspace.tmat_sep.empty())
+        std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
+    size_t st_NA, end_NA, split_ix;
+    switch (trees[curr_tree].col_type)
+    {
+        case Numeric:
+        {
+            if (prediction_data.Xc_indptr == NULL)
+                divide_subset_split(workspace.ix_arr.data(),
+                                    prediction_data.numeric_data + prediction_data.nrows * trees[curr_tree].col_num,
+                                    workspace.st, workspace.end, trees[curr_tree].num_split,
+                                    model_outputs.missing_action, st_NA, end_NA, split_ix);
+            else
+                divide_subset_split(workspace.ix_arr.data(), workspace.st, workspace.end, trees[curr_tree].col_num,
+                                    prediction_data.Xc, prediction_data.Xc_ind, prediction_data.Xc_indptr,
+                                    trees[curr_tree].num_split, model_outputs.missing_action,
+                                    st_NA, end_NA, split_ix);
+            break;
+        }
+
+        case Categorical:
+        {
+            switch(model_outputs.cat_split_type)
+            {
+                case SingleCateg:
+                {
+                    divide_subset_split(workspace.ix_arr.data(),
+                                        prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
+                                        workspace.st, workspace.end, trees[curr_tree].chosen_cat,
+                                         model_outputs.missing_action, st_NA, end_NA, split_ix);
+                    break;
+                }
+
+                case SubSet:
+                {
+                    if (!trees[curr_tree].cat_split.size())
+                        divide_subset_split(workspace.ix_arr.data(),
+                                            prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
+                                            workspace.st, workspace.end,
+                                            model_outputs.missing_action, model_outputs.new_cat_action,
+                                            trees[curr_tree].pct_tree_left < .5, st_NA, end_NA, split_ix);
+                    else
+                        divide_subset_split(workspace.ix_arr.data(),
+                                            prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
+                                            workspace.st, workspace.end, trees[curr_tree].cat_split.data(),
+                                            (int) trees[curr_tree].cat_split.size(),
+                                            model_outputs.missing_action, model_outputs.new_cat_action,
+                                            (bool)(trees[curr_tree].pct_tree_left < .5), st_NA, end_NA, split_ix);
+                    break;
+                }
+            }
+            break;
+        }
+
+        default:
+        {
+            assert(0);
+            break;
+        }
+    }
+
+
+    /* continue splitting recursively */
+    size_t orig_end = workspace.end;
+    if (model_outputs.new_cat_action == Weighted && model_outputs.cat_split_type == SubSet && prediction_data.categ_data != NULL) {
+        if (model_outputs.missing_action == Fail && trees[curr_tree].col_type == Numeric) {
+            st_NA = split_ix;
+            end_NA = split_ix;
+        }
+        goto missing_action_divide;
+    }
+    switch (model_outputs.missing_action)
+    {
+        case Impute:
+        {
+            split_ix = (trees[curr_tree].pct_tree_left >= .5)? end_NA : st_NA;
+        }
+
+        case Fail:
+        {
+            if (split_ix > workspace.st)
+            {
+                workspace.end = split_ix - 1;
+                traverse_tree_sim<PredictionData, ldouble_safe>(
+                                  workspace,
+                                  prediction_data,
+                                  model_outputs,
+                                  trees,
+                                  trees[curr_tree].tree_left,
+                                  as_kernel);
+            }
+
+
+            if (split_ix <= orig_end)
+            {
+                workspace.st  = split_ix;
+                workspace.end = orig_end;
+                traverse_tree_sim<PredictionData, ldouble_safe>(
+                                  workspace,
+                                  prediction_data,
+                                  model_outputs,
+                                  trees,
+                                  trees[curr_tree].tree_right,
+                                  as_kernel);
+            }
+            break;
+        }
+
+        case Divide: /* new_cat_action = 'Weighted' will also fall here */
+        {
+            /* TODO: this one should also have a parameter 'changed_weoghts' like during fitting */
+            missing_action_divide:
+          /* TODO: maybe here it shouldn't copy the whole ix_arr,
+             but then it'd need to re-generate it from outside too */
+            std::vector<double> weights_arr;
+            std::vector<size_t> ix_arr;
+            if (end_NA > workspace.st)
+            {
+                weights_arr.assign(workspace.weights_arr.begin(),
+                                   workspace.weights_arr.begin() + end_NA);
+                ix_arr.assign(workspace.ix_arr.begin(),
+                              workspace.ix_arr.begin() + end_NA);
+            }
+
+            if (end_NA > workspace.st)
+            {
+                workspace.end = end_NA - 1;
+                for (size_t row = st_NA; row < end_NA; row++)
+                    workspace.weights_arr[workspace.ix_arr[row]] *= trees[curr_tree].pct_tree_left;
+                traverse_tree_sim<PredictionData, ldouble_safe>(
+                                  workspace,
+                                  prediction_data,
+                                  model_outputs,
+                                  trees,
+                                  trees[curr_tree].tree_left,
+                                  as_kernel);
+            }
+
+            if (st_NA <= orig_end)
+            {
+                workspace.st = st_NA;
+                workspace.end = orig_end;
+                if (!weights_arr.empty())
+                {
+                    std::copy(weights_arr.begin(),
+                              weights_arr.end(),
+                              workspace.weights_arr.begin());
+                    std::copy(ix_arr.begin(),
+                              ix_arr.end(),
+                              workspace.ix_arr.begin());
+                    weights_arr.clear();
+                    weights_arr.shrink_to_fit();
+                    ix_arr.clear();
+                    ix_arr.shrink_to_fit();
+                }
+
+                for (size_t row = st_NA; row < end_NA; row++)
+                    workspace.weights_arr[workspace.ix_arr[row]] *= (1. - trees[curr_tree].pct_tree_left);
+                traverse_tree_sim<PredictionData, ldouble_safe>(
+                                  workspace,
+                                  prediction_data,
+                                  model_outputs,
+                                  trees,
+                                  trees[curr_tree].tree_right,
+                                  as_kernel);
+            }
+            break;
+        }
+    }
+}
+
+template <class PredictionData, class ldouble_safe>
+void traverse_hplane_sim(WorkerForSimilarity     &workspace,
+                         PredictionData          &prediction_data,
+                         ExtIsoForest            &model_outputs,
+                         std::vector<IsoHPlane>  &hplanes,
+                         size_t                  curr_tree,
+                         const bool              as_kernel)
+{
+    if (interrupt_switch)
+        return;
+    
+    if (workspace.st == workspace.end)
+        return;
+
+    if (workspace.tmat_sep.empty())
+    {
+        std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
+        if (workspace.ix_arr[workspace.st] >= workspace.n_from)
+            return;
+        if (workspace.ix_arr[workspace.end] < workspace.n_from)
+            return;
+    }
+
+    /* Note: the first separation step will not be added here, as it simply consists of adding +1
+       to every combination regardless. It has to be added at the end in 'gather_sim_result' to
+       obtain the average separation depth. */
+    if (hplanes[curr_tree].hplane_left == 0)
+    {
+        if (!as_kernel)
+        {
+            if (!workspace.tmat_sep.empty())
+                increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                      prediction_data.nrows, workspace.tmat_sep.data(),
+                                      workspace.assume_full_distr? 3. : 
+                                      expected_separation_depth((ldouble_safe) hplanes[curr_tree].remainder
+                                                                  + (ldouble_safe)(workspace.end - workspace.st + 1))
+                                      );
+            else if (!workspace.rmat.empty())
+                increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.n_from,
+                                                prediction_data.nrows, workspace.rmat.data(),
+                                                workspace.assume_full_distr? 3. : 
+                                                expected_separation_depth((ldouble_safe) hplanes[curr_tree].remainder
+                                                                            + (ldouble_safe)(workspace.end - workspace.st + 1))
+                                                );
+        }
+
+        else
+        {
+            if (!workspace.tmat_sep.empty())
+            {
+                size_t i_, j_;
+                for (size_t i = workspace.st; i < workspace.end; i++)
+                {
+                    i_ = workspace.ix_arr[i];
+                    for (size_t j = i + 1; j <= workspace.end; j++)
+                    {
+                        j_ = workspace.ix_arr[j];
+                        workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]++;
+                    }
+                }
+            }
+
+            else if (!workspace.rmat.empty())
+            {
+                size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
+                                               std::lower_bound(workspace.ix_arr.begin() + workspace.st,
+                                                                workspace.ix_arr.begin() + workspace.end + 1,
+                                                                workspace.n_from));
+                double *restrict rmat_this;
+                for (size_t i = workspace.st; i < workspace.st + n_group; i++)
+                {
+                    rmat_this = workspace.rmat.data() + workspace.ix_arr[i]*workspace.n_from;
+                    for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
+                    {
+                        rmat_this[workspace.ix_arr[j] - workspace.n_from]++;
+                    }
+                }
+            }
+        }
+        return;
+    }
+
+    else if (curr_tree > 0 && !as_kernel)
+    {
+        if (!workspace.tmat_sep.empty())
+            increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                  prediction_data.nrows, workspace.tmat_sep.data(), -1.);
+        else if (!workspace.rmat.empty())
+            increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.n_from,
+                                            prediction_data.nrows, workspace.rmat.data(), -1.);
+    }
+
+    if (prediction_data.Xc_indptr != NULL && workspace.tmat_sep.size())
+        std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
+
+    /* reconstruct linear combination */
+    size_t ncols_numeric = 0;
+    size_t ncols_categ   = 0;
+    std::fill(workspace.comb_val.begin(), workspace.comb_val.begin() + (workspace.end - workspace.st + 1), 0);
+    double unused;
+    if (prediction_data.categ_data != NULL || prediction_data.Xc_indptr != NULL)
+    {
+        for (size_t col = 0; col < hplanes[curr_tree].col_num.size(); col++)
+        {
+            switch(hplanes[curr_tree].col_type[col])
+            {
+                case Numeric:
+                {
+                    if (prediction_data.Xc_indptr == NULL)
+                        add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
+                                        prediction_data.numeric_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
+                                        hplanes[curr_tree].coef[ncols_numeric], (double)0, hplanes[curr_tree].mean[ncols_numeric],
+                                        (model_outputs.missing_action == Fail)?  unused : hplanes[curr_tree].fill_val[col],
+                                        model_outputs.missing_action, NULL, NULL, false);
+                    else
+                        add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end,
+                                        hplanes[curr_tree].col_num[col], workspace.comb_val.data(),
+                                        prediction_data.Xc, prediction_data.Xc_ind, prediction_data.Xc_indptr,
+                                        hplanes[curr_tree].coef[ncols_numeric], (double)0, hplanes[curr_tree].mean[ncols_numeric],
+                                        (model_outputs.missing_action == Fail)?  unused : hplanes[curr_tree].fill_val[col],
+                                        model_outputs.missing_action, NULL, NULL, false);
+                    ncols_numeric++;
+                    break;
+                }
+
+                case Categorical:
+                {
+                    switch(model_outputs.cat_split_type)
+                    {
+                        case SingleCateg:
+                        {
+                            add_linear_comb<ldouble_safe>(
+                                            workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
+                                            prediction_data.categ_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
+                                            (int)0, NULL, hplanes[curr_tree].fill_new[ncols_categ],
+                                            hplanes[curr_tree].chosen_cat[ncols_categ],
+                                            (model_outputs.missing_action == Fail)?  unused : hplanes[curr_tree].fill_val[col],
+                                            workspace.comb_val[0], NULL, NULL, model_outputs.new_cat_action,
+                                            model_outputs.missing_action, SingleCateg, false);
+                            break;
+                        }
+
+                        case SubSet:
+                        {
+                            add_linear_comb<ldouble_safe>(
+                                            workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
+                                            prediction_data.categ_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
+                                            (int) hplanes[curr_tree].cat_coef[ncols_categ].size(),
+                                            hplanes[curr_tree].cat_coef[ncols_categ].data(), (double) 0, (int) 0,
+                                            (model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
+                                            hplanes[curr_tree].fill_new[ncols_categ], NULL, NULL,
+                                            model_outputs.new_cat_action, model_outputs.missing_action, SubSet, false);
+                            break;
+                        }
+                    }
+                    ncols_categ++;
+                    break;
+                }
+
+                default:
+                {
+                    assert(0);
+                    break;
+                }
+            } 
+        }
+    }
+
+    
+    else /* faster version for numerical-only */
+    {
+        for (size_t col = 0; col < hplanes[curr_tree].col_num.size(); col++)
+            add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
+                            prediction_data.numeric_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
+                            hplanes[curr_tree].coef[col], (double)0, hplanes[curr_tree].mean[col],
+                            (model_outputs.missing_action == Fail)?  unused : hplanes[curr_tree].fill_val[col],
+                            model_outputs.missing_action, NULL, NULL, false);
+    }
+
+    /* divide data */
+    size_t split_ix = divide_subset_split(workspace.ix_arr.data(), workspace.comb_val.data(),
+                                          workspace.st, workspace.end, hplanes[curr_tree].split_point);
+
+    /* continue splitting recursively */
+    size_t orig_end = workspace.end;
+    if (split_ix > workspace.st)
+    {
+        workspace.end = split_ix - 1;
+        traverse_hplane_sim<PredictionData, ldouble_safe>(
+                            workspace,
+                            prediction_data,
+                            model_outputs,
+                            hplanes,
+                            hplanes[curr_tree].hplane_left,
+                            as_kernel);
+    }
+
+    if (split_ix <= orig_end)
+    {
+        workspace.st  = split_ix;
+        workspace.end = orig_end;
+        traverse_hplane_sim<PredictionData, ldouble_safe>(
+                            workspace,
+                            prediction_data,
+                            model_outputs,
+                            hplanes,
+                            hplanes[curr_tree].hplane_right,
+                            as_kernel);
+    }
+
+}
+
+template <class PredictionData, class InputData, class WorkerMemory>
+void gather_sim_result(std::vector<WorkerForSimilarity> *worker_memory,
+                       std::vector<WorkerMemory> *worker_memory_m,
+                       PredictionData *prediction_data, InputData *input_data,
+                       IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+                       double *restrict tmat, double *restrict rmat, size_t n_from,
+                       size_t ntrees, bool assume_full_distr,
+                       bool standardize_dist, bool as_kernel, int nthreads)
+{
+    if (interrupt_switch)
+        return;
+
+    size_t nrows = (prediction_data != NULL)? prediction_data->nrows : input_data->nrows;
+    size_t ncomb = calc_ncomb(nrows);
+    size_t n_to  = (prediction_data != NULL)? (prediction_data->nrows - n_from) : 0;
+
+    #ifdef _OPENMP
+    if (nthreads > 1)
+    {
+        if (worker_memory != NULL)
+        {
+            for (WorkerForSimilarity &w : *worker_memory)
+            {
+                if (!w.tmat_sep.empty())
+                {
+                    #pragma omp parallel for schedule(static) num_threads(nthreads) shared(ncomb, tmat, w, worker_memory)
+                    for (size_t_for ix = 0; ix < (decltype(ix))ncomb; ix++)
+                        tmat[ix] += w.tmat_sep[ix];
+                }
+                else if (!w.rmat.empty())
+                {
+                    #pragma omp parallel for schedule(static) num_threads(nthreads) shared(rmat, w, worker_memory)
+                    for (size_t_for ix = 0; ix < (decltype(ix))w.rmat.size(); ix++)
+                        rmat[ix] += w.rmat[ix];
+                }
+            }
+        }
+
+        else
+        {
+            for (WorkerMemory &w : *worker_memory_m)
+            {
+                if (!w.tmat_sep.empty())
+                {
+                    #pragma omp parallel for schedule(static) num_threads(nthreads) shared(ncomb, tmat, w, worker_memory_m)
+                    for (size_t_for ix = 0; ix < (decltype(ix))ncomb; ix++)
+                        tmat[ix] += w.tmat_sep[ix];
+                }
+            }
+        }
+    }
+    
+    else
+    #endif
+    {
+        if (worker_memory != NULL)
+        {
+            if (!(*worker_memory)[0].tmat_sep.empty())
+                std::copy((*worker_memory)[0].tmat_sep.begin(), (*worker_memory)[0].tmat_sep.end(), tmat);
+            else
+                std::copy((*worker_memory)[0].rmat.begin(), (*worker_memory)[0].rmat.end(), rmat);
+        }
+        
+        else
+        {
+            std::copy((*worker_memory_m)[0].tmat_sep.begin(), (*worker_memory_m)[0].tmat_sep.end(), tmat);
+        }
+    }
+
+    double ntrees_dbl = (double) ntrees;
+    if (standardize_dist)
+    {
+        if (as_kernel)
+        {
+            if (tmat != NULL)
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    tmat[ix] /= ntrees_dbl;
+            else
+                for (size_t ix = 0; ix < (n_from * n_to); ix++)
+                    rmat[ix] /= ntrees_dbl;
+            return;
+        }
+
+
+        /* Note: the separation distances up this point are missing the first hop, which is always
+           a +1 to every combination. Thus, it needs to be added back for the average separation depth.
+           For the standardized metric, it takes the expected divisor as 2(=3-1) instead of 3, given
+           that every combination will always get a +1 at the beginning. Since what's obtained here
+           is a sum across all trees, adding this +1 means adding the number of trees. */
+        double div_trees = ntrees_dbl;
+        if (assume_full_distr)
+        {
+            div_trees *= 2;
+        }
+
+        else if (input_data != NULL)
+        {
+            div_trees *= (expected_separation_depth(input_data->nrows) - 1);
+        }
+
+        else
+        {
+            div_trees *= ((
+                               (model_outputs != NULL)?
+                                expected_separation_depth_hotstart(model_outputs->exp_avg_sep,
+                                                                    model_outputs->orig_sample_size,
+                                                                    model_outputs->orig_sample_size + prediction_data->nrows)
+                                    :
+                                expected_separation_depth_hotstart(model_outputs_ext->exp_avg_sep,
+                                                                    model_outputs_ext->orig_sample_size,
+                                                                    model_outputs_ext->orig_sample_size + prediction_data->nrows)
+                          ) - 1);
+        }
+
+        
+        if (tmat != NULL)
+            #ifndef _WIN32
+            #pragma omp simd
+            #endif
+            for (size_t ix = 0; ix < ncomb; ix++)
+                tmat[ix] = std::exp2( - tmat[ix] / div_trees);
+        else
+            #ifndef _WIN32
+            #pragma omp simd
+            #endif
+            for (size_t ix = 0; ix < (n_from * n_to); ix++)
+                rmat[ix] = std::exp2( - rmat[ix] / div_trees);
+    }
+    
+    else
+    {
+        if (as_kernel) return;
+
+        if (tmat != NULL)
+            #ifndef _WIN32
+            #pragma omp simd
+            #endif
+            for (size_t ix = 0; ix < ncomb; ix++)
+                tmat[ix] = (tmat[ix] + ntrees) / ntrees_dbl;
+        else
+            #ifndef _WIN32
+            #pragma omp simd
+            #endif
+            for (size_t ix = 0; ix < (n_from * n_to); ix++)
+                rmat[ix] = (rmat[ix] + ntrees) / ntrees_dbl;
+    }
+}
+
+template <class PredictionData>
+void initialize_worker_for_sim(WorkerForSimilarity  &workspace,
+                               PredictionData       &prediction_data,
+                               IsoForest            *model_outputs,
+                               ExtIsoForest         *model_outputs_ext,
+                               size_t                n_from,
+                               bool                  assume_full_distr)
+{
+    workspace.st  = 0;
+    workspace.end = prediction_data.nrows - 1;
+    workspace.n_from = n_from;
+    workspace.assume_full_distr = assume_full_distr; /* doesn't need to have one copy per worker */
+
+    if (workspace.ix_arr.empty())
+    {
+        workspace.ix_arr.resize(prediction_data.nrows);
+        std::iota(workspace.ix_arr.begin(), workspace.ix_arr.end(), (size_t)0);
+        if (!n_from)
+            workspace.tmat_sep.resize(calc_ncomb(prediction_data.nrows), 0);
+        else
+            workspace.rmat.resize((prediction_data.nrows - n_from) * n_from, 0);
+    }
+
+    if (model_outputs != NULL &&
+        (model_outputs->missing_action == Divide ||
+         (model_outputs->new_cat_action == Weighted && model_outputs->cat_split_type == SubSet && prediction_data.categ_data != NULL)))
+    {
+        if (workspace.weights_arr.empty())
+            workspace.weights_arr.resize(prediction_data.nrows, 1.);
+        else
+            std::fill(workspace.weights_arr.begin(), workspace.weights_arr.end(), 1.);
+    }
+
+    if (model_outputs_ext != NULL)
+    {
+        if (workspace.comb_val.empty())
+            workspace.comb_val.resize(prediction_data.nrows, 0);
+        else
+            std::fill(workspace.comb_val.begin(), workspace.comb_val.end(), 0);
+    }
+}
+
+template <class real_t, class sparse_ix>
+void calc_similarity_from_indexer
+(
+    real_t *restrict numeric_data, int *restrict categ_data,
+    real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
+    size_t nrows, int nthreads, bool assume_full_distr, bool standardize_dist,
+    IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+    double *restrict tmat, double *restrict rmat, size_t n_from,
+    TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ
+)
+{
+    SignalSwitcher ss;
+    size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
+    std::vector<sparse_ix> terminal_indices(nrows * ntrees);
+    std::unique_ptr<double[]> ignored(new double[nrows]);
+    predict_iforest(numeric_data, categ_data,
+                    is_col_major, ld_numeric, ld_categ,
+                    is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
+                    is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
+                    nrows, nthreads, false,
+                    model_outputs, model_outputs_ext,
+                    ignored.get(), terminal_indices.data(),
+                    (double*)NULL,
+                    indexer);
+    ignored.reset();
+
+    #ifndef _OPENMP
+    nthreads = 1;
+    #endif
+
+    check_interrupt_switch(ss);
+
+    if (n_from == 0)
+    {
+        size_t ncomb = calc_ncomb(nrows);
+        std::fill_n(tmat, ncomb, 0.);
+
+        std::vector<std::vector<double>> sum_separations(nthreads);
+        if (nthreads != 1) {
+            for (auto &v : sum_separations) v.resize(ncomb);
+        }
+
+        std::vector<std::vector<size_t>> thread_argsorted_nodes(nthreads);
+        for (auto &v : thread_argsorted_nodes) v.resize(nrows);
+
+        std::vector<std::vector<size_t>> thread_sorted_nodes(nthreads);
+        for (auto &v : thread_sorted_nodes) v.reserve(nrows); /* <- could shrink to max number of terminal nodes */
+
+
+        bool threw_exception = false;
+        std::exception_ptr ex = NULL;
+        #pragma omp parallel for schedule(static) num_threads(nthreads) \
+                shared(model_outputs, model_outputs_ext, nthreads, indexer, nrows, ncomb, terminal_indices, \
+                       sum_separations, thread_argsorted_nodes, thread_sorted_nodes, tmat, \
+                       threw_exception, ex)
+        for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
+        {
+            if (interrupt_switch || threw_exception) continue;
+
+            if (unlikely(indexer->indices[tree].n_terminal <= 1))
+            {
+                for (auto &el : sum_separations[omp_get_thread_num()]) el += 1.;
+                continue;
+            }
+
+            double *restrict ptr_this_sep = sum_separations[omp_get_thread_num()].data();
+            if (nthreads == 1) ptr_this_sep = tmat;
+            double *restrict node_dist_this = indexer->indices[tree].node_distances.data();
+            double *restrict node_depths_this = indexer->indices[tree].node_depths.data();
+            size_t n_terminal_this = indexer->indices[tree].n_terminal;
+            size_t ncomb_this = calc_ncomb(n_terminal_this);
+            std::vector<IsoTree> *tree_this = (model_outputs != NULL)? &model_outputs->trees[tree] : nullptr;
+            std::vector<IsoHPlane> *hplane_this = (model_outputs_ext != NULL)? &model_outputs_ext->hplanes[tree] : nullptr;
+            sparse_ix *restrict terminal_indices_this = terminal_indices.data() + nrows * tree;
+            size_t i, j;
+            double add_round;
+
+            if (assume_full_distr)
+            {
+                for (size_t el1 = 0; el1 < nrows-1; el1++)
+                {
+                    i = terminal_indices_this[el1];
+                    for (size_t el2 = el1+1; el2 < nrows; el2++)
+                    {
+                        j = terminal_indices_this[el2];
+                        if (unlikely(i == j))
+                            add_round = node_depths_this[i] + 3.;
+                        else
+                            add_round = node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
+                        ptr_this_sep[ix_comb(el1, el2, nrows, ncomb)] += add_round;
+                    }
+                }
+            }
+
+            else
+            {
+                hashed_set<size_t> nodes_w_repeated;
+                try
+                {
+                    nodes_w_repeated.reserve(n_terminal_this);
+                    for (size_t el1 = 0; el1 < nrows-1; el1++)
+                    {
+                        i = terminal_indices_this[el1];
+                        for (size_t el2 = el1+1; el2 < nrows; el2++)
+                        {
+                            j = terminal_indices_this[el2];
+                            if (unlikely(i == j))
+                                nodes_w_repeated.insert(i);
+                            else
+                                ptr_this_sep[ix_comb(el1, el2, nrows, ncomb)]
+                                    +=
+                                node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
+                        }
+                    }
+                }
+
+                catch (...)
+                {
+                    #pragma omp critical
+                    {
+                        if (!threw_exception)
+                        {
+                            threw_exception = true;
+                            ex = std::current_exception();
+                        }
+                    }
+                }
+
+                if (likely(!nodes_w_repeated.empty()))
+                {
+                    std::vector<size_t> *restrict argsorted_nodes = &thread_argsorted_nodes[omp_get_thread_num()];
+                    std::iota(argsorted_nodes->begin(), argsorted_nodes->end(), (size_t)0);
+                    std::sort(argsorted_nodes->begin(), argsorted_nodes->end(),
+                              [&terminal_indices_this](const size_t a, const size_t b)
+                              {return terminal_indices_this[a] < terminal_indices_this[b];});
+                    std::vector<size_t>::iterator curr_begin = argsorted_nodes->begin();
+                    std::vector<size_t>::iterator new_begin;
+
+                    std::vector<size_t> *restrict sorted_nodes = &thread_sorted_nodes[omp_get_thread_num()];
+                    sorted_nodes->assign(nodes_w_repeated.begin(), nodes_w_repeated.end());
+                    std::sort(sorted_nodes->begin(), sorted_nodes->end());
+                    for (size_t node_ix : *sorted_nodes)
+                    {
+                        curr_begin = std::lower_bound(curr_begin, argsorted_nodes->end(),
+                                                      node_ix,
+                                                      [&terminal_indices_this](const size_t &a, const size_t &b)
+                                                      {return (size_t)terminal_indices_this[a] < b;});
+                        new_begin =  std::upper_bound(curr_begin, argsorted_nodes->end(),
+                                                      node_ix,
+                                                      [&terminal_indices_this](const size_t &a, const size_t &b)
+                                                      {return a < (size_t)terminal_indices_this[b];});
+                        size_t n_this = std::distance(curr_begin, new_begin);
+                        double sep_this
+                            =
+                        n_this
+                            +
+                        ((tree_this != NULL)?
+                         (*tree_this)[node_ix].remainder
+                            :
+                         (*hplane_this)[node_ix].remainder);
+                        double sep_this_ = expected_separation_depth(sep_this) + node_depths_this[node_ix];
+
+                        size_t i, j;
+                        for (size_t el1 = 0; el1 < n_this-1; el1++)
+                        {
+                            i = *(curr_begin + el1);
+                            for (size_t el2 = el1+1; el2 < n_this; el2++)
+                            {
+                                j = *(curr_begin + el2);
+                                ptr_this_sep[ix_comb(i, j, nrows, ncomb)] += sep_this_;
+                            }
+                        }
+
+                        curr_begin = new_begin;
+                    }
+                }
+
+            }
+        }
+
+        check_interrupt_switch(ss);
+
+        if (threw_exception)
+            std::rethrow_exception(ex);
+
+        if (nthreads == 1)
+        {
+            /* Here 'tmat' already contains the sum of separations */
+        }
+
+        else
+        {
+            for (int tid = 0; tid < nthreads; tid++)
+            {
+                double *restrict seps_thread = sum_separations[tid].data();
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    tmat[ix] += seps_thread[ix];
+            }
+        }
+
+        check_interrupt_switch(ss);
+
+        if (standardize_dist)
+        {
+            double divisor;
+            if (assume_full_distr)
+                divisor = (double)(ntrees * 2);
+            else
+                divisor = (double)ntrees * ((model_outputs != NULL)? model_outputs->exp_avg_sep : model_outputs_ext->exp_avg_sep);
+
+            if (assume_full_distr)
+            {
+                double ntrees_dbl = (double)ntrees;
+                #ifndef _WIN32
+                #pragma omp simd
+                #endif
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    tmat[ix] = std::exp2( - (tmat[ix] - ntrees_dbl) / divisor);
+            }
+
+            else
+            {
+                #ifndef _WIN32
+                #pragma omp simd
+                #endif
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    tmat[ix] = std::exp2( - tmat[ix] / divisor);
+            }
+        }
+
+        else
+        {
+            double divisor = (double)ntrees;
+            for (size_t ix = 0; ix < ncomb; ix++)
+                tmat[ix] /= divisor;
+        }
+
+        check_interrupt_switch(ss);
+    }
+
+    /* TODO: merge this with the block above, can simplify lots of things by a couple if-elses */
+    else /* has 'rmat' / 'nfrom>0' */
+    {
+        size_t n_to  = nrows - n_from;
+        size_t ncomb = n_from * n_to;
+        std::fill_n(rmat, ncomb, 0.);
+
+        std::vector<std::vector<double>> sum_separations(nthreads);
+        if (nthreads != 1) {
+            for (auto &v : sum_separations) v.resize(ncomb);
+        }
+
+        std::vector<std::vector<size_t>> thread_argsorted_nodes(nthreads);
+        for (auto &v : thread_argsorted_nodes) v.resize(nrows);
+
+        std::vector<std::vector<size_t>> thread_doubly_argsorted(nthreads);
+        for (auto &v : thread_doubly_argsorted) v.reserve(nrows);
+
+        std::vector<std::vector<size_t>> thread_sorted_nodes(nthreads);
+        for (auto &v : thread_sorted_nodes) v.reserve(nrows); /* <- could shrink to max number of terminal nodes */
+
+        bool threw_exception = false;
+        std::exception_ptr ex = NULL;
+        #pragma omp parallel for schedule(static) num_threads(nthreads) \
+                shared(model_outputs, model_outputs_ext, nthreads, indexer, nrows, ncomb, terminal_indices, \
+                       sum_separations, thread_argsorted_nodes, thread_sorted_nodes, thread_doubly_argsorted, rmat, n_to, n_from, \
+                       threw_exception, ex)
+        for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
+        {
+            if (interrupt_switch || threw_exception) continue;
+
+            if (unlikely(indexer->indices[tree].n_terminal <= 1))
+            {
+                for (auto &el : sum_separations[omp_get_thread_num()]) el += 1.;
+                continue;
+            }
+
+            double *restrict ptr_this_sep = sum_separations[omp_get_thread_num()].data();
+            if (nthreads == 1) ptr_this_sep = rmat;
+            double *restrict node_dist_this = indexer->indices[tree].node_distances.data();
+            double *restrict node_depths_this = indexer->indices[tree].node_depths.data();
+            size_t n_terminal_this = indexer->indices[tree].n_terminal;
+            size_t ncomb_this = calc_ncomb(n_terminal_this);
+            std::vector<IsoTree> *tree_this = (model_outputs != NULL)? &model_outputs->trees[tree] : nullptr;
+            std::vector<IsoHPlane> *hplane_this = (model_outputs_ext != NULL)? &model_outputs_ext->hplanes[tree] : nullptr;
+            sparse_ix *restrict terminal_indices_this = terminal_indices.data() + nrows * tree;
+            size_t i, j;
+            double add_round;
+
+            if (assume_full_distr)
+            {
+                for (size_t el1 = 0; el1 < n_from; el1++)
+                {
+                    i = terminal_indices_this[el1];
+                    double *ptr_this_sep_ = ptr_this_sep + el1*n_to;
+                    for (size_t el2 = n_from; el2 < nrows; el2++)
+                    {
+                        j = terminal_indices_this[el2];
+                        if (unlikely(i == j))
+                            add_round = node_depths_this[i] + 3.;
+                        else
+                            add_round = node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
+                        ptr_this_sep_[el2-n_from] += add_round;
+                    }
+                }
+            }
+
+            else
+            {
+                hashed_set<size_t> nodes_w_repeated;
+                try
+                {
+                    nodes_w_repeated.reserve(n_terminal_this);
+                    for (size_t el1 = 0; el1 < n_from; el1++)
+                    {
+                        i = terminal_indices_this[el1];
+                        double *ptr_this_sep_ = ptr_this_sep + el1*n_to;
+                        for (size_t el2 = n_from; el2 < nrows; el2++)
+                        {
+                            j = terminal_indices_this[el2];
+                            if (unlikely(i == j))
+                                nodes_w_repeated.insert(i);
+                            else
+                                ptr_this_sep_[el2-n_from]
+                                    +=
+                                node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
+                        }
+                    }
+
+                    if (likely(!nodes_w_repeated.empty()))
+                    {
+                        std::vector<size_t> *restrict argsorted_nodes = &thread_argsorted_nodes[omp_get_thread_num()];
+                        std::iota(argsorted_nodes->begin(), argsorted_nodes->end(), (size_t)0);
+                        std::sort(argsorted_nodes->begin(), argsorted_nodes->end(),
+                                  [&terminal_indices_this](const size_t a, const size_t b)
+                                  {return terminal_indices_this[a] < terminal_indices_this[b];});
+                        std::vector<size_t>::iterator curr_begin = argsorted_nodes->begin();
+                        std::vector<size_t>::iterator new_begin;
+                        
+                        std::vector<size_t> *restrict sorted_nodes = &thread_sorted_nodes[omp_get_thread_num()];
+                        sorted_nodes->assign(nodes_w_repeated.begin(), nodes_w_repeated.end());
+                        std::sort(sorted_nodes->begin(), sorted_nodes->end());
+                        for (size_t node_ix : *sorted_nodes)
+                        {
+                            curr_begin = std::lower_bound(curr_begin, argsorted_nodes->end(),
+                                                          node_ix,
+                                                          [&terminal_indices_this](const size_t &a, const size_t &b)
+                                                          {return (size_t)terminal_indices_this[a] < b;});
+                            new_begin =  std::upper_bound(curr_begin, argsorted_nodes->end(),
+                                                          node_ix,
+                                                          [&terminal_indices_this](const size_t &a, const size_t &b)
+                                                          {return a < (size_t)terminal_indices_this[b];});
+                            size_t n_this = std::distance(curr_begin, new_begin);
+                            if (unlikely(!n_this)) unexpected_error();
+                            double sep_this
+                                =
+                            n_this
+                                +
+                            ((tree_this != NULL)?
+                             (*tree_this)[node_ix].remainder
+                                :
+                             (*hplane_this)[node_ix].remainder);
+                            double sep_this_ = expected_separation_depth(sep_this) + node_depths_this[node_ix];
+
+                            std::vector<size_t> *restrict doubly_argsorted = &thread_doubly_argsorted[omp_get_thread_num()];
+                            doubly_argsorted->assign(curr_begin, curr_begin + n_this);
+                            std::sort(doubly_argsorted->begin(), doubly_argsorted->end());
+                            std::vector<size_t>::iterator pos_n_from = std::lower_bound(doubly_argsorted->begin(),
+                                                                                        doubly_argsorted->end(),
+                                                                                        n_from);
+                            if (pos_n_from == doubly_argsorted->end()) unexpected_error();
+                            size_t n1 = std::distance(doubly_argsorted->begin(), pos_n_from);
+                            size_t i, j;
+                            double *ptr_this_sep__;
+                            for (size_t el1 = 0; el1 < n1; el1++)
+                            {
+                                i = (*doubly_argsorted)[el1];
+                                ptr_this_sep__ = ptr_this_sep + i*n_to;
+                                for (size_t el2 = n1; el2 < n_this; el2++)
+                                {
+                                    j = (*doubly_argsorted)[el2];
+                                    ptr_this_sep__[j-n_from] += sep_this_;
+                                }
+                            }
+
+                            curr_begin = new_begin;
+                        }
+                    }
+                }
+
+                catch (...)
+                {
+                    #pragma omp critical
+                    {
+                        if (!threw_exception)
+                        {
+                            threw_exception = true;
+                            ex = std::current_exception();
+                        }
+                    }
+                }
+            }
+        }
+
+        check_interrupt_switch(ss);
+
+        if (threw_exception)
+            std::rethrow_exception(ex);
+
+        if (nthreads == 1)
+        {
+            /* Here 'rmat' already contains the sum of separations */
+        }
+
+        else
+        {
+            for (int tid = 0; tid < nthreads; tid++)
+            {
+                double *restrict seps_thread = sum_separations[tid].data();
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    rmat[ix] += seps_thread[ix];
+            }
+        }
+
+        check_interrupt_switch(ss);
+
+        if (standardize_dist)
+        {
+            double divisor;
+            if (assume_full_distr)
+                divisor = (double)(ntrees * 2);
+            else
+                divisor = (double)ntrees * ((model_outputs != NULL)? model_outputs->exp_avg_sep : model_outputs_ext->exp_avg_sep);
+
+            if (assume_full_distr)
+            {
+                double ntrees_dbl = (double)ntrees;
+                #ifndef _WIN32
+                #pragma omp simd
+                #endif
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    rmat[ix] = std::exp2( - (rmat[ix] - ntrees_dbl) / divisor);
+            }
+
+            else
+            {
+                #ifndef _WIN32
+                #pragma omp simd
+                #endif
+                for (size_t ix = 0; ix < ncomb; ix++)
+                    rmat[ix] = std::exp2( - rmat[ix] / divisor);
+            }
+        }
+
+        else
+        {
+            double divisor = (double)ntrees;
+            for (size_t ix = 0; ix < ncomb; ix++)
+                rmat[ix] /= divisor;
+        }
+
+        check_interrupt_switch(ss);
+    }
+}
+
+template <class real_t, class sparse_ix>
+void calc_similarity_from_indexer_with_references
+(
+    real_t *restrict numeric_data, int *restrict categ_data,
+    real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
+    size_t nrows, int nthreads, bool standardize_dist,
+    IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+    double *restrict rmat,
+    TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ
+)
+{
+    size_t n_ref = get_number_of_reference_points(*indexer);
+    if (unlikely(!n_ref)) unexpected_error();
+
+    SignalSwitcher ss;
+
+    size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
+    std::vector<sparse_ix> terminal_indices(nrows * ntrees);
+    std::unique_ptr<double[]> ignored(new double[nrows]);
+    predict_iforest(numeric_data, categ_data,
+                    is_col_major, ld_numeric, ld_categ,
+                    is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
+                    is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
+                    nrows, nthreads, false,
+                    model_outputs, model_outputs_ext,
+                    ignored.get(), terminal_indices.data(),
+                    (double*)NULL,
+                    indexer);
+    ignored.reset();
+
+    #ifndef _OPENMP
+    nthreads = 1;
+    #endif
+
+    check_interrupt_switch(ss);
+
+    #pragma omp parallel for schedule(static) num_threads(nthreads) \
+            shared(rmat, terminal_indices, nrows, n_ref, indexer, ntrees)
+    for (size_t_for row = 0; row < (decltype(row))nrows; row++)
+    {
+        if (interrupt_switch) continue;
+
+        size_t i, j;
+        size_t n_terminal_this;
+        size_t ncomb_this;
+        size_t *restrict ref_this;
+        sparse_ix *restrict ind_this;
+        double *restrict node_depths_this;
+        double *restrict node_dist_this;
+        double *rmat_this = rmat + row*n_ref;
+        memset(rmat_this, 0, n_ref*sizeof(double));
+        for (size_t tree = 0; tree < ntrees; tree++)
+        {
+            ref_this = indexer->indices[tree].reference_points.data();
+            ind_this = terminal_indices.data() + tree*nrows;
+            node_depths_this = indexer->indices[tree].node_depths.data();
+            n_terminal_this = indexer->indices[tree].n_terminal;
+            node_dist_this = indexer->indices[tree].node_distances.data();
+            ncomb_this = calc_ncomb(n_terminal_this);
+            for (size_t ref = 0; ref < n_ref; ref++)
+            {
+                i = ind_this[row];
+                j = ref_this[ref];
+
+                if (unlikely(i == j))
+                    rmat_this[ref] += node_depths_this[i] + 3.;
+                else
+                    rmat_this[ref] += node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
+            }
+        }
+    }
+
+    check_interrupt_switch(ss);
+
+    size_t size_rmat = nrows * n_ref;
+    if (standardize_dist)
+    {
+        double ntrees_dbl = (double)ntrees;
+        double div_trees = (double)(mult2(ntrees));
+        #ifndef _WIN32
+        #pragma omp simd
+        #endif
+        for (size_t ix = 0; ix < size_rmat; ix++)
+            rmat[ix] = std::exp2( - (rmat[ix] - ntrees_dbl) / div_trees);
+    }
+
+    else
+    {
+        double div_trees = (double)ntrees;
+        for (size_t ix = 0; ix < size_rmat; ix++)
+            rmat[ix] /= div_trees;
+    }
+
+    check_interrupt_switch(ss);
+}
+
+template <class real_t, class sparse_ix>
+void kernel_to_references(TreesIndexer &indexer,
+                          IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
+                          real_t *restrict numeric_data, int *restrict categ_data,
+                          real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
+                          bool is_col_major, size_t ld_numeric, size_t ld_categ,
+                          size_t nrows, int nthreads,
+                          double *restrict rmat,
+                          bool standardize)
+{
+    size_t ntrees = indexer.indices.size();
+    size_t n_ref = indexer.indices.front().reference_points.size();
+
+    SignalSwitcher ss;
+
+    std::unique_ptr<sparse_ix[]> terminal_indices(new sparse_ix[nrows*ntrees]);
+    std::unique_ptr<double[]> ignored(new double[nrows]);
+    predict_iforest(numeric_data, categ_data,
+                    is_col_major, ld_numeric, ld_categ,
+                    is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
+                    is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
+                    nrows, nthreads, false,
+                    model_outputs, model_outputs_ext,
+                    ignored.get(), terminal_indices.get(),
+                    (double*)NULL,
+                    &indexer);
+    ignored.reset();
+
+    check_interrupt_switch(ss);
+
+    #pragma omp parallel for schedule(static) num_threads(nthreads) \
+            shared(indexer, terminal_indices, nrows, ntrees, n_ref, rmat)
+    for (size_t_for row = 0; row < (decltype(row))nrows; row++)
+    {
+        if (interrupt_switch) continue;
+
+        SingleTreeIndex *restrict index_node;
+        size_t idx_this;
+        sparse_ix *restrict terminal_indices_this = terminal_indices.get() + row;
+        double *restrict rmat_this = rmat + row*n_ref;
+        memset(rmat_this, 0, n_ref*sizeof(double));
+
+        for (size_t tree = 0; tree < ntrees; tree++)
+        {
+            idx_this = terminal_indices_this[tree*nrows];
+            index_node = &indexer.indices[tree];
+            for (size_t ind = index_node->reference_indptr[idx_this];
+                 ind < index_node->reference_indptr[idx_this + 1];
+                 ind++)
+            {
+                rmat_this[index_node->reference_mapping[ind]]++;
+            }
+        }
+    }
+
+    check_interrupt_switch(ss);
+
+    if (standardize)
+    {
+        double ntrees_dbl = (double)ntrees;
+        for (size_t ix = 0; ix < nrows*n_ref; ix++)
+            rmat[ix] /= ntrees_dbl;
+    }
+
+    check_interrupt_switch(ss);
+}