sparsekmeans 0.1__tar.gz → 0.2__tar.gz

sparsekmeans-0.2/PKG-INFO

@@ -0,0 +1,12 @@
+Metadata-Version: 2.1
+Name: sparsekmeans
+Version: 0.2
+Home-page: https://github.com/cjlin1/sparsekmeans
+Author: Chih-Jen Lin, He-Zhe Lin, Khoi Nguyen Pham Dang
+Author-email: cjlin@csie.ntu.edu.tw
+License: MIT
+Requires-Python: >=3.10
+License-File: LICENSE
+Requires-Dist: numpy
+Requires-Dist: python-graphblas
+Requires-Dist: scipy

{sparsekmeans-0.1 → sparsekmeans-0.2}/README.md

@@ -12,7 +12,11 @@ pip install sparsekmeans
 
 ## Usage
 
-We support two popular K-means algorithms: Lloyd's method and Elkan's method. Use the following steps to cluster a data set:
+We support two popular K-means algorithms: Lloyd's method and Elkan's method. While both require distance calculations between data points and centroids, Elkan's method reduces some computations by exploiting information from the previous iteration, specifically using the triangle inequality to maintain two distance-related bounds.
+
+When the data set is small (e.g., the number of samples is less than 10,000), it is recommended to use Lloyd's. Lloyd's relies primarily on sparse matrix multiplications, whereas Elkan's mechanism to reduce distance calculations may cause non-negligible overhead for small problems. Conversely, for large data sets or the situation of a huge number of clusters (K), Elkan's method may significantly outperforms Lloyd's.
+
+Use the following steps to cluster a data set:
 
 ```
 from sparsekmeans import LloydKmeans, ElkanKmeans
@@ -63,9 +67,8 @@ import urllib.request
 import bz2
 
 with urllib.request.urlopen("https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass/news20_tfidf_train.svm.bz2") as r:
-    r = bz2.BZ2File(r)
-    r = io.TextIOWrapper(io.BufferedReader(r))
-    _, X = svm_read_problem(r, return_scipy=True)
+    with bz2.open(r, 'rt') as f:
+        _, X = svm_read_problem(f, return_scipy=True)
 
 kmeans = LloydKmeans(n_clusters=100)
 labels = kmeans.fit(X)

{sparsekmeans-0.1 → sparsekmeans-0.2}/setup.py

@@ -2,7 +2,7 @@ from setuptools import setup
 
 PACKAGE_DIR = ["sparsekmeans"]
 PACKAGE_NAME = "sparsekmeans"
-VERSION = "0.1"
+VERSION = "0.2"
 
 
 # license parameters
@@ -16,8 +16,8 @@ setup(
     packages=PACKAGE_DIR,
     version=VERSION,
     python_requires=">=3.10",
-    install_requires=["numpy","python-graphblas","scipy","libsvm-official>=3.36.0"],
-    description="A package for efficient K-means clustering on sparse dataset",
+    install_requires=["numpy","python-graphblas","scipy"],
+    description="",
     long_description_content_type="",
     author="Chih-Jen Lin, He-Zhe Lin, Khoi Nguyen Pham Dang",
     author_email="cjlin@csie.ntu.edu.tw",

{sparsekmeans-0.1 → sparsekmeans-0.2}/sparsekmeans/sparse_kmeans.py

@@ -3,6 +3,7 @@ import sys
 import numpy as np
 import scipy.sparse as sparse
 import time
+import warnings
 from typing import Union
 
 import graphblas as gb
@@ -44,7 +45,7 @@ def check_centroids_density(centroids: gb.Matrix):
         is_centroid_dense = False
         centroids_density = centroids.nvals / (centroids.nrows * centroids.ncols)
         if centroids_density > gamma:
-            is_centroid_dense = False
+            is_centroid_dense = True
             centroids(mask=~centroids.S) << 0
 
     return centroids, is_centroid_dense
@@ -58,7 +59,7 @@ def predict_labels(X: gb.Matrix, centroids: gb.Matrix, is_centroid_dense: bool,
     c_squared_norms = squared_row_norms(centroids, n_threads)
     XCt = Matrix(dtypes.FP64, nrows=n_samples, ncols=n_clusters)
 
-    # For X * Ct, we use CSR x (fullc or CSC) for the best efficiency.
+    # For X * Ct, we use CSR (for X format) x (fullc or CSC for C format) for the best efficiency.
     # We use row format to store centroids in other places, so must convert them here to column format.
     if is_centroid_dense:
         centroids = centroids.ss.export("fullc")
@@ -197,6 +198,13 @@ class SparseKmeans:
         n_local_trials = 2 + int(np.log(n_clusters))
 
         for j in range(1, n_clusters):
+
+            if current_potential < 1e-12:
+                warnings.warn(f"The specified number of clusters is larger than the number of distinct points.\nChange the number of clusters to the number of distinct points.", category=UserWarning)
+                centroids = centroids[:j, :]
+                self.n_clusters = j
+                break
+
             # Choose centroids candidates by sampling with probability smallest_sq_dist/current_potential, where
             # current_potential is the sum of smallest_sq_dist
             rand_vals = self.random_state.uniform(size=n_local_trials) * current_potential
@@ -241,7 +249,6 @@ class SparseKmeans:
         # Allowing the kmeans procedure to allocate and initialize internal variables
         self._setup_internal_state(X)
 
-        # Calculate the mean feature variance
         self.mean_feature_variance = self._cal_feature_variance(X)
         if self.mean_feature_variance <= sys.float_info.min:
             return
@@ -344,20 +351,29 @@ class SparseKmeans:
         empty_clusters = np.where(cluster_sizes == 0)[0]
         n_empty = len(empty_clusters)
 
-        # To handle empty clusters, we follow the setting in scikit-learn
-        # We consider points with the largest distances to their assigned clusters, and reassign them to the empty clusters
+        # To handle empty clusters, we consider points with the largest distances to their assigned clusters, and reassign them to the empty clusters
+        # If a selected point belongs to a cluster with only one point, we skip it and move on to the next furthest point.
         # Each empty cluster gets one point
         # We also need to update cluster_sizes
         if n_empty > 0:
-            far_samples_idx = np.argpartition(self.sample_centroids_closest_distance, -n_empty)[
-                : -n_empty - 1 : -1
-            ]
-            affected_clusters, counts = np.unique(
-                self.labels[far_samples_idx], return_counts=True
-            )
-            cluster_sizes[affected_clusters] -= counts
-            cluster_sizes[empty_clusters] = 1
-            self.labels[far_samples_idx] = empty_clusters
+            far_samples_idx = np.argsort(self.sample_centroids_closest_distance)[:: -1]
+            i, j = 0, 0
+            
+            while i < n_samples - n_empty and j < n_empty:
+                origin_cluster = self.labels[far_samples_idx[i]]
+                i += 1
+
+                if cluster_sizes[origin_cluster] == 1:
+                    continue
+
+                target_empty_cluster = empty_clusters[j]
+                cluster_sizes[origin_cluster] -= 1
+                cluster_sizes[target_empty_cluster] = 1
+                self.labels[far_samples_idx[i]] = target_empty_cluster
+                j += 1
+            
+            if i == n_samples - n_empty:
+                raise ValueError("Not enough distinct points to assign to empty clusters.")
 
         # Get the weight of each sample in its corresponding cluster
         weights = 1 / cluster_sizes

sparsekmeans-0.2/sparsekmeans.egg-info/PKG-INFO

@@ -0,0 +1,12 @@
+Metadata-Version: 2.1
+Name: sparsekmeans
+Version: 0.2
+Home-page: https://github.com/cjlin1/sparsekmeans
+Author: Chih-Jen Lin, He-Zhe Lin, Khoi Nguyen Pham Dang
+Author-email: cjlin@csie.ntu.edu.tw
+License: MIT
+Requires-Python: >=3.10
+License-File: LICENSE
+Requires-Dist: numpy
+Requires-Dist: python-graphblas
+Requires-Dist: scipy

{sparsekmeans-0.1 → sparsekmeans-0.2}/sparsekmeans.egg-info/requires.txt

@@ -1,4 +1,3 @@
 numpy
 python-graphblas
 scipy
-libsvm-official>=3.36.0

sparsekmeans-0.1/PKG-INFO

@@ -1,22 +0,0 @@
-Metadata-Version: 2.4
-Name: sparsekmeans
-Version: 0.1
-Summary: A package for efficient K-means clustering on sparse dataset
-Home-page: https://github.com/cjlin1/sparsekmeans
-Author: Chih-Jen Lin, He-Zhe Lin, Khoi Nguyen Pham Dang
-Author-email: cjlin@csie.ntu.edu.tw
-License: MIT
-Requires-Python: >=3.10
-License-File: LICENSE
-Requires-Dist: numpy
-Requires-Dist: python-graphblas
-Requires-Dist: scipy
-Requires-Dist: libsvm-official>=3.36.0
-Dynamic: author
-Dynamic: author-email
-Dynamic: home-page
-Dynamic: license
-Dynamic: license-file
-Dynamic: requires-dist
-Dynamic: requires-python
-Dynamic: summary

sparsekmeans-0.1/sparsekmeans.egg-info/PKG-INFO

@@ -1,22 +0,0 @@
-Metadata-Version: 2.4
-Name: sparsekmeans
-Version: 0.1
-Summary: A package for efficient K-means clustering on sparse dataset
-Home-page: https://github.com/cjlin1/sparsekmeans
-Author: Chih-Jen Lin, He-Zhe Lin, Khoi Nguyen Pham Dang
-Author-email: cjlin@csie.ntu.edu.tw
-License: MIT
-Requires-Python: >=3.10
-License-File: LICENSE
-Requires-Dist: numpy
-Requires-Dist: python-graphblas
-Requires-Dist: scipy
-Requires-Dist: libsvm-official>=3.36.0
-Dynamic: author
-Dynamic: author-email
-Dynamic: home-page
-Dynamic: license
-Dynamic: license-file
-Dynamic: requires-dist
-Dynamic: requires-python
-Dynamic: summary