replay-rec 0.17.1__py3-none-any.whl → 0.17.1rc0__py3-none-any.whl

replay/__init__.py

@@ -1,2 +1,2 @@
 """ RecSys library """
-__version__ = "0.17.1"
+__version__ = "0.17.1.preview"

replay/experimental/metrics/__init__.py

@@ -0,0 +1,61 @@
+"""
+Most metrics require dataframe with recommendations
+and dataframe with ground truth values —
+which objects each user interacted with.
+
+- recommendations (Union[pandas.DataFrame, spark.DataFrame]):
+    predictions of a recommender system,
+    DataFrame with columns ``[user_id, item_id, relevance]``
+- ground_truth (Union[pandas.DataFrame, spark.DataFrame]):
+    test data, DataFrame with columns
+    ``[user_id, item_id, timestamp, relevance]``
+
+Metric is calculated for all users, presented in ``ground_truth``
+for accurate metric calculation in case when the recommender system generated
+recommendation not for all users.  It is assumed, that all users,
+we want to calculate metric for, have positive interactions.
+
+But if we have users, who observed the recommendations, but have not responded,
+those users will be ignored and metric will be overestimated.
+For such case we propose additional optional parameter ``ground_truth_users``,
+the dataframe with all users, which should be considered during the metric calculation.
+
+- ground_truth_users (Optional[Union[pandas.DataFrame, spark.DataFrame]]):
+    full list of users to calculate metric for, DataFrame with ``user_id`` column
+
+Every metric is calculated using top ``K`` items for each user.
+It is also possible to calculate metrics
+using multiple values for ``K`` simultaneously.
+In this case the result will be a dictionary and not a number.
+
+Make sure your recommendations do not contain user-item duplicates
+as duplicates could lead to the wrong calculation results.
+
+- k (Union[Iterable[int], int]):
+    a single number or a list, specifying the
+    truncation length for recommendation list for each user
+
+By default, metrics are averaged by users,
+but you can alternatively use method ``metric.median``.
+Also, you can get the lower bound
+of ``conf_interval`` for a given ``alpha``.
+
+Diversity metrics require extra parameters on initialization stage,
+but do not use ``ground_truth`` parameter.
+
+For each metric, a formula for its calculation is given, because this is
+important for the correct comparison of algorithms, as mentioned in our
+`article <https://arxiv.org/abs/2206.12858>`_.
+"""
+from replay.experimental.metrics.base_metric import Metric, NCISMetric
+from replay.experimental.metrics.coverage import Coverage
+from replay.experimental.metrics.hitrate import HitRate
+from replay.experimental.metrics.map import MAP
+from replay.experimental.metrics.mrr import MRR
+from replay.experimental.metrics.ncis_precision import NCISPrecision
+from replay.experimental.metrics.ndcg import NDCG
+from replay.experimental.metrics.precision import Precision
+from replay.experimental.metrics.recall import Recall
+from replay.experimental.metrics.rocauc import RocAuc
+from replay.experimental.metrics.surprisal import Surprisal
+from replay.experimental.metrics.unexpectedness import Unexpectedness

replay/experimental/metrics/base_metric.py

@@ -0,0 +1,601 @@
+"""
+Base classes for quality and diversity metrics.
+"""
+import logging
+from abc import ABC, abstractmethod
+from typing import Dict, List, Optional, Union
+
+from scipy.stats import norm
+
+from replay.utils import PYSPARK_AVAILABLE, DataFrameLike, IntOrList, NumType, PandasDataFrame, SparkDataFrame
+from replay.utils.session_handler import State
+from replay.utils.spark_utils import convert2spark, get_top_k_recs
+
+if PYSPARK_AVAILABLE:
+    from pyspark.sql import (
+        Column,
+        Window,
+        functions as sf,
+        types as st,
+    )
+    from pyspark.sql.column import _to_java_column, _to_seq
+    from pyspark.sql.types import DataType
+
+
+def fill_na_with_empty_array(df: SparkDataFrame, col_name: str, element_type: DataType) -> SparkDataFrame:
+    """
+    Fill empty values in array column with empty array of `element_type` values.
+    :param df: dataframe with `col_name` column of ArrayType(`element_type`)
+    :param col_name: name of a column to fill missing values
+    :param element_type: DataType of an array element
+    :return: df with `col_name` na values filled with empty arrays
+    """
+    return df.withColumn(
+        col_name,
+        sf.coalesce(
+            col_name,
+            sf.array().cast(st.ArrayType(element_type)),
+        ),
+    )
+
+
+def preprocess_gt(
+    ground_truth: DataFrameLike,
+    ground_truth_users: Optional[DataFrameLike] = None,
+) -> SparkDataFrame:
+    """
+    Preprocess `ground_truth` data before metric calculation
+    :param ground_truth: spark dataframe with columns ``[user_idx, item_idx, relevance]``
+    :param ground_truth_users: spark dataframe with column ``[user_idx]``
+    :return: spark dataframe with columns ``[user_idx, ground_truth]``
+    """
+    ground_truth = convert2spark(ground_truth)
+    ground_truth_users = convert2spark(ground_truth_users)
+
+    true_items_by_users = ground_truth.groupby("user_idx").agg(sf.collect_set("item_idx").alias("ground_truth"))
+    if ground_truth_users is not None:
+        true_items_by_users = true_items_by_users.join(ground_truth_users, on="user_idx", how="right")
+        true_items_by_users = fill_na_with_empty_array(
+            true_items_by_users,
+            "ground_truth",
+            ground_truth.schema["item_idx"].dataType,
+        )
+
+    return true_items_by_users
+
+
+def drop_duplicates(recommendations: DataFrameLike) -> SparkDataFrame:
+    """
+    Filter duplicated predictions by choosing the most relevant
+    """
+    return (
+        recommendations.withColumn(
+            "_num",
+            sf.row_number().over(Window.partitionBy("user_idx", "item_idx").orderBy(sf.col("relevance").desc())),
+        )
+        .where(sf.col("_num") == 1)
+        .drop("_num")
+    )
+
+
+def filter_sort(recommendations: SparkDataFrame, extra_column: Optional[str] = None) -> SparkDataFrame:
+    """
+    Filters duplicated predictions by choosing items with the highest relevance,
+    Sorts items in predictions by its relevance,
+    If `extra_column` is not None return DataFrame with extra_column e.g. item weight.
+
+    :param recommendations: recommendation list
+    :param extra_column: column in recommendations
+        which will be return besides ``[user_idx, item_idx]``
+    :return: ``[user_idx, item_idx]`` if extra_column = None
+        or ``[user_idx, item_idx, extra_column]`` if extra_column exists.
+    """
+    item_type = recommendations.schema["item_idx"].dataType
+    extra_column_type = recommendations.schema[extra_column].dataType if extra_column else None
+
+    recommendations = drop_duplicates(recommendations)
+
+    recommendations = (
+        recommendations.groupby("user_idx")
+        .agg(
+            sf.collect_list(sf.struct(*[c for c in ["relevance", "item_idx", extra_column] if c is not None])).alias(
+                "pred_list"
+            )
+        )
+        .withColumn("pred_list", sf.reverse(sf.array_sort("pred_list")))
+    )
+
+    selection = ["user_idx", sf.col("pred_list.item_idx").cast(st.ArrayType(item_type, True)).alias("pred")]
+    if extra_column:
+        selection.append(
+            sf.col(f"pred_list.{extra_column}").cast(st.ArrayType(extra_column_type, True)).alias(extra_column)
+        )
+
+    recommendations = recommendations.select(*selection)
+
+    return recommendations
+
+
+def get_enriched_recommendations(
+    recommendations: DataFrameLike,
+    ground_truth: DataFrameLike,
+    max_k: int,
+    ground_truth_users: Optional[DataFrameLike] = None,
+) -> SparkDataFrame:
+    """
+    Leave max_k recommendations for each user,
+    merge recommendations and ground truth into a single DataFrame
+    and aggregate items into lists so that each user has only one record.
+
+    :param recommendations: recommendation list
+    :param ground_truth: test data
+    :param max_k: maximal k value to calculate the metric for.
+        `max_k` most relevant predictions are left for each user
+    :param ground_truth_users: list of users to consider in metric calculation.
+        if None, only the users from ground_truth are considered.
+    :return:  ``[user_idx, pred, ground_truth]``
+    """
+    recommendations = convert2spark(recommendations)
+    # if there are duplicates in recommendations,
+    # we will leave fewer than k recommendations after sort_udf
+    recommendations = get_top_k_recs(recommendations, k=max_k)
+
+    true_items_by_users = preprocess_gt(ground_truth, ground_truth_users)
+    joined = filter_sort(recommendations).join(true_items_by_users, how="right", on=["user_idx"])
+
+    return fill_na_with_empty_array(joined, "pred", recommendations.schema["item_idx"].dataType)
+
+
+def process_k(func):
+    """Decorator that converts k to list and unpacks result"""
+
+    def wrap(self, recs: SparkDataFrame, k: IntOrList, *args):
+        k_list = [k] if isinstance(k, int) else k
+
+        res = func(self, recs, k_list, *args)
+
+        if isinstance(k, int):
+            return res[k]
+        return res
+
+    return wrap
+
+
+class Metric(ABC):
+    """Base metric class"""
+
+    _logger: Optional[logging.Logger] = None
+    _scala_udf_name: Optional[str] = None
+
+    def __init__(self, use_scala_udf: bool = False) -> None:
+        self._use_scala_udf = use_scala_udf
+
+    @property
+    def logger(self) -> logging.Logger:
+        """
+        :returns: get library logger
+        """
+        if self._logger is None:
+            self._logger = logging.getLogger("replay")
+        return self._logger
+
+    @property
+    def scala_udf_name(self) -> str:
+        """Returns UDF name from `org.apache.spark.replay.utils.ScalaPySparkUDFs`"""
+        if self._scala_udf_name:
+            return self._scala_udf_name
+        else:
+            msg = f"Scala UDF not implemented for {type(self).__name__} class!"
+            raise NotImplementedError(msg)
+
+    def __str__(self):
+        return type(self).__name__
+
+    def __call__(
+        self,
+        recommendations: DataFrameLike,
+        ground_truth: DataFrameLike,
+        k: IntOrList,
+        ground_truth_users: Optional[DataFrameLike] = None,
+    ) -> Union[Dict[int, NumType], NumType]:
+        """
+        :param recommendations: model predictions in a
+            DataFrame ``[user_idx, item_idx, relevance]``
+        :param ground_truth: test data
+            ``[user_idx, item_idx, timestamp, relevance]``
+        :param k: depth cut-off. Truncates recommendation lists to top-k items.
+        :param ground_truth_users: list of users to consider in metric calculation.
+            if None, only the users from ground_truth are considered.
+        :return: metric value
+        """
+        recs = get_enriched_recommendations(
+            recommendations,
+            ground_truth,
+            max_k=k if isinstance(k, int) else max(k),
+            ground_truth_users=ground_truth_users,
+        )
+        return self._mean(recs, k)
+
+    @process_k
+    def _conf_interval(self, recs: SparkDataFrame, k_list: list, alpha: float):
+        res = {}
+        quantile = norm.ppf((1 + alpha) / 2)
+        for k in k_list:
+            distribution = self._get_metric_distribution(recs, k)
+            value = (
+                distribution.agg(
+                    sf.stddev("value").alias("std"),
+                    sf.count("value").alias("count"),
+                )
+                .select(
+                    sf.when(
+                        sf.isnan(sf.col("std")) | sf.col("std").isNull(),
+                        sf.lit(0.0),
+                    )
+                    .otherwise(sf.col("std"))
+                    .cast("float")
+                    .alias("std"),
+                    "count",
+                )
+                .first()
+            )
+            res[k] = quantile * value["std"] / (value["count"] ** 0.5)
+        return res
+
+    @process_k
+    def _median(self, recs: SparkDataFrame, k_list: list):
+        res = {}
+        for k in k_list:
+            distribution = self._get_metric_distribution(recs, k)
+            value = distribution.agg(sf.expr("percentile_approx(value, 0.5)").alias("value")).first()["value"]
+            res[k] = value
+        return res
+
+    @process_k
+    def _mean(self, recs: SparkDataFrame, k_list: list):
+        res = {}
+        for k in k_list:
+            distribution = self._get_metric_distribution(recs, k)
+            value = distribution.agg(sf.avg("value").alias("value")).first()["value"]
+            res[k] = value
+        return res
+
+    def _get_metric_distribution(self, recs: SparkDataFrame, k: int) -> SparkDataFrame:
+        """
+        :param recs: recommendations
+        :param k: depth cut-off
+        :return: metric distribution for different cut-offs and users
+        """
+        if self._use_scala_udf:
+            metric_value_col = self.get_scala_udf(self.scala_udf_name, [sf.lit(k).alias("k"), *recs.columns[1:]]).alias(
+                "value"
+            )
+            return recs.select("user_idx", metric_value_col)
+
+        cur_class = self.__class__
+        distribution = recs.rdd.flatMap(lambda x: [(x[0], float(cur_class._get_metric_value_by_user(k, *x[1:])))]).toDF(
+            f"user_idx {recs.schema['user_idx'].dataType.typeName()}, value double"
+        )
+        return distribution
+
+    @staticmethod
+    @abstractmethod
+    def _get_metric_value_by_user(k, pred, ground_truth) -> float:
+        """
+        Metric calculation for one user.
+
+        :param k: depth cut-off
+        :param pred: recommendations
+        :param ground_truth: test data
+        :return: metric value for current user
+        """
+
+    def user_distribution(
+        self,
+        log: DataFrameLike,
+        recommendations: DataFrameLike,
+        ground_truth: DataFrameLike,
+        k: IntOrList,
+        ground_truth_users: Optional[DataFrameLike] = None,
+    ) -> PandasDataFrame:
+        """
+        Get mean value of metric for all users with the same number of ratings.
+
+        :param log: history DataFrame to calculate number of ratings per user
+        :param recommendations: prediction DataFrame
+        :param ground_truth: test data
+        :param k: depth cut-off
+        :param ground_truth_users: list of users to consider in metric calculation.
+            if None, only the users from ground_truth are considered.
+        :return: pandas DataFrame
+        """
+        log = convert2spark(log)
+        count = log.groupBy("user_idx").count()
+        if hasattr(self, "_get_enriched_recommendations"):
+            recs = self._get_enriched_recommendations(
+                recommendations,
+                ground_truth,
+                max_k=k if isinstance(k, int) else max(k),
+                ground_truth_users=ground_truth_users,
+            )
+        else:
+            recs = get_enriched_recommendations(
+                recommendations,
+                ground_truth,
+                max_k=k if isinstance(k, int) else max(k),
+                ground_truth_users=ground_truth_users,
+            )
+        k_list = [k] if isinstance(k, int) else k
+        res = PandasDataFrame()
+        for cut_off in k_list:
+            dist = self._get_metric_distribution(recs, cut_off)
+            val = count.join(dist, on="user_idx", how="right").fillna(0, subset="count")
+            val = (
+                val.groupBy("count")
+                .agg(sf.avg("value").alias("value"))
+                .orderBy(["count"])
+                .select("count", "value")
+                .toPandas()
+            )
+            res = res.append(val, ignore_index=True)
+        return res
+
+    @staticmethod
+    def get_scala_udf(udf_name: str, params: List) -> Column:
+        """
+        Returns expression of calling scala UDF as column
+
+        :param udf_name: UDF name from `org.apache.spark.replay.utils.ScalaPySparkUDFs`
+        :param params: list of UDF params in right order
+        :return: column expression
+        """
+        sc = State().session.sparkContext
+        scala_udf = getattr(sc._jvm.org.apache.spark.replay.utils.ScalaPySparkUDFs, udf_name)()
+        return Column(scala_udf.apply(_to_seq(sc, params, _to_java_column)))
+
+
+class RecOnlyMetric(Metric):
+    """Base class for metrics that do not need holdout data"""
+
+    @abstractmethod
+    def __init__(self, log: DataFrameLike, *args, **kwargs):
+        pass
+
+    @abstractmethod
+    def _get_enriched_recommendations(
+        self,
+        recommendations: DataFrameLike,
+        ground_truth: Optional[DataFrameLike],
+        max_k: int,
+        ground_truth_users: Optional[DataFrameLike] = None,
+    ) -> SparkDataFrame:
+        pass
+
+    def __call__(
+        self,
+        recommendations: DataFrameLike,
+        k: IntOrList,
+        ground_truth_users: Optional[DataFrameLike] = None,
+    ) -> Union[Dict[int, NumType], NumType]:
+        """
+        :param recommendations: predictions of a model,
+            DataFrame  ``[user_idx, item_idx, relevance]``
+        :param k: depth cut-off
+        :param ground_truth_users: list of users to consider in metric calculation.
+            if None, only the users from ground_truth are considered.
+        :return: metric value
+        """
+        recs = self._get_enriched_recommendations(
+            recommendations,
+            None,
+            max_k=k if isinstance(k, int) else max(k),
+            ground_truth_users=ground_truth_users,
+        )
+        return self._mean(recs, k)
+
+    @staticmethod
+    @abstractmethod
+    def _get_metric_value_by_user(k, *args) -> float:
+        """
+        Metric calculation for one user.
+
+        :param k: depth cut-off
+        :param *args: extra parameters, returned by
+            '''self._get_enriched_recommendations''' method
+        :return: metric value for current user
+        """
+
+
+class NCISMetric(Metric):
+    """
+    RePlay implements Normalized Capped Importance Sampling for metric calculation with ``NCISMetric`` class.
+    This method is mostly applied to RL-based recommendation systems to perform counterfactual evaluation, but could be
+    used for any kind of recommender systems. See an article
+    `Offline A/B testing for Recommender Systems <http://arxiv.org/abs/1801.07030>` for details.
+
+    *Reward* (metric value for a user-item pair) is weighed by
+    the ratio of *current policy score* (current relevance) on *previous policy score* (historical relevance).
+
+    The *weight* is clipped by the *threshold* and put into interval :math:`[\\frac{1}{threshold}, threshold]`.
+    Activation function (e.g. softmax, sigmoid) could be applied to the scores before weights calculation.
+
+    Normalization weight for recommended item is calculated as follows:
+
+    .. math::
+        w_{ui} = \\frac{f(\\pi^t_ui, pi^t_u)}{f(\\pi^p_ui, pi^p_u)}
+
+    Where:
+
+    :math:`\\pi^t_{ui}` - current policy value (predicted relevance) of the user-item interaction
+
+    :math:`\\pi^p_{ui}` - previous policy value (historical relevance) of the user-item interaction.
+    Only values for user-item pairs present in current recommendations are used for calculation.
+
+    :math:`\\pi_u` - all predicted /historical policy values for selected user :math:`u`
+
+    :math:`f(\\pi_{ui}, \\pi_u)` - activation function applied to policy values (optional)
+
+    :math:`w_{ui}` - weight of user-item interaction for normalized metric calculation before clipping
+
+
+    Calculated weights are clipped as follows:
+
+    .. math::
+        \\hat{w_{ui}} = min(max(\\frac{1}{threshold}, w_{ui}), threshold)
+
+    Normalization metric value for a user is calculated as follows:
+
+    .. math::
+        R_u = \\frac{r_{ui} \\hat{w_{ui}}}{\\sum_{i}\\hat{w_{ui}}}
+
+    Where:
+
+    :math:`r_ui` - metric value (reward) for user-item interaction
+
+    :math:`R_u` - metric value (reward) for user :math:`u`
+
+    Weight calculation is implemented in ``_get_enriched_recommendations`` method.
+    """
+
+    def __init__(
+        self,
+        prev_policy_weights: DataFrameLike,
+        threshold: float = 10.0,
+        activation: Optional[str] = None,
+        use_scala_udf: bool = False,
+    ):
+        """
+        :param prev_policy_weights: historical item of user-item relevance (previous policy values)
+        :threshold: capping threshold, applied after activation,
+            relevance values are cropped to interval [1/`threshold`, `threshold`]
+        :activation: activation function, applied over relevance values.
+            "logit"/"sigmoid", "softmax" or None
+        """
+        self._use_scala_udf = use_scala_udf
+        self.prev_policy_weights = convert2spark(prev_policy_weights).withColumnRenamed("relevance", "prev_relevance")
+        self.threshold = threshold
+        if activation is None or activation in ("logit", "sigmoid", "softmax"):
+            self.activation = activation
+            if activation == "softmax":
+                self.logger.info(
+                    "For accurate softmax calculation pass only one `k` value in the NCISMetric metrics `call`"
+                )
+        else:
+            msg = f"Unexpected `activation` - {activation}"
+            raise ValueError(msg)
+        if threshold <= 0:
+            msg = "Threshold should be positive real number"
+            raise ValueError(msg)
+
+    @staticmethod
+    def _softmax_by_user(df: SparkDataFrame, col_name: str) -> SparkDataFrame:
+        """
+        Subtract minimal value (relevance) by user from `col_name`
+        and apply softmax by user to `col_name`.
+        """
+        return (
+            df.withColumn(
+                "_min_rel_user",
+                sf.min(col_name).over(Window.partitionBy("user_idx")),
+            )
+            .withColumn(col_name, sf.exp(sf.col(col_name) - sf.col("_min_rel_user")))
+            .withColumn(
+                col_name,
+                sf.col(col_name) / sf.sum(col_name).over(Window.partitionBy("user_idx")),
+            )
+            .drop("_min_rel_user")
+        )
+
+    @staticmethod
+    def _sigmoid(df: SparkDataFrame, col_name: str) -> SparkDataFrame:
+        """
+        Apply sigmoid/logistic function to column `col_name`
+        """
+        return df.withColumn(col_name, sf.lit(1.0) / (sf.lit(1.0) + sf.exp(-sf.col(col_name))))
+
+    @staticmethod
+    def _weigh_and_clip(
+        df: SparkDataFrame,
+        threshold: float,
+        target_policy_col: str = "relevance",
+        prev_policy_col: str = "prev_relevance",
+    ):
+        """
+        Clip weights to fit into interval [1/threshold, threshold].
+        """
+        lower, upper = 1 / threshold, threshold
+        return (
+            df.withColumn(
+                "weight_unbounded",
+                sf.col(target_policy_col) / sf.col(prev_policy_col),
+            )
+            .withColumn(
+                "weight",
+                sf.when(sf.col(prev_policy_col) == sf.lit(0.0), sf.lit(upper))
+                .when(sf.col("weight_unbounded") < sf.lit(lower), sf.lit(lower))
+                .when(sf.col("weight_unbounded") > sf.lit(upper), sf.lit(upper))
+                .otherwise(sf.col("weight_unbounded")),
+            )
+            .select("user_idx", "item_idx", "relevance", "weight")
+        )
+
+    def _reweighing(self, recommendations):
+        if self.activation == "softmax":
+            recommendations = self._softmax_by_user(recommendations, col_name="prev_relevance")
+            recommendations = self._softmax_by_user(recommendations, col_name="relevance")
+        elif self.activation in ["logit", "sigmoid"]:
+            recommendations = self._sigmoid(recommendations, col_name="prev_relevance")
+            recommendations = self._sigmoid(recommendations, col_name="relevance")
+
+        return self._weigh_and_clip(recommendations, self.threshold)
+
+    def _get_enriched_recommendations(
+        self,
+        recommendations: DataFrameLike,
+        ground_truth: DataFrameLike,
+        max_k: int,
+        ground_truth_users: Optional[DataFrameLike] = None,
+    ) -> SparkDataFrame:
+        """
+        Merge recommendations and ground truth into a single DataFrame
+        and aggregate items into lists so that each user has only one record.
+
+        :param recommendations: recommendation list
+        :param ground_truth: test data
+        :param max_k: maximal k value to calculate the metric for.
+            `max_k` most relevant predictions are left for each user
+        :param ground_truth_users: list of users to consider in metric calculation.
+            if None, only the users from ground_truth are considered.
+        :return:  ``[user_idx, pred, ground_truth]``
+        """
+        recommendations = convert2spark(recommendations)
+        ground_truth = convert2spark(ground_truth)
+        ground_truth_users = convert2spark(ground_truth_users)
+
+        true_items_by_users = ground_truth.groupby("user_idx").agg(sf.collect_set("item_idx").alias("ground_truth"))
+
+        group_on = ["item_idx"]
+        if "user_idx" in self.prev_policy_weights.columns:
+            group_on.append("user_idx")
+        recommendations = get_top_k_recs(recommendations, k=max_k)
+
+        recommendations = recommendations.join(self.prev_policy_weights, on=group_on, how="left").na.fill(
+            0.0, subset=["prev_relevance"]
+        )
+
+        recommendations = self._reweighing(recommendations)
+
+        weight_type = recommendations.schema["weight"].dataType
+        item_type = ground_truth.schema["item_idx"].dataType
+
+        recommendations = filter_sort(recommendations, "weight")
+
+        if ground_truth_users is not None:
+            true_items_by_users = true_items_by_users.join(ground_truth_users, on="user_idx", how="right")
+
+        recommendations = recommendations.join(true_items_by_users, how="right", on=["user_idx"])
+        return fill_na_with_empty_array(
+            fill_na_with_empty_array(recommendations, "pred", item_type),
+            "weight",
+            weight_type,
+        )