validmind 2.4.10__py3-none-any.whl → 2.5.1__py3-none-any.whl

validmind/tests/model_validation/sklearn/PrecisionRecallCurve.py

@@ -65,7 +65,7 @@ class PrecisionRecallCurve(Metric):
             raise SkipTestError("Skipping PrecisionRecallCurve for Foundation models")
 
         y_true = self.inputs.dataset.y
-        y_pred = self.inputs.model.predict_proba(self.inputs.dataset.x)
+        y_pred = self.inputs.dataset.y_prob(self.inputs.model)
 
         # PR curve is only supported for binary classification
         if len(np.unique(y_true)) > 2:

validmind/tests/model_validation/sklearn/RobustnessDiagnosis.py

@@ -12,6 +12,7 @@ import pandas as pd
 import seaborn as sns
 from sklearn import metrics
 
+from validmind.errors import MissingOrInvalidModelPredictFnError
 from validmind.vm_models import (
     Figure,
     ResultSummary,
@@ -22,6 +23,7 @@ from validmind.vm_models import (
 )
 
 
+# TODO: make this support regression and classification as well as more performance metrics
 @dataclass
 class RobustnessDiagnosis(ThresholdTest):
     """
@@ -39,13 +41,13 @@ class RobustnessDiagnosis(ThresholdTest):
 
     This test is conducted by adding Gaussian noise, proportional to a particular standard deviation scale, to numeric
     input features of both the training and testing datasets. The model performance in the face of these perturbed
-    features is then evaluated using metrics (default: 'accuracy'). This process is iterated over a range of scale
-    factors. The resulting accuracy trend against the amount of noise introduced is illustrated with a line chart. A
-    predetermined threshold determines what level of accuracy decay due to perturbation is considered acceptable.
+    features is then evaluated using the ROC_AUC score. This process is iterated over a range of scale
+    factors. The resulting auc trend against the amount of noise introduced is illustrated with a line chart. A
+    predetermined threshold determines what level of auc decay due to perturbation is considered acceptable.
 
     **Signs of High Risk**:
-    - Substantial decreases in accuracy when noise is introduced to feature inputs.
-    - The decay in accuracy surpasses the configured threshold, indicating that the model is not robust against input
+    - Substantial decreases in auc when noise is introduced to feature inputs.
+    - The decay in auc surpasses the configured threshold, indicating that the model is not robust against input
     noise.
     - Instances where one or more elements provided in the features list don't match with the training dataset's
     numerical feature columns.
@@ -57,15 +59,12 @@ class RobustnessDiagnosis(ThresholdTest):
     - Detailed results visualization helps in interpreting the outcome of robustness testing.
 
     **Limitations**:
+    - The default threshold for auc decay is set to 0.05, which is unlikely to be optimal for most use cases and
+    should be adjusted based on domain expertise to suit the needs of the specific model.
     - Only numerical features are perturbed, leaving out non-numerical features, which can lead to an incomplete
     analysis of robustness.
-    - The default metric used is accuracy, which might not always give the best measure of a model's success,
-    particularly for imbalanced datasets.
     - The test is contingent on the assumption that the added Gaussian noise sufficiently represents potential data
     corruption or incompleteness in real-world scenarios.
-    - There might be a requirement to fine-tune the set decay threshold for accuracy with the help of domain knowledge
-    or specific project requisites.
-    - The robustness test might not deliver the expected results for datasets with a text column.
     """
 
     name = "robustness"
@@ -73,9 +72,9 @@ class RobustnessDiagnosis(ThresholdTest):
     default_params = {
         "features_columns": None,
         "scaling_factor_std_dev_list": [0.0, 0.1, 0.2, 0.3, 0.4, 0.5],
-        "accuracy_decay_threshold": 4,
+        "auc_decay_threshold": 0.05,
     }
-    tasks = ["classification", "text_classification"]
+    tasks = ["classification"]
     tags = [
         "sklearn",
         "binary_classification",
@@ -84,17 +83,15 @@ class RobustnessDiagnosis(ThresholdTest):
         "visualization",
     ]
 
-    default_metrics = {"accuracy": metrics.accuracy_score}
-
     def run(self):
         # Validate X std deviation parameter
         if "scaling_factor_std_dev_list" not in self.params:
             raise ValueError("scaling_factor_std_dev_list must be provided in params")
         x_std_dev_list = self.params["scaling_factor_std_dev_list"]
 
-        if self.params["accuracy_decay_threshold"] is None:
-            raise ValueError("accuracy_decay_threshold must be provided in params")
-        accuracy_threshold = self.params["accuracy_decay_threshold"]
+        if self.params["auc_decay_threshold"] is None:
+            raise ValueError("auc_decay_threshold must be provided in params")
+        auc_threshold = self.params["auc_decay_threshold"]
 
         if self.inputs.model is None:
             raise ValueError("model must of provided to run this test")
@@ -131,9 +128,7 @@ class RobustnessDiagnosis(ThresholdTest):
         test_results = []
         test_figures = []
 
-        results_headers = ["Perturbation Size", "Dataset Type", "Records"] + list(
-            self.default_metrics.keys()
-        )
+        results_headers = ["Perturbation Size", "Dataset Type", "Records", "AUC"]
         results = {k: [] for k in results_headers}
         # Iterate scaling factor for the standard deviation list
         for x_std_dev in x_std_dev_list:
@@ -159,32 +154,32 @@ class RobustnessDiagnosis(ThresholdTest):
         test_figures.append(
             Figure(
                 for_object=self,
-                key=f"{self.name}:accuracy",
+                key=f"{self.name}:auc",
                 figure=fig,
                 metadata={
-                    "metric": "accuracy",
+                    "metric": "AUC",
                     "features_list": features_list,
                 },
             )
         )
 
-        train_acc = df.loc[(df["Dataset Type"] == "Training"), "accuracy"].values[0]
-        test_acc = df.loc[(df["Dataset Type"] == "Test"), "accuracy"].values[0]
+        train_auc = df.loc[(df["Dataset Type"] == "Training"), "AUC"].values[0]
+        test_auc = df.loc[(df["Dataset Type"] == "Test"), "AUC"].values[0]
 
         df["Passed"] = np.where(
             (df["Dataset Type"] == "Training")
-            & (df["accuracy"] >= (train_acc - accuracy_threshold)),
+            & (df["AUC"] >= (train_auc - auc_threshold)),
             True,
             np.where(
                 (df["Dataset Type"] == "Test")
-                & (df["accuracy"] >= (test_acc - accuracy_threshold)),
+                & (df["AUC"] >= (test_auc - auc_threshold)),
                 True,
                 False,
             ),
         )
         test_results.append(
             ThresholdTestResult(
-                test_name="accuracy",
+                test_name="AUC",
                 column=features_list,
                 passed=True,
                 values={"records": df.to_dict("records")},
@@ -194,7 +189,7 @@ class RobustnessDiagnosis(ThresholdTest):
             test_results, passed=df["Passed"].all(), figures=test_figures
         )
 
-    def summary(self, results: List[ThresholdTestResult], all_passed: bool):
+    def summary(self, results: List[ThresholdTestResult], _):
         results_table = [
             record for result in results for record in result.values["records"]
         ]
@@ -229,9 +224,13 @@ class RobustnessDiagnosis(ThresholdTest):
         results["Dataset Type"].append(dataset_type)
         results["Perturbation Size"].append(x_std_dev)
         results["Records"].append(df.shape[0])
-        y_prediction = self.inputs.model.predict(df)
-        for metric, metric_fn in self.default_metrics.items():
-            results[metric].append(metric_fn(y_true, y_prediction) * 100)
+
+        try:
+            y_proba = self.inputs.model.predict_proba(df)
+        except MissingOrInvalidModelPredictFnError:
+            y_proba = self.inputs.model.predict(df)
+
+        results["AUC"].append(metrics.roc_auc_score(y_true, y_proba))
 
     def _add_noise_std_dev(
         self, values: List[float], x_std_dev: float
@@ -256,14 +255,14 @@ class RobustnessDiagnosis(ThresholdTest):
 
     def _plot_robustness(self, results: dict, features_columns: List[str]):
         """
-        Plots the model's accuracy under feature perturbations.
+        Plots the model's auc under feature perturbations.
         Args:
             results (dict): A dictionary containing the results of the evaluation.
                 It has the following keys:
                     - 'Dataset Type': the type of dataset evaluated, e.g. 'Training' or 'Test'.
                     - 'Perturbation Size': the size of the perturbation applied to the features.
                     - 'Records': the number of records evaluated.
-                    - Any other metric used for evaluation as keys, e.g. 'accuracy', 'precision', 'recall'.
+                    - 'auc': the ROC AUC score obtained for the evaluation.
                 The values of each key are lists containing the results for each evaluation.
             features_columns (list[str]): A list containing the names of the features perturbed.
         Returns:
@@ -277,7 +276,7 @@ class RobustnessDiagnosis(ThresholdTest):
         sns.lineplot(
             data=df,
             x="Perturbation Size",
-            y="accuracy",
+            y="AUC",
             hue="Dataset Type",
             style="Dataset Type",
             linewidth=3,
@@ -288,7 +287,7 @@ class RobustnessDiagnosis(ThresholdTest):
             ax=ax,
         )
         ax.tick_params(axis="x")
-        ax.set_ylabel("Accuracy", weight="bold", fontsize=18)
+        ax.set_ylabel("AUC", weight="bold", fontsize=18)
         ax.legend(fontsize=18)
         ax.set_xlabel(
             "Perturbation Size (X * Standard Deviation)", weight="bold", fontsize=18
@@ -321,9 +320,9 @@ class RobustnessDiagnosis(ThresholdTest):
             assert isinstance(test_result.values, dict)
             assert "records" in test_result.values
 
-            # For unperturbed training dataset, accuracy should be present
+            # For unperturbed training dataset, auc should be present
             if (
                 test_result.column == self.params["features_columns"]
                 and 0.0 in test_result.values["records"][0]["Perturbation Size"]
             ):
-                assert "accuracy" in test_result.values["records"][0]
+                assert "AUC" in test_result.values["records"][0]

validmind/tests/ongoing_monitoring/FeatureDrift.py

@@ -0,0 +1,182 @@
+# Copyright © 2023-2024 ValidMind Inc. All rights reserved.
+# See the LICENSE file in the root of this repository for details.
+# SPDX-License-Identifier: AGPL-3.0 AND ValidMind Commercial
+
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+
+from validmind import tags, tasks
+
+
+@tags("visualization")
+@tasks("monitoring")
+def FeatureDrift(
+    datasets, bins=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], feature_columns=None
+):
+    """
+    **Purpose**:
+
+    The Feature Drift test aims to evaluate how much the distribution of features has shifted over time between two
+    datasets, typically training and monitoring datasets. It uses the Population Stability Index (PSI) to quantify this
+    change, providing insights into the model's robustness and the necessity for retraining or feature engineering.
+
+    **Test Mechanism**:
+
+    This test calculates the PSI by:
+    - Bucketing the distributions of each feature in both datasets.
+    - Comparing the percentage of observations in each bucket between the two datasets.
+    - Aggregating the differences across all buckets for each feature to produce the PSI score for that feature.
+
+    The PSI score is interpreted as:
+    - PSI < 0.1: No significant population change.
+    - PSI < 0.2: Moderate population change.
+    - PSI >= 0.2: Significant population change.
+
+    **Signs of High Risk**:
+
+    - PSI >= 0.2 for any feature, indicating a significant distribution shift.
+    - Consistently high PSI scores across multiple features.
+    - Sudden spikes in PSI in recent monitoring data compared to historical data.
+
+    **Strengths**:
+
+    - Provides a quantitative measure of feature distribution changes.
+    - Easily interpretable thresholds for decision-making.
+    - Helps in early detection of data drift, prompting timely interventions.
+
+    **Limitations**:
+
+    - May not capture more intricate changes in data distribution nuances.
+    - Assumes that bucket thresholds (quantiles) adequately represent distribution shifts.
+    - PSI score interpretation can be overly simplistic for complex datasets.
+    """
+
+    # Feature columns for both datasets should be the same if not given
+    default_feature_columns = datasets[0].feature_columns
+    feature_columns = feature_columns or default_feature_columns
+
+    x_train_df = datasets[0].x_df()
+    x_test_df = datasets[1].x_df()
+
+    quantiles_train = x_train_df[feature_columns].quantile(
+        bins, method="single", interpolation="nearest"
+    )
+    PSI_QUANTILES = quantiles_train.to_dict()
+
+    PSI_BUCKET_FRAC, col, n = get_psi_buckets(
+        x_test_df, x_train_df, feature_columns, bins, PSI_QUANTILES
+    )
+
+    def nest(d: dict) -> dict:
+        result = {}
+        for key, value in d.items():
+            target = result
+            for k in key[:-1]:  # traverse all keys but the last
+                target = target.setdefault(k, {})
+            target[key[-1]] = value
+        return result
+
+    PSI_BUCKET_FRAC = nest(PSI_BUCKET_FRAC)
+
+    PSI_SCORES = {}
+    for col in feature_columns:
+        psi = 0
+        for n in bins:
+            actual = PSI_BUCKET_FRAC["test"][col][n]
+            expected = PSI_BUCKET_FRAC["train"][col][n]
+            psi_of_bucket = (actual - expected) * np.log(
+                (actual + 1e-6) / (expected + 1e-6)
+            )
+            psi += psi_of_bucket
+        PSI_SCORES[col] = psi
+
+    psi_df = pd.DataFrame(list(PSI_SCORES.items()), columns=["Features", "PSI Score"])
+
+    psi_df.sort_values(by=["PSI Score"], inplace=True, ascending=False)
+
+    psi_table = [
+        {"Features": values["Features"], "PSI Score": values["PSI Score"]}
+        for i, values in enumerate(psi_df.to_dict(orient="records"))
+    ]
+
+    save_fig = plot_hist(PSI_BUCKET_FRAC, bins)
+
+    final_psi = pd.DataFrame(psi_table)
+
+    return (final_psi, *save_fig)
+
+
+def get_psi_buckets(x_test_df, x_train_df, feature_columns, bins, PSI_QUANTILES):
+    DATA = {"test": x_test_df, "train": x_train_df}
+    PSI_BUCKET_FRAC = {}
+    for table in DATA.keys():
+        total_count = DATA[table].shape[0]
+        for col in feature_columns:
+            count_sum = 0
+            for n in bins:
+                if n == 0:
+                    bucket_count = (DATA[table][col] < PSI_QUANTILES[col][n]).sum()
+                elif n < 9:
+                    bucket_count = (
+                        total_count
+                        - count_sum
+                        - ((DATA[table][col] >= PSI_QUANTILES[col][n]).sum())
+                    )
+                elif n == 9:
+                    bucket_count = total_count - count_sum
+                count_sum += bucket_count
+                PSI_BUCKET_FRAC[table, col, n] = bucket_count / total_count
+    return PSI_BUCKET_FRAC, col, n
+
+
+def plot_hist(PSI_BUCKET_FRAC, bins):
+    bin_table_psi = pd.DataFrame(PSI_BUCKET_FRAC)
+    save_fig = []
+    for i in range(len(bin_table_psi)):
+
+        x = pd.DataFrame(
+            bin_table_psi.iloc[i]["test"].items(),
+            columns=["Bin", "Population % Reference"],
+        )
+        y = pd.DataFrame(
+            bin_table_psi.iloc[i]["train"].items(),
+            columns=["Bin", "Population % Monitoring"],
+        )
+        xy = x.merge(y, on="Bin")
+        xy.index = xy["Bin"]
+        xy = xy.drop(columns="Bin", axis=1)
+        feature_name = bin_table_psi.index[i]
+
+        n = len(bins)
+        r = np.arange(n)
+        width = 0.25
+
+        fig = plt.figure()
+
+        plt.bar(
+            r,
+            xy["Population % Reference"],
+            color="b",
+            width=width,
+            edgecolor="black",
+            label="Reference {0}".format(feature_name),
+        )
+        plt.bar(
+            r + width,
+            xy["Population % Monitoring"],
+            color="g",
+            width=width,
+            edgecolor="black",
+            label="Monitoring {0}".format(feature_name),
+        )
+
+        plt.xlabel("Bin")
+        plt.ylabel("Population %")
+        plt.title("Histogram of Population Differences {0}".format(feature_name))
+        plt.legend()
+        plt.tight_layout()
+        plt.close()
+        save_fig.append(fig)
+    return save_fig

validmind/tests/ongoing_monitoring/PredictionAcrossEachFeature.py

@@ -0,0 +1,76 @@
+# Copyright © 2023-2024 ValidMind Inc. All rights reserved.
+# See the LICENSE file in the root of this repository for details.
+# SPDX-License-Identifier: AGPL-3.0 AND ValidMind Commercial
+
+
+import matplotlib.pyplot as plt
+
+from validmind import tags, tasks
+
+
+@tags("visualization")
+@tasks("monitoring")
+def PredictionAcrossEachFeature(datasets, model):
+    """
+    **Purpose:**
+    This test shows visually the prediction using reference data and monitoring data across each individual feature. If
+    there are significant differences in predictions across feature values from reference to monitoring dataset, then
+    further investigation is needed as the model is producing predictions that are different than what was observed
+    during the training of the model.
+
+    **Test Mechanism:**
+    The test creates scatter plots for each feature, comparing the reference dataset (used for training) with the
+    monitoring dataset (used in production). Each plot has two subplots: one for the reference data and one for the
+    monitoring data, visualizing the prediction probabilities. This allows for a visual comparison of the model's
+    behavior across different datasets.
+
+    **Signs of High Risk:**
+    - Significant discrepancies between the reference and monitoring subplots for the same feature
+    - Unexpected patterns or trends in monitoring data that weren't present in reference data
+
+    **Strengths:**
+    - Provides a clear visual representation of model performance across different features
+    - Allows for easy identification of features where the model's predictions have changed
+    - Facilitates quick detection of potential issues with the model when deployed in production
+
+    **Limitations:**
+    - Interpretation of scatter plots can be subjective and may require expertise
+    - Visualizations do not provide quantitative metrics for objective evaluation
+    - May not capture all types of distribution changes or issues with the model's predictions
+    """
+
+    """
+    This test shows visually the prediction using reference data and monitoring data
+    across each individual feature. If there are significant differences in predictions
+    across feature values from reference to monitoring dataset then futher investigation
+    is needed as the model is producing predictions that are different then what was
+    observed during the training of the model.
+    """
+
+    df_reference = datasets[0]._df
+    df_monitoring = datasets[1]._df
+
+    figures_to_save = []
+    for column in df_reference:
+        prediction_prob_column = f"{model.input_id}_probabilities"
+        prediction_column = f"{model.input_id}_prediction"
+        if column == prediction_prob_column or column == prediction_column:
+            pass
+        else:
+            fig, axs = plt.subplots(1, 2, figsize=(20, 10), sharey="row")
+
+            ax1, ax2 = axs
+
+            ax1.scatter(df_reference[column], df_reference[prediction_prob_column])
+            ax2.scatter(df_monitoring[column], df_monitoring[prediction_prob_column])
+
+            ax1.set_title("Reference")
+            ax1.set_xlabel(column)
+            ax1.set_ylabel("Prediction Value")
+
+            ax2.set_title("Monitoring")
+            ax2.set_xlabel(column)
+            figures_to_save.append(fig)
+            plt.close()
+
+    return tuple(figures_to_save)

validmind/tests/ongoing_monitoring/PredictionCorrelation.py

@@ -0,0 +1,91 @@
+# Copyright © 2023-2024 ValidMind Inc. All rights reserved.
+# See the LICENSE file in the root of this repository for details.
+# SPDX-License-Identifier: AGPL-3.0 AND ValidMind Commercial
+
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+from validmind import tags, tasks
+
+
+@tags("visualization")
+@tasks("monitoring")
+def PredictionCorrelation(datasets, model):
+    """
+    **Purpose:**
+    The test is used to assess the correlation pairs for each feature between model predictions from reference and
+    monitoring datasets. The primary goal is to detect significant changes in these pairs, which may signal target
+    drift, leading to lower model performance.
+
+    **Test Mechanism:**
+    The test calculates the correlation of each feature with model predictions for both reference and monitoring
+    datasets. The test then compares these correlations side-by-side via a bar plot and a correlation table. Features
+    with significant changes in correlation pairs highlight potential risks of model drift.
+
+    **Signs of High Risk:**
+    - Significant changes in correlation pairs between the reference and monitoring predictions.
+    - Notable correlation differences indicating a potential shift in the relationship between features and the target
+    variable.
+
+    **Strengths:**
+    - Allows for visual identification of drift in feature relationships with model predictions.
+    - Comparison via a clear bar plot assists in understanding model stability over time.
+    - Helps in early detection of target drift, enabling timely interventions.
+
+    **Limitations:**
+    - May require substantial reference and monitoring data for accurate comparison.
+    - Correlation does not imply causation, and other factors might influence changes.
+    - The method solely focuses on linear relationships, potentially missing non-linear interactions.
+    """
+
+    prediction_prob_column = f"{model.input_id}_probabilities"
+    prediction_column = f"{model.input_id}_prediction"
+
+    df_corr = datasets[0]._df.corr()
+    df_corr = df_corr[[prediction_prob_column]]
+
+    df_corr2 = datasets[1]._df.corr()
+    df_corr2 = df_corr2[[prediction_prob_column]]
+
+    corr_final = df_corr.merge(df_corr2, left_index=True, right_index=True)
+    corr_final.columns = ["Reference Predictions", "Monitoring Predictions"]
+    corr_final = corr_final.drop(index=[prediction_column, prediction_prob_column])
+
+    n = len(corr_final)
+    r = np.arange(n)
+    width = 0.25
+
+    fig = plt.figure()
+
+    plt.bar(
+        r,
+        corr_final["Reference Predictions"],
+        color="b",
+        width=width,
+        edgecolor="black",
+        label="Reference Prediction Correlation",
+    )
+    plt.bar(
+        r + width,
+        corr_final["Monitoring Predictions"],
+        color="g",
+        width=width,
+        edgecolor="black",
+        label="Monitoring Prediction Correlation",
+    )
+
+    plt.xlabel("Features")
+    plt.ylabel("Correlation")
+    plt.title("Correlation between Predictions and Features")
+
+    features = corr_final.index.to_list()
+    plt.xticks(r + width / 2, features, rotation=45)
+    plt.legend()
+    plt.tight_layout()
+
+    corr_final["Features"] = corr_final.index
+    corr_final = corr_final[
+        ["Features", "Reference Predictions", "Monitoring Predictions"]
+    ]
+    return ({"Correlation Pair Table": corr_final}, fig)

validmind/tests/ongoing_monitoring/TargetPredictionDistributionPlot.py

@@ -0,0 +1,57 @@
+# Copyright © 2023-2024 ValidMind Inc. All rights reserved.
+# See the LICENSE file in the root of this repository for details.
+# SPDX-License-Identifier: AGPL-3.0 AND ValidMind Commercial
+
+import matplotlib.pyplot as plt
+import seaborn as sns
+
+from validmind import tags, tasks
+
+
+@tags("visualization")
+@tasks("monitoring")
+def TargetPredictionDistributionPlot(datasets, model):
+    """
+    **Purpose:**
+    This test provides the prediction distributions from the reference dataset and the new monitoring dataset. If there
+    are significant differences in the distributions, it might indicate different underlying data characteristics that
+    warrant further investigation into the root causes.
+
+    **Test Mechanism:**
+    The methodology involves generating Kernel Density Estimation (KDE) plots for the prediction probabilities from
+    both the reference and monitoring datasets. By comparing these KDE plots, one can visually assess any significant
+    differences in the prediction distributions between the two datasets.
+
+    **Signs of High Risk:**
+    - Significant divergence between the distribution curves of the reference and monitoring predictions
+    - Unusual shifts or bimodal distribution in the monitoring predictions compared to the reference predictions
+
+    **Strengths:**
+    - Visual representation makes it easy to spot differences in prediction distributions
+    - Useful for identifying potential data drift or changes in underlying data characteristics
+    - Simple and efficient to implement using standard plotting libraries
+
+    **Limitations:**
+    - Subjective interpretation of the visual plots
+    - Might not pinpoint the exact cause of distribution changes
+    - Less effective if the differences in distributions are subtle and not easily visible
+    """
+
+    pred_ref = datasets[0].y_prob_df(model)
+    pred_ref.columns = ["Reference Prediction"]
+    pred_monitor = datasets[1].y_prob_df(model)
+    pred_monitor.columns = ["Monitoring Prediction"]
+
+    fig = plt.figure()
+    plot = sns.kdeplot(
+        pred_ref["Reference Prediction"], shade=True, label="Reference Prediction"
+    )
+    plot = sns.kdeplot(
+        pred_monitor["Monitoring Prediction"], shade=True, label="Monitor Prediction"
+    )
+    plot.set(
+        xlabel="Prediction", title="Distribution of Reference & Monitor Predictions"
+    )
+    plot.legend()
+
+    return fig