.. GENERATED FROM PYTHON SOURCE LINES 62-70 The number of claims (``ClaimNb``) is a positive integer that can be modeled as a Poisson distribution. It is then assumed to be the number of discrete events occurring with a constant rate in a given time interval (``Exposure``, in units of years). Here we want to model the frequency ``y = ClaimNb / Exposure`` conditionally on ``X`` via a (scaled) Poisson distribution, and use ``Exposure`` as ``sample_weight``. .. GENERATED FROM PYTHON SOURCE LINES 70-88 .. code-block:: default df["Frequency"] = df["ClaimNb"] / df["Exposure"] print("Average Frequency = {}" .format(np.average(df["Frequency"], weights=df["Exposure"]))) print("Fraction of exposure with zero claims = {0:.1%}" .format(df.loc[df["ClaimNb"] == 0, "Exposure"].sum() / df["Exposure"].sum())) fig, (ax0, ax1, ax2) = plt.subplots(ncols=3, figsize=(16, 4)) ax0.set_title("Number of claims") _ = df["ClaimNb"].hist(bins=30, log=True, ax=ax0) ax1.set_title("Exposure in years") _ = df["Exposure"].hist(bins=30, log=True, ax=ax1) ax2.set_title("Frequency (number of claims per year)") _ = df["Frequency"].hist(bins=30, log=True, ax=ax2) .. image:: /auto_examples/linear_model/images/sphx_glr_plot_poisson_regression_non_normal_loss_001.png :alt: Number of claims, Exposure in years, Frequency (number of claims per year) :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Average Frequency = 0.10070308464041304 Fraction of exposure with zero claims = 93.9% .. GENERATED FROM PYTHON SOURCE LINES 89-95 The remaining columns can be used to predict the frequency of claim events. Those columns are very heterogeneous with a mix of categorical and numeric variables with different scales, possibly very unevenly distributed. In order to fit linear models with those predictors it is therefore necessary to perform standard feature transformations as follows: .. GENERATED FROM PYTHON SOURCE LINES 95-121 .. code-block:: default from sklearn.pipeline import make_pipeline from sklearn.preprocessing import FunctionTransformer, OneHotEncoder from sklearn.preprocessing import StandardScaler, KBinsDiscretizer from sklearn.compose import ColumnTransformer log_scale_transformer = make_pipeline( FunctionTransformer(np.log, validate=False), StandardScaler() ) linear_model_preprocessor = ColumnTransformer( [ ("passthrough_numeric", "passthrough", ["BonusMalus"]), ("binned_numeric", KBinsDiscretizer(n_bins=10), ["VehAge", "DrivAge"]), ("log_scaled_numeric", log_scale_transformer, ["Density"]), ("onehot_categorical", OneHotEncoder(), ["VehBrand", "VehPower", "VehGas", "Region", "Area"]), ], remainder="drop", ) .. GENERATED FROM PYTHON SOURCE LINES 122-133 A constant prediction baseline ------------------------------ It is worth noting that more than 93% of policyholders have zero claims. If we were to convert this problem into a binary classification task, it would be significantly imbalanced, and even a simplistic model that would only predict mean can achieve an accuracy of 93%. To evaluate the pertinence of the used metrics, we will consider as a baseline a "dummy" estimator that constantly predicts the mean frequency of the training sample. .. GENERATED FROM PYTHON SOURCE LINES 133-147 .. code-block:: default from sklearn.dummy import DummyRegressor from sklearn.pipeline import Pipeline from sklearn.model_selection import train_test_split df_train, df_test = train_test_split(df, test_size=0.33, random_state=0) dummy = Pipeline([ ("preprocessor", linear_model_preprocessor), ("regressor", DummyRegressor(strategy='mean')), ]).fit(df_train, df_train["Frequency"], regressor__sample_weight=df_train["Exposure"]) .. GENERATED FROM PYTHON SOURCE LINES 148-150 Let's compute the performance of this constant prediction baseline with 3 different regression metrics: .. GENERATED FROM PYTHON SOURCE LINES 150-185 .. code-block:: default from sklearn.metrics import mean_squared_error from sklearn.metrics import mean_absolute_error from sklearn.metrics import mean_poisson_deviance def score_estimator(estimator, df_test): """Score an estimator on the test set.""" y_pred = estimator.predict(df_test) print("MSE: %.3f" % mean_squared_error(df_test["Frequency"], y_pred, sample_weight=df_test["Exposure"])) print("MAE: %.3f" % mean_absolute_error(df_test["Frequency"], y_pred, sample_weight=df_test["Exposure"])) # Ignore non-positive predictions, as they are invalid for # the Poisson deviance. mask = y_pred > 0 if (~mask).any(): n_masked, n_samples = (~mask).sum(), mask.shape[0] print(f"WARNING: Estimator yields invalid, non-positive predictions " f" for {n_masked} samples out of {n_samples}. These predictions " f"are ignored when computing the Poisson deviance.") print("mean Poisson deviance: %.3f" % mean_poisson_deviance(df_test["Frequency"][mask], y_pred[mask], sample_weight=df_test["Exposure"][mask])) print("Constant mean frequency evaluation:") score_estimator(dummy, df_test) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Constant mean frequency evaluation: MSE: 0.564 MAE: 0.189 mean Poisson deviance: 0.625 .. GENERATED FROM PYTHON SOURCE LINES 186-193 (Generalized) linear models --------------------------- We start by modeling the target variable with the (l2 penalized) least squares linear regression model, more comonly known as Ridge regression. We use a low penalization `alpha`, as we expect such a linear model to under-fit on such a large dataset. .. GENERATED FROM PYTHON SOURCE LINES 193-203 .. code-block:: default from sklearn.linear_model import Ridge ridge_glm = Pipeline([ ("preprocessor", linear_model_preprocessor), ("regressor", Ridge(alpha=1e-6)), ]).fit(df_train, df_train["Frequency"], regressor__sample_weight=df_train["Exposure"]) .. GENERATED FROM PYTHON SOURCE LINES 204-210 The Poisson deviance cannot be computed on non-positive values predicted by the model. For models that do return a few non-positive predictions (e.g. :class:`~sklearn.linear_model.Ridge`) we ignore the corresponding samples, meaning that the obtained Poisson deviance is approximate. An alternative approach could be to use :class:`~sklearn.compose.TransformedTargetRegressor` meta-estimator to map ``y_pred`` to a strictly positive domain. .. GENERATED FROM PYTHON SOURCE LINES 210-214 .. code-block:: default print("Ridge evaluation:") score_estimator(ridge_glm, df_test) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Ridge evaluation: MSE: 0.560 MAE: 0.187 WARNING: Estimator yields invalid, non-positive predictions for 600 samples out of 223745. These predictions are ignored when computing the Poisson deviance. mean Poisson deviance: 0.599 .. GENERATED FROM PYTHON SOURCE LINES 215-225 Next we fit the Poisson regressor on the target variable. We set the regularization strength ``alpha`` to approximately 1e-6 over number of samples (i.e. `1e-12`) in order to mimic the Ridge regressor whose L2 penalty term scales differently with the number of samples. Since the Poisson regressor internally models the log of the expected target value instead of the expected value directly (log vs identity link function), the relationship between X and y is not exactly linear anymore. Therefore the Poisson regressor is called a Generalized Linear Model (GLM) rather than a vanilla linear model as is the case for Ridge regression. .. GENERATED FROM PYTHON SOURCE LINES 225-240 .. code-block:: default from sklearn.linear_model import PoissonRegressor n_samples = df_train.shape[0] poisson_glm = Pipeline([ ("preprocessor", linear_model_preprocessor), ("regressor", PoissonRegressor(alpha=1e-12, max_iter=300)) ]) poisson_glm.fit(df_train, df_train["Frequency"], regressor__sample_weight=df_train["Exposure"]) print("PoissonRegressor evaluation:") score_estimator(poisson_glm, df_test) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none PoissonRegressor evaluation: MSE: 0.560 MAE: 0.186 mean Poisson deviance: 0.594 .. GENERATED FROM PYTHON SOURCE LINES 241-260 Gradient Boosting Regression Trees for Poisson regression --------------------------------------------------------- Finally, we will consider a non-linear model, namely Gradient Boosting Regression Trees. Tree-based models do not require the categorical data to be one-hot encoded: instead, we can encode each category label with an arbitrary integer using :class:`~sklearn.preprocessing.OrdinalEncoder`. With this encoding, the trees will treat the categorical features as ordered features, which might not be always a desired behavior. However this effect is limited for deep enough trees which are able to recover the categorical nature of the features. The main advantage of the :class:`~sklearn.preprocessing.OrdinalEncoder` over the :class:`~sklearn.preprocessing.OneHotEncoder` is that it will make training faster. Gradient Boosting also gives the possibility to fit the trees with a Poisson loss (with an implicit log-link function) instead of the default least-squares loss. Here we only fit trees with the Poisson loss to keep this example concise. .. GENERATED FROM PYTHON SOURCE LINES 260-285 .. code-block:: default from sklearn.ensemble import HistGradientBoostingRegressor from sklearn.preprocessing import OrdinalEncoder tree_preprocessor = ColumnTransformer( [ ("categorical", OrdinalEncoder(), ["VehBrand", "VehPower", "VehGas", "Region", "Area"]), ("numeric", "passthrough", ["VehAge", "DrivAge", "BonusMalus", "Density"]), ], remainder="drop", ) poisson_gbrt = Pipeline([ ("preprocessor", tree_preprocessor), ("regressor", HistGradientBoostingRegressor(loss="poisson", max_leaf_nodes=128)), ]) poisson_gbrt.fit(df_train, df_train["Frequency"], regressor__sample_weight=df_train["Exposure"]) print("Poisson Gradient Boosted Trees evaluation:") score_estimator(poisson_gbrt, df_test) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Poisson Gradient Boosted Trees evaluation: MSE: 0.563 MAE: 0.183 mean Poisson deviance: 0.574 .. GENERATED FROM PYTHON SOURCE LINES 286-297 Like the Poisson GLM above, the gradient boosted trees model minimizes the Poisson deviance. However, because of a higher predictive power, it reaches lower values of Poisson deviance. Evaluating models with a single train / test split is prone to random fluctuations. If computing resources allow, it should be verified that cross-validated performance metrics would lead to similar conclusions. The qualitative difference between these models can also be visualized by comparing the histogram of observed target values with that of predicted values: .. GENERATED FROM PYTHON SOURCE LINES 297-325 .. code-block:: default fig, axes = plt.subplots(nrows=2, ncols=4, figsize=(16, 6), sharey=True) fig.subplots_adjust(bottom=0.2) n_bins = 20 for row_idx, label, df in zip(range(2), ["train", "test"], [df_train, df_test]): df["Frequency"].hist(bins=np.linspace(-1, 30, n_bins), ax=axes[row_idx, 0]) axes[row_idx, 0].set_title("Data") axes[row_idx, 0].set_yscale('log') axes[row_idx, 0].set_xlabel("y (observed Frequency)") axes[row_idx, 0].set_ylim([1e1, 5e5]) axes[row_idx, 0].set_ylabel(label + " samples") for idx, model in enumerate([ridge_glm, poisson_glm, poisson_gbrt]): y_pred = model.predict(df) pd.Series(y_pred).hist(bins=np.linspace(-1, 4, n_bins), ax=axes[row_idx, idx+1]) axes[row_idx, idx + 1].set( title=model[-1].__class__.__name__, yscale='log', xlabel="y_pred (predicted expected Frequency)" ) plt.tight_layout() .. image:: /auto_examples/linear_model/images/sphx_glr_plot_poisson_regression_non_normal_loss_002.png :alt: Data, Ridge, PoissonRegressor, HistGradientBoostingRegressor, Data, Ridge, PoissonRegressor, HistGradientBoostingRegressor :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 326-360 The experimental data presents a long tail distribution for ``y``. In all models, we predict the expected frequency of a random variable, so we will have necessarily fewer extreme values than for the observed realizations of that random variable. This explains that the mode of the histograms of model predictions doesn't necessarily correspond to the smallest value. Additionally, the normal distribution used in ``Ridge`` has a constant variance, while for the Poisson distribution used in ``PoissonRegressor`` and ``HistGradientBoostingRegressor``, the variance is proportional to the predicted expected value. Thus, among the considered estimators, ``PoissonRegressor`` and ``HistGradientBoostingRegressor`` are a-priori better suited for modeling the long tail distribution of the non-negative data as compared to the ``Ridge`` model which makes a wrong assumption on the distribution of the target variable. The ``HistGradientBoostingRegressor`` estimator has the most flexibility and is able to predict higher expected values. Note that we could have used the least squares loss for the ``HistGradientBoostingRegressor`` model. This would wrongly assume a normal distributed response variable as does the `Ridge` model, and possibly also lead to slightly negative predictions. However the gradient boosted trees would still perform relatively well and in particular better than ``PoissonRegressor`` thanks to the flexibility of the trees combined with the large number of training samples. Evaluation of the calibration of predictions -------------------------------------------- To ensure that estimators yield reasonable predictions for different policyholder types, we can bin test samples according to ``y_pred`` returned by each model. Then for each bin, we compare the mean predicted ``y_pred``, with the mean observed target: .. GENERATED FROM PYTHON SOURCE LINES 360-437 .. code-block:: default from sklearn.utils import gen_even_slices def _mean_frequency_by_risk_group(y_true, y_pred, sample_weight=None, n_bins=100): """Compare predictions and observations for bins ordered by y_pred. We order the samples by ``y_pred`` and split it in bins. In each bin the observed mean is compared with the predicted mean. Parameters ---------- y_true: array-like of shape (n_samples,) Ground truth (correct) target values. y_pred: array-like of shape (n_samples,) Estimated target values. sample_weight : array-like of shape (n_samples,) Sample weights. n_bins: int Number of bins to use. Returns ------- bin_centers: ndarray of shape (n_bins,) bin centers y_true_bin: ndarray of shape (n_bins,) average y_pred for each bin y_pred_bin: ndarray of shape (n_bins,) average y_pred for each bin """ idx_sort = np.argsort(y_pred) bin_centers = np.arange(0, 1, 1/n_bins) + 0.5/n_bins y_pred_bin = np.zeros(n_bins) y_true_bin = np.zeros(n_bins) for n, sl in enumerate(gen_even_slices(len(y_true), n_bins)): weights = sample_weight[idx_sort][sl] y_pred_bin[n] = np.average( y_pred[idx_sort][sl], weights=weights ) y_true_bin[n] = np.average( y_true[idx_sort][sl], weights=weights ) return bin_centers, y_true_bin, y_pred_bin print(f"Actual number of claims: {df_test['ClaimNb'].sum()}") fig, ax = plt.subplots(nrows=2, ncols=2, figsize=(12, 8)) plt.subplots_adjust(wspace=0.3) for axi, model in zip(ax.ravel(), [ridge_glm, poisson_glm, poisson_gbrt, dummy]): y_pred = model.predict(df_test) y_true = df_test["Frequency"].values exposure = df_test["Exposure"].values q, y_true_seg, y_pred_seg = _mean_frequency_by_risk_group( y_true, y_pred, sample_weight=exposure, n_bins=10) # Name of the model after the estimator used in the last step of the # pipeline. print(f"Predicted number of claims by {model[-1]}: " f"{np.sum(y_pred * exposure):.1f}") axi.plot(q, y_pred_seg, marker='x', linestyle="--", label="predictions") axi.plot(q, y_true_seg, marker='o', linestyle="--", label="observations") axi.set_xlim(0, 1.0) axi.set_ylim(0, 0.5) axi.set( title=model[-1], xlabel='Fraction of samples sorted by y_pred', ylabel='Mean Frequency (y_pred)' ) axi.legend() plt.tight_layout() .. image:: /auto_examples/linear_model/images/sphx_glr_plot_poisson_regression_non_normal_loss_003.png :alt: Ridge(alpha=1e-06), PoissonRegressor(alpha=1e-12, max_iter=300), HistGradientBoostingRegressor(loss='poisson', max_leaf_nodes=128), DummyRegressor() :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Actual number of claims: 11935.0 Predicted number of claims by Ridge(alpha=1e-06): 11940.4 Predicted number of claims by PoissonRegressor(alpha=1e-12, max_iter=300): 11936.4 Predicted number of claims by HistGradientBoostingRegressor(loss='poisson', max_leaf_nodes=128): 12093.7 Predicted number of claims by DummyRegressor(): 11931.2 .. GENERATED FROM PYTHON SOURCE LINES 438-469 The dummy regression model predicts a constant frequency. This model does not attribute the same tied rank to all samples but is none-the-less globally well calibrated (to estimate the mean frequency of the entire population). The ``Ridge`` regression model can predict very low expected frequencies that do not match the data. It can therefore severely under-estimate the risk for some policyholders. ``PoissonRegressor`` and ``HistGradientBoostingRegressor`` show better consistency between predicted and observed targets, especially for low predicted target values. The sum of all predictions also confirms the calibration issue of the ``Ridge`` model: it under-estimates by more than 3% the total number of claims in the test set while the other three models can approximately recover the total number of claims of the test portfolio. Evaluation of the ranking power ------------------------------- For some business applications, we are interested in the ability of the model to rank the riskiest from the safest policyholders, irrespective of the absolute value of the prediction. In this case, the model evaluation would cast the problem as a ranking problem rather than a regression problem. To compare the 3 models from this perspective, one can plot the cumulative proportion of claims vs the cumulative proportion of exposure for the test samples order by the model predictions, from safest to riskiest according to each model. This plot is called a Lorenz curve and can be summarized by the Gini index: .. GENERATED FROM PYTHON SOURCE LINES 469-516 .. code-block:: default from sklearn.metrics import auc def lorenz_curve(y_true, y_pred, exposure): y_true, y_pred = np.asarray(y_true), np.asarray(y_pred) exposure = np.asarray(exposure) # order samples by increasing predicted risk: ranking = np.argsort(y_pred) ranked_frequencies = y_true[ranking] ranked_exposure = exposure[ranking] cumulated_claims = np.cumsum(ranked_frequencies * ranked_exposure) cumulated_claims /= cumulated_claims[-1] cumulated_exposure = np.cumsum(ranked_exposure) cumulated_exposure /= cumulated_exposure[-1] return cumulated_exposure, cumulated_claims fig, ax = plt.subplots(figsize=(8, 8)) for model in [dummy, ridge_glm, poisson_glm, poisson_gbrt]: y_pred = model.predict(df_test) cum_exposure, cum_claims = lorenz_curve(df_test["Frequency"], y_pred, df_test["Exposure"]) gini = 1 - 2 * auc(cum_exposure, cum_claims) label = "{} (Gini: {:.2f})".format(model[-1], gini) ax.plot(cum_exposure, cum_claims, linestyle="-", label=label) # Oracle model: y_pred == y_test cum_exposure, cum_claims = lorenz_curve(df_test["Frequency"], df_test["Frequency"], df_test["Exposure"]) gini = 1 - 2 * auc(cum_exposure, cum_claims) label = "Oracle (Gini: {:.2f})".format(gini) ax.plot(cum_exposure, cum_claims, linestyle="-.", color="gray", label=label) # Random Baseline ax.plot([0, 1], [0, 1], linestyle="--", color="black", label="Random baseline") ax.set( title="Lorenz curves by model", xlabel='Cumulative proportion of exposure (from safest to riskiest)', ylabel='Cumulative proportion of claims' ) ax.legend(loc="upper left") .. image:: /auto_examples/linear_model/images/sphx_glr_plot_poisson_regression_non_normal_loss_004.png :alt: Lorenz curves by model :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out Out: .. code-block:: none