Fix the error proportion in regression models

[SanGoku95]

Problem

The method self.error_tree.estimator_.tree_.value does not correctly return the error count when the model analyzed is a regression model. Instead, it provides normalized class proportions, which leads to incorrect calculations of n_errors for regression tasks in the get_error_leaf_summary function.

Reproduction of error

import numpy as np
import pandas as pd
from sklearn.datasets import fetch_california_housing
from sklearn.tree import DecisionTreeRegressor
from sklearn.model_selection import train_test_split
from mealy.error_analyzer import ErrorAnalyzer

# Fetch California Housing Dataset
housing = fetch_california_housing()
X = pd.DataFrame(housing.data, columns=housing.feature_names)
y_true = housing.target

# Train-Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y_true, test_size=0.3, random_state=42)

# Fit Primary Model
primary_model = DecisionTreeRegressor(max_depth=2, random_state=42)
primary_model.fit(X_train, y_train)

# Initialize ErrorAnalyzer
error_analyzer = ErrorAnalyzer(primary_model, feature_names=housing.feature_names, param_grid={"max_depth": [2]}, random_state=42)
error_analyzer.epsilon = 0.5
error_analyzer.fit(X_test, y_test)

# Validate Issue: Inspect a Node
leaf_id = 2
n_errors = error_analyzer.get_error_leaf_summary(leaf_id)[0]["n_errors"]

# Manual Verification
y_pred = primary_model.predict(X_test)
node_samples = error_analyzer.error_tree.estimator_.apply(X_test) == leaf_id
node_errors = (np.abs(y_test[node_samples] - y_pred[node_samples]) > error_analyzer.epsilon).sum()

# The assertion highlights the discrepancy
assert n_errors != node_errors  # n_errors from the library is incorrect for regression tasks

Root Cause

For regression tasks, self.error_tree.estimator_.tree_.value returns proportions, not absolute counts, for the class predictions. This behavior is appropriate for classification tasks but not for regression, leading to incorrect calculations of n_errors.

Solution

The fix involves correctly handling regression-specific logic by:

1. Multiplying the normalized proportion (self.error_tree.estimator_.tree_.value) by the total number of samples in the node (self.error_tree.estimator_.tree_.n_node_samples).
2. Ensuring consistent behavior for both regression and classification models.

[SanGoku95]

@du-phan @simonamaggio @AgatheG

[AgatheG]

@du-phan Could you merge this one?

[du-phan]

Unfortunately I don't have merge rights

[SanGoku95]

@simonamaggio

[AgatheG]

@instanceofme

[SanGoku95]

@instanceofme