1{
2  "op": "replace",
3  "filename": "scikit-learn/custom_featsel.py",
4  "start_line": 86,
5  "end_line": 112,
6  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Score each feature for its relevance to the classification target.\n\n    This function receives the training data and must return a score for\n    each feature indicating its importance. Higher scores = more important.\n    The top-k features (by score) will be selected for classification.\n\n    Args:\n        X: Training feature matrix of shape (n_samples, n_features).\n           All values are non-negative floats (TF-IDF, pixel intensities,\n           or pre-processed numeric features depending on the dataset).\n        y: Training labels of shape (n_samples,), integer class labels.\n\n    Returns:\n        scores: 1-D numpy array of shape (n_features,) with non-negative\n                importance scores. Features with higher scores are selected.\n\n    Notes:\n        - X values are guaranteed non-negative (suitable for chi2, etc.)\n        - Must handle both binary and multi-class problems.\n        - Must handle high-dimensional data (up to 10000 features).\n        - Returned scores must be finite (no NaN or Inf).\n        - Available imports: numpy, scipy, sklearn (see top of file).\n    \"\"\"\n    # Compute F-statistic (linear separability)\n    f_scores, _ = f_classif(X, y)\n    f_scores = np.nan_to_num(f_scores, nan=0.0)\n    \n    # Compute mutual information (non-linear dependency)\n    mi_scores = mutual_info_classif(X, y, discrete_features=False, random_state=42, n_neighbors=5)\n    mi_scores = np.nan_to_num(mi_scores, nan=0.0)\n    \n    # Normalize each to [0,1] range (avoid division by zero)\n    f_max = f_scores.max()\n    if f_max > 0:\n        f_scores = f_scores / f_max\n    mi_max = mi_scores.max()\n    if mi_max > 0:\n        mi_scores = mi_scores / mi_max\n    \n    # Combine using geometric mean (captures both linear and non-linear)\n    # Add small epsilon to avoid zero scores\n    eps = 1e-12\n    scores = np.sqrt((f_scores + eps) * (mi_scores + eps))\n    \n    return scores"
7}

1{
2  "op": "replace",
3  "filename": "scikit-learn/custom_featsel.py",
4  "start_line": 86,
5  "end_line": 131,
6  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Score each feature for its relevance to the classification target.\n\n    This function receives the training data and must return a score for\n    each feature indicating its importance. Higher scores = more important.\n    The top-k features (by score) will be selected for classification.\n\n    Args:\n        X: Training feature matrix of shape (n_samples, n_features).\n           All values are non-negative floats (TF-IDF, pixel intensities,\n           or pre-processed numeric features depending on the dataset).\n        y: Training labels of shape (n_samples,), integer class labels.\n\n    Returns:\n        scores: 1-D numpy array of shape (n_features,) with non-negative\n                importance scores. Features with higher scores are selected.\n\n    Notes:\n        - X values are guaranteed non-negative (suitable for chi2, etc.)\n        - Must handle both binary and multi-class problems.\n        - Must handle high-dimensional data (up to 10000 features).\n        - Returned scores must be finite (no NaN or Inf).\n        - Available imports: numpy, scipy, sklearn (see top of file).\n    \"\"\"\n    # Compute three different univariate scores\n    # Chi-squared statistic (good for count-like data)\n    chi2_scores, _ = chi2(X, y)\n    chi2_scores = np.nan_to_num(chi2_scores, nan=0.0)\n    \n    # ANOVA F-statistic (good for continuous linear relationships)\n    f_scores, _ = f_classif(X, y)\n    f_scores = np.nan_to_num(f_scores, nan=0.0)\n    \n    # Mutual information (good for non-linear dependencies)\n    mi_scores = mutual_info_classif(X, y, discrete_features=False, random_state=42, n_neighbors=5)\n    mi_scores = np.nan_to_num(mi_scores, nan=0.0)\n    \n    # Normalize each to [0,1] range (avoid division by zero)\n    def normalize(arr):\n        arr = arr.copy()\n        max_val = arr.max()\n        if max_val > 0:\n            arr = arr / max_val\n        return arr\n    \n    chi2_norm = normalize(chi2_scores)\n    f_norm = normalize(f_scores)\n    mi_norm = normalize(mi_scores)\n    \n    # Take the maximum of the three normalized scores per feature\n    # This adaptively picks the best signal for each feature\n    scores = np.maximum.reduce([chi2_norm, f_norm, mi_norm])\n    \n    # Ensure non-negative (should already be)\n    scores = np.maximum(scores, 0.0)\n    \n    return scores"
7}

1{
2  "op": "replace",
3  "filename": "scikit-learn/custom_featsel.py",
4  "start_line": 86,
5  "end_line": 142,
6  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Score each feature for its relevance to the classification target.\n\n    This function receives the training data and must return a score for\n    each feature indicating its importance. Higher scores = more important.\n    The top-k features (by score) will be selected for classification.\n\n    Args:\n        X: Training feature matrix of shape (n_samples, n_features).\n           All values are non-negative floats (TF-IDF, pixel intensities,\n           or pre-processed numeric features depending on the dataset).\n        y: Training labels of shape (n_samples,), integer class labels.\n\n    Returns:\n        scores: 1-D numpy array of shape (n_features,) with non-negative\n                importance scores. Features with higher scores are selected.\n\n    Notes:\n        - X values are guaranteed non-negative (suitable for chi2, etc.)\n        - Must handle both binary and multi-class problems.\n        - Must handle high-dimensional data (up to 10000 features).\n        - Returned scores must be finite (no NaN or Inf).\n        - Available imports: numpy, scipy, sklearn (see top of file).\n    \"\"\"\n    # Compute three different univariate scores\n    chi2_scores, _ = chi2(X, y)\n    chi2_scores = np.nan_to_num(chi2_scores, nan=0.0)\n    \n    f_scores, _ = f_classif(X, y)\n    f_scores = np.nan_to_num(f_scores, nan=0.0)\n    \n    mi_scores = mutual_info_classif(X, y, discrete_features=False, random_state=42, n_neighbors=5)\n    mi_scores = np.nan_to_num(mi_scores, nan=0.0)\n    \n    # Normalize each to [0,1] range\n    def normalize(arr):\n        arr = arr.copy()\n        max_val = arr.max()\n        if max_val > 0:\n            arr = arr / max_val\n        return arr\n    \n    chi2_norm = normalize(chi2_scores)\n    f_norm = normalize(f_scores)\n    mi_norm = normalize(mi_scores)\n    \n    # Compute sparsity of each feature (fraction of zero values)\n    sparsity = (X == 0).mean(axis=0)\n    \n    # Adaptive weighting: for sparse features (text), trust chi2 more;\n    # for dense features (continuous), trust f_classif more;\n    # mutual information is always valuable but weighted by non-linearity.\n    # Estimate linearity via correlation between chi2 and f scores?\n    # Simple: weight chi2 by sparsity, f by (1 - sparsity), mi by 0.5 constant.\n    w_chi2 = sparsity\n    w_f = 1.0 - sparsity\n    w_mi = 0.5  # constant contribution\n    \n    # Combine weighted scores\n    scores = w_chi2 * chi2_norm + w_f * f_norm + w_mi * mi_norm\n    \n    # Ensure non-negative\n    scores = np.maximum(scores, 0.0)\n    \n    return scores"
7}

1{
2  "n": 3
3}