1{
2  "end_line": 112,
3  "filename": "custom_featsel.py",
4  "op": "replace",
5  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Ensemble of univariate methods to capture diverse relationships.\"\"\"\n    from scipy.stats import rankdata\n    from sklearn.tree import DecisionTreeClassifier\n    from sklearn.feature_selection import f_classif, chi2\n    import numpy as np\n\n    n_samples, n_features = X.shape\n    \n    # 1. Chi-squared (optimal for sparse/categorical dependencies)\n    c_scores, _ = chi2(X, y)\n    c_ranks = rankdata(np.nan_to_num(c_scores, nan=0.0))\n    \n    # 2. ANOVA F-value (optimal for linear mean separability)\n    f_scores, _ = f_classif(X, y)\n    f_ranks = rankdata(np.nan_to_num(f_scores, nan=0.0))\n    \n    # 3. Adaptive-Binning Mutual Information via single-feature Decision Trees\n    # Captures highly non-linear or multi-modal dependencies with optimal bins.\n    ig_scores = np.zeros(n_features)\n    clf = DecisionTreeClassifier(max_depth=4, criterion='entropy', random_state=42)\n    \n    for i in range(n_features):\n        X_col = np.ascontiguousarray(X[:, i:i+1])\n        clf.fit(X_col, y)\n        imp = clf.tree_.compute_feature_importances(normalize=False)\n        if len(imp) > 0:\n            ig_scores[i] = imp[0]\n            \n    ig_ranks = rankdata(ig_scores)\n    \n    # 4. Ensemble Voting: Uniform Borda count aggregation\n    # Creates an extremely robust metric across diverse data modalities.\n    return (c_ranks + f_ranks + ig_ranks) / n_features",
6  "start_line": 86
7}

1{
2  "start_line": 86,
3  "filename": "scikit-learn/custom_featsel.py",
4  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Ensemble of univariate methods to capture diverse relationships.\"\"\"\n    from scipy.stats import rankdata\n    from sklearn.tree import DecisionTreeClassifier\n    from sklearn.feature_selection import f_classif, chi2\n    import numpy as np\n\n    n_samples, n_features = X.shape\n    \n    # 1. Chi-squared (optimal for sparse/categorical dependencies)\n    c_scores, _ = chi2(X, y)\n    c_ranks = rankdata(np.nan_to_num(c_scores, nan=0.0))\n    \n    # 2. ANOVA F-value (optimal for linear mean separability)\n    f_scores, _ = f_classif(X, y)\n    f_ranks = rankdata(np.nan_to_num(f_scores, nan=0.0))\n    \n    # 3. Adaptive-Binning Mutual Information via single-feature Decision Trees\n    # Captures highly non-linear or multi-modal dependencies with optimal bins.\n    ig_scores = np.zeros(n_features)\n    clf = DecisionTreeClassifier(max_depth=4, criterion='entropy', random_state=42)\n    \n    for i in range(n_features):\n        X_col = np.ascontiguousarray(X[:, i:i+1])\n        clf.fit(X_col, y)\n        imp = clf.tree_.compute_feature_importances(normalize=False)\n        if len(imp) > 0:\n            ig_scores[i] = imp[0]\n            \n    ig_ranks = rankdata(ig_scores)\n    \n    # 4. Ensemble Voting: Uniform Borda count aggregation\n    # Creates an extremely robust metric across diverse data modalities.\n    return (c_ranks + f_ranks + ig_ranks) / n_features",
5  "end_line": 112,
6  "op": "replace"
7}

1{
2  "op": "replace",
3  "end_line": 119,
4  "start_line": 86,
5  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Robust unified feature scoring across diverse modalities (Text, Image, Tabular).\n    \n    Combines linear (ANOVA F), categorical/sparse (Chi-squared), and non-linear \n    dispersion metrics (ANOVA on median deviations), using a max-pool over \n    robustly standardized scores to auto-select the best metric per feature.\n    \"\"\"\n    import numpy as np\n    from sklearn.feature_selection import f_classif, chi2\n    \n    # 1. Linear differences in mean (optimal for continuous variables, Vision)\n    f_scores, _ = f_classif(X, y)\n    f_scores = np.nan_to_num(f_scores)\n    \n    # 2. Sparse / Categorical dependencies (optimal for Text/TF-IDF)\n    # X is guaranteed non-negative\n    c_scores, _ = chi2(X, y)\n    c_scores = np.nan_to_num(c_scores)\n    \n    # 3. Non-linear bimodal dependencies (optimal for XOR-like tabular data, Madelon)\n    # Measures if the variance / absolute deviation from median differs by class\n    X_median = np.median(X, axis=0)\n    X_dev = np.abs(X - X_median)\n    dev_scores, _ = f_classif(X_dev, y)\n    dev_scores = np.nan_to_num(dev_scores)\n    \n    # 4. Scale and combine using maximum\n    # By robustly scaling and taking the maximum, we let a feature shine if it \n    # has a strong signal in ANY of the modalities, while suppressing metrics \n    # that produce flat noise for that specific dataset.\n    def robust_scale(s):\n        med = np.median(s)\n        mad = np.median(np.abs(s - med))\n        if mad > 0:\n            return np.maximum(0, (s - med) / mad)\n        std = np.std(s)\n        if std > 0:\n            return np.maximum(0, (s - np.mean(s)) / std)\n        return s\n        \n    s_f = robust_scale(f_scores)\n    s_c = robust_scale(c_scores)\n    s_dev = robust_scale(dev_scores)\n    \n    combined_scores = np.maximum.reduce([s_f, s_c, s_dev])\n    return combined_scores",
6  "filename": "scikit-learn/custom_featsel.py"
7}

1{
2  "start_line": 86,
3  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Robust unified feature scoring using Signal-Amplified Consensus.\n    \n    Computes multiple complementary univariate metrics (Linear, Sparse, Scale, Rank)\n    and merges them using cubic signal amplification. This dynamically zeroes out\n    noise metrics on any given dataset while boosting the most discriminative metric.\n    \"\"\"\n    import numpy as np\n    from sklearn.feature_selection import f_classif, chi2\n    from scipy.stats import rankdata\n\n    # 1. Linear mean shifts (optimal for normal continuous variables like images)\n    f_scores, _ = f_classif(X, y)\n    \n    # 2. Sparse/categorical multinomial dependencies (optimal for text TF-IDF)\n    c_scores, _ = chi2(X, y)\n    \n    # 3. Variance/Magnitude shifts (captures bimodal/XOR interactions)\n    X_med = np.median(X, axis=0)\n    dev_scores, _ = f_classif(np.abs(X - X_med), y)\n    \n    # 4. Non-parametric rank shifts (Kruskal-Wallis style, robust to tabular outliers)\n    # Fast column-wise approximation of rankdata\n    X_ranks = np.argsort(np.argsort(X, axis=0), axis=0)\n    k_scores, _ = f_classif(X_ranks, y)\n    \n    scores_list = [f_scores, c_scores, dev_scores, k_scores]\n    fused_score = np.zeros(X.shape[1])\n    \n    def process_score(s):\n        s = np.nan_to_num(s)\n        med = np.median(s)\n        mad = np.median(np.abs(s - med))\n        \n        # Robust standardization to extract statistical significance (Z-score)\n        if mad > 1e-8:\n            z = (s - med) / (mad * 1.4826)\n        else:\n            std = np.std(s)\n            if std > 1e-8:\n                z = (s - np.mean(s)) / std\n            else:\n                z = np.zeros_like(s)\n                \n        # Keep only the upper tail of the statistic\n        z = np.maximum(0, z)\n        # Power transformation heavily amplifies the signal of confident metrics\n        # and suppresses metrics that only produce noisy, low-Z uniform scores.\n        return z ** 3\n\n    for s in scores_list:\n        fused_score += process_score(s)\n        \n    return fused_score",
4  "filename": "scikit-learn/custom_featsel.py",
5  "op": "replace",
6  "end_line": 131
7}

1{
2  "start_line": 86,
3  "filename": "scikit-learn/custom_featsel.py",
4  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Robust unified feature scoring using Signal-Amplified Consensus.\n    \n    Computes mutually complementary univariate metrics (Linear, Sparse, Scale)\n    and merges them using cubic signal amplification. This dynamically zeroes out\n    noise metrics on any given dataset while boosting the most discriminative metric.\n    \"\"\"\n    import numpy as np\n    from sklearn.feature_selection import f_classif, chi2\n\n    # 1. Linear mean shifts (optimal for normal continuous variables like images)\n    f_scores, _ = f_classif(X, y)\n    \n    # 2. Sparse/categorical multinomial dependencies (optimal for text TF-IDF)\n    c_scores, _ = chi2(X, y)\n    \n    # 3. Variance/Magnitude shifts (captures bimodal/XOR interactions like Madelon)\n    X_med = np.median(X, axis=0)\n    dev_scores, _ = f_classif(np.abs(X - X_med), y)\n    \n    scores_list = [f_scores, c_scores, dev_scores]\n    fused_score = np.zeros(X.shape[1])\n    \n    def process_score(s):\n        s = np.nan_to_num(s)\n        med = np.median(s)\n        mad = np.median(np.abs(s - med))\n        \n        # Robust standardization to extract statistical significance (Z-score)\n        if mad > 1e-8:\n            z = (s - med) / (mad * 1.4826)\n        else:\n            std = np.std(s)\n            if std > 1e-8:\n                z = (s - np.mean(s)) / std\n            else:\n                z = np.zeros_like(s)\n                \n        # Keep only the upper tail of the statistic\n        z = np.maximum(0, z)\n        # Power transformation heavily amplifies the signal of confident metrics\n        # and suppresses metrics that only produce noisy, low-Z uniform scores.\n        return z ** 3\n\n    for s in scores_list:\n        fused_score += process_score(s)\n        \n    return fused_score",
5  "op": "replace",
6  "end_line": 139
7}

1{
2  "content": "def score_features(X: np.ndarray, y: np.ndarray) -> np.ndarray:\n    \"\"\"Probe-Assisted Task-Aware Feature Selection.\n    \n    Generates candidate feature scores from complementary univariate statistics:\n      1. ANOVA F-value (Linear/Continuous)\n      2. Chi-Squared (Multinomial/Sparse Count)\n      3. Dispersion ANOVA (Non-linear XOR/Variance shifts)\n      \n    Conducts fast internal cross-validation using a proxy classifier to dynamically \n    route the dataset to the optimal statistical test, ensuring maximal downstream \n    performance regardless of the data modality.\n    \"\"\"\n    import numpy as np\n    from sklearn.feature_selection import f_classif, chi2\n    from sklearn.linear_model import LogisticRegression\n    from sklearn.preprocessing import StandardScaler\n    from sklearn.model_selection import StratifiedShuffleSplit\n\n    # 1. Linear differences in mean\n    f_scores, _ = f_classif(X, y)\n    \n    # 2. Sparse/categorical multinomial dependencies\n    c_scores, _ = chi2(X, y)\n    \n    # 3. Variance/Magnitude shifts (captures bimodal/XOR interactions)\n    X_med = np.median(X, axis=0)\n    dev_scores, _ = f_classif(np.abs(X - X_med), y)\n    \n    f_scores = np.nan_to_num(f_scores)\n    c_scores = np.nan_to_num(c_scores)\n    dev_scores = np.nan_to_num(dev_scores)\n    \n    candidates = [\n        (\"f_classif\", f_scores),\n        (\"chi2\", c_scores),\n        (\"dev_scores\", dev_scores)\n    ]\n    \n    n_features = X.shape[1]\n    # Set proxy K dynamically matching typical dataset scale heuristics\n    if n_features >= 10000:\n        proxy_k = 500\n    elif n_features >= 700:\n        proxy_k = 200\n    else:\n        proxy_k = 20\n    proxy_k = min(n_features, proxy_k)\n    \n    # Internal split for proxy validation\n    sss = StratifiedShuffleSplit(n_splits=1, test_size=0.2, random_state=42)\n    train_idx, val_idx = next(sss.split(X, y))\n    \n    X_t, y_t = X[train_idx], y[train_idx]\n    X_v, y_v = X[val_idx], y[val_idx]\n    \n    best_acc = -1.0\n    best_scores = f_scores\n    \n    for name, scores in candidates:\n        top_indices = np.argsort(scores)[::-1][:proxy_k]\n        \n        X_t_sel = X_t[:, top_indices]\n        X_v_sel = X_v[:, top_indices]\n        \n        scaler = StandardScaler()\n        X_t_sel = scaler.fit_transform(X_t_sel)\n        X_v_sel = scaler.transform(X_v_sel)\n        \n        clf = LogisticRegression(\n            max_iter=200, \n            solver='lbfgs', \n            multi_class='auto', \n            random_state=42\n        )\n        try:\n            clf.fit(X_t_sel, y_t)\n            acc = clf.score(X_v_sel, y_v)\n        except Exception:\n            acc = 0.0\n            \n        if acc > best_acc:\n            best_acc = acc\n            best_scores = scores\n            \n    return best_scores",
3  "filename": "scikit-learn/custom_featsel.py",
4  "op": "replace",
5  "start_line": 86,
6  "end_line": 133
7}