HF.117 | 统计学约束手段在深度学习过程中的应用

import numpy as npimport pandas as pdfrom scipy.stats import pearsonrfrom sklearn.ensemble import RandomForestRegressorfrom sklearn.feature_selection import mutual_info_regressionimport matplotlib.pyplot as pltimport matplotlib.font_manager as fm
# 0. 数据生成函数def generate_mock_data(n_samples=500):      np.random.seed(42)      n_features = 50      X_data = np.random.rand(n_samples, n_features)      feature_names = [f'Clim_{i}' for i in range(n_features)]      df = pd.DataFrame(X_data, columns=feature_names)      df['Q_t1'] = np.random.rand(n_samples)      df['Q_t2'] = df['Q_t1'] * 0.8 + np.random.normal(0, 0.1, n_samples)      df['Q_t3'] = np.random.rand(n_samples)      y = (3 * df['Clim_0'] +               np.sin(10 * df['Clim_5']) +               2 * df['Q_t1'] +               np.random.normal(0, 0.2, n_samples))      df['target'] = y      return df
# 1. 准备数据df = generate_mock_data(n_samples=300)X = df.drop(columns=['target'])y = df['target']
# 2. 计算 PCCpcc_scores = {}for col in X.columns:      r, _ = pearsonr(X[col], y)      pcc_scores[col] = abs(r)
# 3. 计算MImi_values = mutual_info_regression(X, y, random_state=42)mi_scores = dict(zip(X.columns, mi_values))
# 4. 计算 VIMrf = RandomForestRegressor(n_estimators=100, random_state=42)rf.fit(X, y)vim_scores = dict(zip(X.columns, rf.feature_importances_))
# 5. 归一化def normalize_dict(d):      vals = list(d.values())      min_v, max_v = min(vals), max(vals)      if max_v - min_v == 0: return d      return {k: (v - min_v) / (max_v - min_v) for k, v in d.items()}
pcc_norm = normalize_dict(pcc_scores)mi_norm = normalize_dict(mi_scores)vim_norm = normalize_dict(vim_scores)
# 6. 选 Top-P 并进行 DS 融合p = 10def get_top_keys(d, n):      return [k for k, v in sorted(d.items(), key=lambda x: x[1], reverse=True)[:n]]  A1 = get_top_keys(pcc_norm, p)A2 = get_top_keys(mi_norm, p)A3 = get_top_keys(vim_norm, p)
all_preds = set(A1) | set(A2) | set(A3)ds_scores = {}
for pred in all_preds:      score = 0      if pred in A1: score += pcc_norm[pred]      if pred in A2: score += mi_norm[pred]      if pred in A3: score += vim_norm[pred]      ds_scores[pred] = score
# 排序ds_sorted = sorted(ds_scores.items(), key=lambda x: x[1], reverse=True)
# 打印结果供参考print("DS 综合排序结果 (前 10):")for name, score in ds_sorted[:10]:      print(f"


    
{name}: {score}")
# 绘图top_10 = ds_sorted[:10]names = [x[0] for x in top_10]scores = [x[1] for x in top_10]
# 倒序以便在水平条形图中，最高的在上面names.reverse()scores.reverse()plt.figure(figsize=(10, 6))plt.barh(names, scores, color='skyblue')plt.xlabel('DS Score (Normalized Sum)')plt.title('Top 10 Features by DS Fusion Score')plt.grid(axis='x', linestyle='--', alpha=0.7)plt.tight_layout()plt.show()