python matplotlib sklearn 数据分析，主要是引入了shap ，可以对于影响因素全面分析

# 文件名: predict.py (已升级，处理实时文本向量化) import pickle import sys import numpy as np import pandas as pd import shap from sentence_transformers import SentenceTransformer MODEL_FILE = 'model_v3.pkl' SCALER_FILE = 'scaler_v3.pkl' LANGUAGE_MODEL = 'bge-base-zh-v1.5' # 在全局加载模型，避免每次调用都重复加载 try: with open(MODEL_FILE, 'rb') as f_model: MODEL = pickle.load(f_model) with open(SCALER_FILE, 'rb') as f_scaler: SCALER = pickle.load(f_scaler) LANG_MODEL = SentenceTransformer(LANGUAGE_MODEL) print("信息：模型、Scaler和语言模型已成功加载。") except FileNotFoundError: print(f"错误：找不到 '{MODEL_FILE}' 或 '{SCALER_FILE}'。请先运行步骤1, 2, 3。") sys.exit(1) def predict_new_video(title, duration, skip_rate, shot_count, has_subtitle): # --- 1. 处理数值特征 --- avg_shot_duration = duration / shot_count if shot_count > 0 else duration numerical_features_cols = ['duration', 's3_skip_rate', 'shot_count', 'has_subtitle', 'avg_shot_duration'] numerical_data = np.array([[duration, skip_rate, shot_count, has_subtitle, avg_shot_duration]]) numerical_data_scaled = SCALER.transform(numerical_data) df_numerical = pd.DataFrame(numerical_data_scaled, columns=numerical_features_cols) # --- 2. 处理文本特征 --- title_embedding = LANG_MODEL.encode([title]) embedding_cols = [f'title_vec_{i}' for i in range(title_embedding.shape[1])] df_embedding = pd.DataFrame(title_embedding, columns=embedding_cols) # --- 3. 合并特征并预测 --- input_df = pd.concat([df_numerical, df_embedding], axis=1) predicted_ratio = MODEL.predict(input_df)[0] prediction_text = f"📈 预测平均完播率：{predicted_ratio:.1%}" # --- 4. SHAP 归因分析 --- explainer = shap.TreeExplainer(MODEL) shap_values = explainer.shap_values(input_df) base_value = explainer.expected_value if isinstance(base_value, np.ndarray): base_value = base_value[0] # 由于特征太多，我们只关注数值特征和标题整体的贡献 # 计算所有标题向量特征的SHAP值总和 title_vector_contribution = np.sum(shap_values[0][len(numerical_features_cols):]) feature_contributions = {} for i, feature in enumerate(numerical_features_cols): feature_contributions[feature] = shap_values[0][i] feature_contributions['title_semantics'] = title_vector_contribution # 将标题作为一个整体 sorted_contributions = sorted(feature_contributions.items(), key=lambda item: abs(item[1]), reverse=True) diagnosis_report = [f"模型预测基准值: {base_value:.1%}", f"最终预测值: {predicted_ratio:.1%}\n", "主要影响因素分析:"] for feature, value in sorted_contributions: effect = "提升" if value > 0 else "拉低" diagnosis_report.append(f" - 特征 '{feature}' 将预测值 {effect} 了 {abs(value):.1%}") return prediction_text, "\n".join(diagnosis_report) if __name__ == '__main__': if len(sys.argv) != 6: print("\n使用方法: python predict.py \"<标题>\" <时长> <跳过率> <镜头数> <字幕(1/0)>") print("示例: python predict.py \"十分钟学会人像摄影，保姆级教程！\" 15 0.45 15 1\n") else: try: title_arg = sys.argv[1] duration_arg = float(sys.argv[2]) skip_rate_arg = float(sys.argv[3]) shot_count_arg = int(sys.argv[4]) has_subtitle_arg = int(sys.argv[5]) prediction, diagnosis = predict_new_video(title_arg, duration_arg, skip_rate_arg, shot_count_arg, has_subtitle_arg) print("-" * 50) print(prediction) print("\n💡 SHAP 归因诊断报告：") print(diagnosis) print("-" * 50) except (ValueError, IndexError): print("错误：请输入有效的参数。")