使用Optuna优化协同过滤算法：一个超参数调优实践

pip install optuna surprise pandas numpy

from surprise import Dataset, Reader, KNNBasic
from surprise.model_selection import cross_validate
import pandas as pd

# 加载数据
data = pd.read_csv('movie_ratings.csv')
reader = Reader(rating_scale=(1, 5))
dataset = Dataset.load_from_df(data[['user_id', 'movie_id', 'rating']], reader)

# 基线模型
baseline_model = KNNBasic(
    k=40,
    sim_options={'name': 'cosine', 'user_based': True}
)

# 交叉验证
baseline_results = cross_validate(baseline_model, dataset, measures=['RMSE'], cv=5, verbose=True)
print(f"基线RMSE: {baseline_results['test_rmse'].mean():.4f}")

import optuna
from surprise import accuracy
from surprise.model_selection import train_test_split

def objective(trial):
    # 定义超参数搜索空间
    k = trial.suggest_int('k', 5, 100)
    sim_name = trial.suggest_categorical('sim_name', ['cosine', 'msd', 'pearson'])
    user_based = trial.suggest_categorical('user_based', [True, False])
    
    # 创建模型
    model = KNNBasic(
        k=k,
        sim_options={'name': sim_name, 'user_based': user_based}
    )
    
    # 数据集划分
    trainset, testset = train_test_split(dataset, test_size=0.2)
    
    # 训练与评估
    model.fit(trainset)
    predictions = model.test(testset)
    rmse = accuracy.rmse(predictions, verbose=False)
    
    return rmse

# 创建Optuna研究
study = optuna.create_study(direction='minimize')
study.optimize(objective, n_trials=100)

# 输出最佳参数
print("最佳参数:", study.best_params)
print("最佳RMSE:", study.best_value)

from surprise.model_selection import cross_validate
from optuna.trial import TrialState

def objective_with_pruning(trial):
    k = trial.suggest_int('k', 5, 100)
    sim_name = trial.suggest_categorical('sim_name', ['cosine', 'msd', 'pearson'])
    
    model = KNNBasic(
        k=k,
        sim_options={'name': sim_name, 'user_based': True}
    )
    
    # 使用中间值进行剪枝
    for fold in range(3):
        cv_results = cross_validate(model, dataset, measures=['RMSE'], cv=3, verbose=False)
        current_rmse = cv_results['test_rmse'].mean()
        
        # 报告中间结果
        trial.report(current_rmse, fold)
        
        # 处理剪枝
        if trial.should_prune():
            raise optuna.TrialPruned()
    
    return current_rmse

# 使用TPE采样器
sampler = optuna.samplers.TPESampler()
study = optuna.create_study(direction='minimize', sampler=sampler)
study.optimize(objective_with_pruning, n_trials=50)

import multiprocessing

if __name__ == '__main__':
    # 设置并行试验
    study = optuna.create_study(direction='minimize')
    study.optimize(objective, n_trials=100, n_jobs=multiprocessing.cpu_count()-1)

# 集成学习示例
from surprise import KNNWithMeans

def create_ensemble(params_list):
    models = []
    for params in params_list:
        models.append(KNNWithMeans(**params))
    return models

# 使用Optuna优化得到的多个最佳参数组合
best_params_list = [study.best_params for study in studies]
ensemble_models = create_ensemble(best_params_list)