电商平台个性化推荐中相似度优化与可学习模型融合研究

Research on the Similarity Optimization and Learnable Model Integration for Personalized Recommendation in E-Commerce Platforms

摘要: 随着电子商务的发展，个性化推荐系统在提升用户满意度、点击率和转化率方面起着重要作用。然而，传统的协同过滤方法往往面临数据稀疏性和用户评分偏差等问题，限制了其预测精度和排序效果。本研究旨在通过将相似度度量与残差学习框架相结合，优化电子商务场景下的个性化推荐性能。本文基于MovieLens 100k数据集，探索了在不同相似度方法下结合手工特征与可学习模型的推荐方法。首先提取了用户活跃度(Activity)、评分序列相关性(DFA)和声誉(Reputation)三个手工特征，并在Cosine、Jaccard、Pearson和Spearman四种相似度基础上，使用了XGBoost等六种可学习模型对预测残差进行建模与校正。实验结果显示：在不同的相似度方法下，XGBoost模型性能最优，其中Jaccard相似度下的XGBoost在RMSE、MAE指标上表现最佳，并且单个手工特征使用的效果有限，而特征组合能够显著提升推荐性能。此外，将Jaccard相似度结合XGBoost可学习模型的推荐模型与常用推荐算法进行对比，发现其在RMSE和MAE上显著优于传统方法，且NDCG@10指标最高，验证了手工特征与可学习模型结合在电商推荐中的有效性。本文研究结果为电商平台提供了优化个性化推荐策略的可行方案。

Abstract: With the rapid development of e-commerce, personalized recommendation systems play a critical role in enhancing user satisfaction, click-through rates, and conversion rates. However, traditional collaborative filtering methods often suffer from data sparsity and user rating bias, which limit prediction accuracy and ranking performance. To address these challenges, this study proposes a hybrid recommendation framework that integrates similarity measures with a residual learning mechanism to optimize personalized recommendation performance in e-commerce scenarios. Based on the MovieLens 100K dataset, we investigate recommendation approaches that combine handcrafted user features with learnable models under different similarity metrics. Specifically, three handcrafted features—user activity, rating sequence correlation measured by detrended fluctuation analysis (DFA), and reputation—are extracted. On the basis of four similarity measures (Cosine, Jaccard, Pearson, and Spearman), six learnable models, including XGBoost, are employed to model and correct the prediction residuals. Experimental results demonstrate that XGBoost consistently achieves superior performance across different similarity measures. In particular, the combination of Jaccard similarity and XGBoost yields the lowest RMSE and MAE values. Moreover, using a single handcrafted feature yields limited improvement, whereas the combination of all features significantly enhances recommendation performance. Furthermore, compared with commonly used recommendation algorithms, the proposed hybrid model significantly outperforms traditional methods in terms of RMSE and MAE, while achieving the highest NDCG@10. These findings validate the effectiveness of integrating handcrafted features with learnable models for e-commerce recommendation and provide empirical support for optimizing personalized recommendation strategies.