一种面向稀疏数据的自适应视图优化对比学习推荐方法

An Adaptive View Optimization Contrastive Learning Recommendation Method for Sparse Data

摘要: 随着电子商务的快速发展，中小电商已成为数字经济的重要参与主体，推荐系统作为电商平台的核心引擎直接影响用户购物体验。然而，中小电商普遍面临用户交互数据稀疏、相似商品难以精准区分的困境，传统协同过滤算法在数据稀疏条件下推荐效果显著下降。针对上述问题，本文提出了一种面向稀疏数据的自适应视图优化对比学习推荐方法(简称：AVOCL)。该方法根据用户交互数据的稀疏程度动态调整图视图增强策略，在稀疏交互条件下自动生成高质量训练样本；同时通过特征距离筛选困难负样本并动态优化区分权重，使模型训练聚焦于最具辨别价值的高难度负样本。在多个公开数据集上.的实验结果表明，本文方法在用户交互稀疏条件下(平均交互不足5条)，推荐点击率与成交转化率较现有对比学习方法提升10%以上，有效缓解了数据稀疏带来的推荐效果下降问题。

Abstract: With the rapid development of e-commerce, small and medium-sized e-commerce enterprises have become important participants in the digital economy, and recommendation systems serve as the core engine of e-commerce platforms, directly affecting user shopping experience. However, these enterprises generally face two major challenges: sparse user interaction data and difficulty in accurately distinguishing similar products. Traditional collaborative filtering algorithms significantly degrade in performance under data-sparse conditions. To address these problems, this paper proposes an Adaptive View Optimization Contrastive Learning recommendation method for sparse data (abbreviated as AVOCL). The method dynamically adjusts the graph view enhancement strategy based on the sparsity level of user interaction data, automatically generating high-quality training samples under sparse interaction conditions. Meanwhile, it screens hard negative samples through feature distance and dynamically optimizes differentiation weights, focusing model training on the most valuable hard negative samples. Experimental results on multiple public datasets show that under sparse interaction conditions (average less than 5 interactions per user), the proposed method improves recommendation click-through rate and conversion rate by more than 10% compared to existing contrastive learning methods, effectively alleviating the performance degradation caused by data sparsity.