Abstract:

In the information feeds consumption scenario, it is an essential problem to personalize content recommendation using the user’s implicit behavior feedback. Due to behavioral inertia, users usually only browse the feed stream, so the interactive behavior data is relatively sparse, which leads to the low performance of traditional methods, expecially in terms of personalization. Aiming at this problem, this paper designs the weight conversion method of implicit feedback, and then proposes the LFM-XGB-LR hybrid model. It uses the LFM algorithm to generate the embedded vector, and combines the advantages of XGB in feature crossing and LR in discrete calculation. Experiment and analysis results show that LFM embedding improves individuation of the recommendation, and the hybrid model has improvements on each evaluation index.

Key words: Latent Factor Model（LFM）, implicit vector embedding, implicit feedback, hybrid model, recommender system

摘要：

在信息流消费场景中，利用用户的隐式行为反馈，对用户进行个性化内容推荐是核心问题。而由于行为惯性的问题，用户通常只是浏览feed流，互动行为数据稀疏，导致传统方法在个性化等方面性能不高。针对该问题，设计了隐式反馈的权重转化方法，提出LFM-XGB-LR融合模型，利用LFM生成嵌入向量，结合了XGB在特征交叉和LR在离散计算上的优势。实验结果表明，基于LFM的嵌入改善了模型个性化的问题，该融合模型在各项指标上均有稳定提升。

关键词: 隐语义模型, 隐向量嵌入, 隐式反馈, 融合模型, 推荐系统