融合时间信息的序列商品推荐模型-《计算机技术与发展》

文章信息/Info

Title:: A Sequential Product Recommendation Model Integrating Time Information

Author(s):: XU Hong-yan; DANG Yi-ming; FENG Yong; WANG Rong-bing; School of Information,Liaoning University,Shenyang 110036,China

Keywords:: sequential recommendation; long-short term interest; time information; multi-headed self-attention mechanism; deep learning

摘要:: 针对基于序列的推荐方法通常忽略用户的多种兴趣倾向,并且不能很好地获取用户在短期序列中的兴趣变化,从而导致推荐结果多样性不足的问题,提出了一种融合时间信息的序列商品推荐模型。首先,将用户的历史交互行为区分为短期序列与长期序列,分别采取不同的方法进行建模。对于短期序列,在传统的门控循环单元( GRU) 结构中加入时间门,单独处理序列中的时间信息,同时利用多头自注意力机制捕获用户在同一会话中不同的兴趣方向;对于长期序列,采用 DeepFM 模型进行建模。最后,利用自适应的门控结构融合用户的长短期兴趣,并根据得到的兴趣向量计算商品的得分,排序后进行推荐。在淘宝数据集上的对比实验表明,该模型相较于主流的协同过滤模型,基于 RNN、DNN 的推荐模型以及 BINN 模型在命中率、平均倒数排名两个指标上都具有显著优势。

Abstract:: In order to solve the problem that sequence-based recommendation methods usually ignore users’ various interest tendenciesand fail to obtain users’ interest changes in short - term sequence, which leads to insufficient diversity of recommendation results, asequential product recommendation model integrating time information is proposed. For short-term sequence,a time gate is added intoGRU to process time information separately,and multiple self - attention mechanism is used to capture different directions of interest.Then DeepFM is used to model long-term sequence. Finally,an adaptive gating structure is applied to integrate long-term and short-term interests. The experiments on Taobao dataset show that compared with the mainstream collaborative filtering model, therecommendation model based on RNN and DNN and the BINN model,the proposed model has significant advantages in terms of hit ratioand average reciprocal ranking.