基于时间感知与协同挖掘的下一购物篮推荐方法

Next-Basket Recommendation Based on Time Perception and Collaborative Mining

下一购物篮推荐是电商平台上最重要的任务之一,旨在挖掘用户的购物习惯及其兴趣进化的特征。现有的下一购物篮推荐方法存在以下不足:一是仅基于购物篮先后位置建模无法捕捉购物篮时间间隔的差异性;二是基于RNN(recurrent neural network)的方法无法捕获个性化商品频率信息并缺乏可解释性。上述不足限制了电商推荐的准确率且无法提供给用户直观的推荐理由。因此,提出了一种基于时间感知和协同挖掘的下一购物篮推荐方法。该方法对购物篮时间进行建模,将购物篮划分为表征用户短期兴趣的多个组别,并采用层次时间衰退实现用户兴趣进化挖掘;同时对用户表达进行最近邻聚类,基于协同过滤思想增加可解释性。实验表明,该方法能有效建模用户的兴趣进化并提供可解释性,在多种评价指标上优于主流方法。

The next-basket recommendation is one of the most important tasks on the e-commerce platform, which aims to explore the characteristics of users’ shopping habits and the evolution of their interests. The existing basket methods have the following shortcomings. First, only modeling based on the sequential position of baskets cannot capture the difference of basket time interval;Second, the RNN(recurrent neural network)based methods cannot capture the personalized item frequency information and lack interpretability. The above shortcomings limit the accuracy of e-commerce recommendation and cannot provide users with intuitive reasons for recommendation. Therefore, a next-basket recommendation method based on time perception and collaborative mining was proposed. In this method, the baskets were divided into multiple groups representing users’ short-term interests by modeling the basket time directly, and a hierarchical time decay was used to realize user interest evolutionary mining. At the same time, the nearest neighbor clustering of user expression was carried out to increase the interpretability based on the idea of collaborative filtering.Experiments proved that the proposed method can effectively model the evolution of users’ interest and provide interpretability,and is superior to the mainstream methods in a variety of evaluation metrics.