Abstract
Existing recommender systems usually generate personalized recommendation lists based on the estimation of the preference scores over user-item pairs, while ignoring the impacts of the entire display list that plays a central part in the decision making process of a user. This leaves us an opportunity to generate better recommendation results by considering the impacts of all offered choices. However, such an extension cannot be handled efficiently by traditional top- $k$ list recommendation methods, due to the entire list dependency issue which means a complete list of items is needed before we can precisely measure any item preference among the list. In this paper, we propose a Co-displayed Items Aware (CDIA) list generation approach, which is based on the reinforcement learning architecture, and can efficiently generate high-utility lists. Specifically, we propose CDIA-Sim to predict users’ preferences, which considers the impacts of the co-displayed items. Then, to overcome the entire list dependency issue in the list recommendation task, we utilize the reinforcement learning technique and design CDIA-RL to generate high-utility lists. Experimental results show that CDIA-Sim achieves significant improvements in modeling user-item preferences, and CDIA-RL can yield lists efficiently and effectively, illustrating better performance than other competitors.
Highlights
In recommender systems, traditional point-wise or pair-wise methods [1]–[3] generate recommendation lists for each user u by estimating p(i|u) or p(i > j|u)
The results show that compared with the traditional user-item preference estimation models that ignore the surrounding choices, Co-displayed Items Aware (CDIA)-Sim can significantly improve the prediction accuracy on users’ behaviors, and CDIA-RL can efficiently generate much better high-utility recommendation lists than other list recommendation methods
Here. though we focus on the impacts of the co-displayed items and the item display position information is degraded to some extent in all the four datasets as discussed in Section IV-A, we still receive slight performance improvements in multilayer perceptron (MLP)+Pos sometimes
Summary
Traditional point-wise or pair-wise methods [1]–[3] generate recommendation lists for each user u by estimating p(i|u) or p(i > j|u). The top-k items with the highest estimated probabilities are selected [3], [4]. Most recent works focus on how to improve the precision of the estimated p(i|u) or p(i > j|u) [3], [5], [6]. They all follow a greedy top-k item selection process when recommending a new list. As these methods do not consider the impacts of the entire list, we call these methods list-free recommendation methods. In the e-commerce recommender system, the displayed products are comparable in aspects of price, brand, reviews, etc, and there are complex relationships such as complementation or substitution [7] among products, making the users’ behaviors highly related to what they can see
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
