Sequential Evaluation and Generation Framework for Combinatorial Recommender System

Typical recommender systems push K items at once in the result page in the form of a feed, in which the selection and the order of the items are important for user experience. In this paper, we formalize the K-item recommendation problem as taking an unordered set of candidate items as input, and exporting an ordered list of selected items as output. The goal is to maximize the overall utility, e.g. the click through rate, of the whole list. As one solution to the K-item recommendation problem under this proposition, we proposed a new ranking framework called the Evaluator-Generator framework. In this framework, the Evaluator is trained on user logs to precisely predict the expected feedback of each item by fully considering its intra-list correlations with other co-exposed items. On the other hand, the Generator will generate different sequences from which the Evaluator will choose one sequence as the final recommendation. In our experiments, both the offline analysis and the online test show the effectiveness of our proposed framework. Furthermore, we show that the offline behavior of the Evaluator is consistent with the realistic online environment.