Multi-objective optimization for long tail recommendation

摘要

Recommender systems are tools to suggest items to target users. Accuracy-focused recommender systems tend to recommend popular items, while suggesting items with few ratings (long tail items) is also of great importance in practice. Recommending long tail items may cause an accuracy loss of recommendation results. Thus, it is necessary to have a recommendation framework that recommends unpopular items meanwhile minimizing the accuracy loss. In this paper, we formulate a multi-objective framework for long tail items recommendation. Under this framework, two contradictory objective functions are designed to describe the abilities of recommender system to recommend accurate and unpopular items, respectively. To optimize these two objective functions, a novel multi-objective evolutionary algorithm is proposed. This multi-objective evolutionary algorithm aims to find a set of tradeoff solutions by optimizing two objective functions simultaneously. Experiments show that the proposed framework is effective to suggest accurate and novel items. The proposed recommendation algorithm could suggest many high-quality recommendation lists for the target user based on the concept of Pareto dominance in one run.