Enhancing recommender ensemble by estimating input fitness

Abstract

Recommendation system is designed to tackle the information overload problem. The performance of a single recommendation system can be significantly improved if ensemble methods are used. In some recent works researchers seek ways to leverage past history for enhancing the performance of ensemble methods, The core concept rooted in this strand of research is based on exploiting user meta-data for enriching the input spaces for ensembler, however, in many user-agnostic recommendation scenarios such as ad-hoc searching where the user IDs and meta-data are absent, which makes such meta-data-driven ensembling method infeasible. In this work, we proposed a novel user-agnostic ensembling model Rec4Rec, which learns the representations of base recommenders from the input sequences via attention mechanism, in such a manner, the patterns and characteristics of input sequences are memorized within base recommender representations. The base recommender fitnesses are estimated through computing distance with input sequences and can be readily integrated with outputs of base recommenders for ensemble learning, no user-IDs or meta-data are required here as opposed to other meta-data-driven ensemble approaches. Experiment results show that Rec4Rec significantly outperforms the stacking ensembler. The embedding visualization proves that the proposed ensemble method can effectively distinguish the base recommenders. The superior performance of Rec4Rec demonstrates the importance of building representations for base recommenders of an ensemble in ad-hoc inquiries.