Local Collaborative Autoencoders

Top-N recommendation is a challenging research problem because complex user-item interactions should be adequately considered to achieve high-quality results. The local approach has been successful in this task by utilizing multiple local models to capture diverse user preferences from different sub-communities. However, previous studies have only partially explored the potential of local models, failing to identify many small and coherent sub-communities. In this paper, we present Local Collaborative Autoencoders (LOCA), a generalized framework that fully takes advantage of local approaches. Specifically, LOCA adopts different neighborhood ranges at training and inference, respectively. In addition, LOCA equips with a novel sub-community discovery method, maximizing the coverage of a union of local models and employing a large number of diverse local models effectively. By adopting autoencoders as the base models, LOCA effectively captures non-linear patterns for representing meaningful user-item interactions within sub-communities. Our experimental results show that LOCA is highly scalable and achieves state-of-the-art performance on several public benchmarks, by 2.99-4.70% in Recall and 1.02-7.95% in NDCG, respectively.