When Search Meets Recommendation: Learning Disentangled Search Representation for Recommendation

Abstract

Modern online service providers such as online shopping platforms often\nprovide both search and recommendation (S&R) services to meet different user\nneeds. Rarely has there been any effective means of incorporating user behavior\ndata from both S&R services. Most existing approaches either simply treat S&R\nbehaviors separately, or jointly optimize them by aggregating data from both\nservices, ignoring the fact that user intents in S&R can be distinctively\ndifferent. In our paper, we propose a Search-Enhanced framework for the\nSequential Recommendation (SESRec) that leverages users' search interests for\nrecommendation, by disentangling similar and dissimilar representations within\nS&R behaviors. Specifically, SESRec first aligns query and item embeddings\nbased on users' query-item interactions for the computations of their\nsimilarities. Two transformer encoders are used to learn the contextual\nrepresentations of S&R behaviors independently. Then a contrastive learning\ntask is designed to supervise the disentanglement of similar and dissimilar\nrepresentations from behavior sequences of S&R. Finally, we extract user\ninterests by the attention mechanism from three perspectives, i.e., the\ncontextual representations, the two separated behaviors containing similar and\ndissimilar interests. Extensive experiments on both industrial and public\ndatasets demonstrate that SESRec consistently outperforms state-of-the-art\nmodels. Empirical studies further validate that SESRec successfully disentangle\nsimilar and dissimilar user interests from their S&R behaviors.\n