FIMN: Fusing Multi-Interest and Scenario-Mutual Network for Multi-Scenario Recommendation

Wei Xu, Fei Fei, RuiwenXu

Abstract: Recently, multi-scenario learning has achieved flourishing development in recommendation and retrieval systems in E-commerce platforms. Current numerous models have been proposed that attempt to use a unified model to serve multiple scenarios. However, three critical challenges still remain to be carefully addressed. First, users in different scenarios explicitly have different behavior interests, which is vital for modeling but has been neglected in previous works. Second, it is intuitive that relationships between scenarios is intricate as various scenarios generally have commonalities and specific characteristics, while previous solutions neglect the complicated interrelations among scenarios. Moreover, current state-of-the-art unified models may not work well in all scenarios, since they usually face head scenario domination phenomenon due to the different data distribution. To resolve these problems, we propose a novel approach named Fusing multi-Interest and scenario-Mutual Network (FIMN), which mainly consists of four modules. FIMN performs explicitly multi-interest fusing corresponding to specific scenario and learns correlations across scenarios dynamically, meanwhile the scenario distribution discrepancy problem can be mitigated. Extensive experiments show the superiority of FIMN towards the state-of-the-art methods. FIMN has been successfully deployed in our online retrieval platform.