Disentangling Scenario-wise Interest Network for Multi-scenario Recommendation

With rapid development of the E-commerce platforms, multiple shopping scenarios are rapidly emerged to cater to the diversified needs of users. Currently, numerous models have been proposed to use a unified framework to serve multiple scenarios. However, most of those methods aggregate historical behavior sequences as a whole when modeling users' interests, neglect the discrepancy of user behaviors in different scenarios. To accurately capture the migration of interests across different scenarios, we propose a novel Disentangling Scenario-wise Interest Network, which explicitly incorporates scenario related context into user behavior sequences and learns users' scenario-wise interests. DSWIN consists of three key modules: 1) Global Interest Aggregation (GIA), which fuses users' global behaviors and scenario-aware context information with useful aggregation mechanism; 2) Local Interest Resolution (LIR), which explicitly extracts users' scenario-wise interests in each scenario through self-attention; 3) Interest Disentangling Module (IDM), which disentangles and distinguishes users' interests in different scenarios with self-supervised learning. Extensive experimental results and online A/B testing validate the effectiveness of DSWIN compared with state-of-the-art methods.