Reproduce “Unified Dual-Intent Translation for Joint Modeling of Search and Recommendation” on OpenResearch

About this paper

Yuting Zhang, Yiqing Wu, Ruidong Han, Ying Sun, Yongchun Zhu, Xiang Li, Wei Lin, Fuzhen Zhuang and 2 more

Researchers from ICT CAS, Meituan, and HKUST developed UDITSR, a unified framework that improves search and recommendation by modeling users' unchanging inherent intents and changing demand intents. The model leverages explicit search queries to supervise the learning of implicit demand intents in recommendation, leading to average performance improvements of 6.22% in search and 3.06% in recommendation on Meituan datasets, and a 1.46% GMV increase in online A/B tests.

Key insights

• Explicit search queries can effectively supervise the learning of implicit demand intents in recommendation scenarios, bridging the gap between explicit and implicit intent modeling.
• User behavior is driven by a combination of unchanging inherent preferences and dynamic, context-dependent demand intents.
• Modeling the interaction between inherent intent and demand intent as a 'translation' (inherent + demand = item) provides a more interpretable and effective way to predict item interactions.

Summary by alphaXiv.