Authors: Zhichao Feng, Junjiie Xie, Kaiyuan Li, Yu Qin, Pengfei Wang, Qianzhong Li, Bin Yin, Xiang Li, Wei Lin, Shangguang Wang

Published on: March 19, 2024

Impact Score: 7.4

Arxiv code: Arxiv:2403.12566

Summary

• What is new: Introduction of a Context-based Fast Recommendation Strategy to efficiently handle long user behavior sequences by selecting relevant sub-sequences.
• Why this is important: Existing models struggle with long-term dependencies or are too complex for Meituan Waimai’s recommender system.
• What the research proposes: A novel method that identifies contexts similar to the target context to find corresponding Points of Interest (PoIs) without needing to select a sub-sequence for each PoI, using prototype-based approaches and JS divergence for measuring similarity.
• Results: Since implementation in 2023, a 4.6% increase in Click-Through Rate (CTR) and a 4.2% increase in Gross Merchandise Value (GMV).

Technical Details

Technological frameworks used: Prototype-based approach, Temporal graph

Models used: Context selection for sub-sequences, CTR and CTCVR scoring with target attention

Data used: User behavior sequences, PoI attributes (categories, prices)

Potential Impact

Online food delivery platforms, recommender system-dependent businesses

Want to implement this idea in a business?

We have generated a startup concept here: ContextQuick.