Reproduce “Time Matters: A Novel Real-Time Long- and Short-term User Interest Model for Click-Through Rate Prediction” on OpenResearch

About this paper

Xian-Jin Gui

Researchers at Nanjing University's LAMDA Group developed the Time-aware Long- and Short-term user Interest model (TLSI) to predict click-through rates by explicitly modeling periodic and time-point patterns in user interests. The model achieved an AUC of 0.9251 on the Taobao dataset and 0.9676 on a large industrial video platform dataset, outperforming state-of-the-art baselines like SLi-Rec by a notable margin.

Key insights

• User interests are not static but exhibit distinct, personalized "periodic" and "time-point" patterns that are critical for accurate real-time CTR prediction.
• Explicitly modeling the correlation between activated interest and the specific occurrence time, rather than just sequential order, yields substantial performance improvements.
• A comprehensive framework that dynamically tailors long-term interest, employs a time-aware short-term interest model, and adaptively fuses these components provides a more robust representation of user preferences.

Summary by alphaXiv.