Streaming CTR Prediction: Rethinking Recommendation Task for Real-World Streaming Data

This paper presents a novel perspective on Click-Through Rate (CTR) prediction tasks, commonly used in recommendation systems. Rather than considering CTR prediction as a static task, the authors advocate for a “Streaming CTR Prediction” task, highlighting the dynamic nature of real-world data. They point out significant challenges in streaming data, such as distribution shifts, temporal non-stationarity, and biases, which can impact the effectiveness of prediction models. To overcome these challenges, they propose dedicated benchmark settings and metrics for streaming data, as well as two simple methods that significantly improve the CTR models’ effectiveness in the streaming scenario.

 

Publication date: July 14, 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2307.07509.pdf