ATRD: A Comprehensive Framework to Implement Debiasing and Calibration Simultaneously in Recommender Systems

Yunkai Chen, Chenhao Li, Zhongya Xue, Xibo Song, youyou zhang, YichunLi, Erpeng Qi, JingTao

Keywords: Recommender Systems, Debiasing and Calibration, Traffic proportion, Traffic-redistribution

Abstract: Accurate prediction of rates (e.g., click-through rate and conversion rate) and values (e.g., watch time and pay amount) is a fundamental pursuit of modern recommender systems. Due to training sample bias and model training error, few online models are able to deliver an absolutely precise prediction which can fully align with the ideal data distribution. Existing research has developed two technical approaches, i.e., debiasing and calibration, which address training sample bias and model training error respectively, failing to optimize both types of errors simultaneously. In this paper, we propose the Adaptive Traffic Redistribution (ATRD) framework, which implements debiasing and calibration from a comprehensive perspective. Firstly, we propose parallel sampling and traffic minimal connected graph to construct a series of comparable samples of item traffic proportion and the corresponding efficiency. Secondly, we fit the function which maps item traffic proportion to its efficiency and solve for the primal traffic proportion. Thirdly, we apply step exploration and subgradient descent to derive the correction factor for online traffic adjustment. Theoretically, the proposed framework can ensure that the exposure of items is more commensurate with their true efficiency by traffic redistribution, leading to the optimization of recommendation results. Online experiments validate the effectiveness of the proposed method and demonstrate significant improvements in business metrics.

Primary Area: optimization

Submission Number: 8354