MITIGATING POSITION BIAS IN CLICK PREDICTOR MODELS: A NOVEL DOWNSAMPLING APPROACH FOR ENHANCED ACCURACY AND EFFICIENCY

Abstract

In the rapidly evolving landscape of digital information retrieval and online advertising, response predictor models have become indispensable tools for understanding user behavior and optimizing content delivery. These models aim to estimate the likelihood of a user response on a displayed item, such as a search result or advertisement, thereby enabling more efficient and targeted allocation strategies. However, the accuracy and reliability of response predictor models are significantly challenged by position bias – a phenomenon where the position of an item in a list influences the probability of user responding to it though either click, conversion, etc., independent of its relevance or quality.

This paper introduces a novel downsampling method designed to mitigate the effects of position bias while addressing the challenges posed by large-scale interaction data and resource constraints. Our approach preserves all positive samples and filters out a substantial amount of poor negative samples, maintaining the integrity of valuable information necessary for accurate predictions. By leveraging the insight that item response probabilities should be uniform across positions in the absence of bias, we propose a random filtering strategy that optimizes the preservation of valuable interaction data.

We demonstrate the effectiveness of our method through extensive offline and online experiments, showing that it not only complements existing bias mitigation techniques but also enhances model accuracy in resource-constrained environments. Our findings suggest that integrating this downsampling method with new training strategies leads to improved prediction performance.