Nudging Neural Click Prediction Models to Pay Attention to Position

By E. K. Taniskidou et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. ABSTRACT
2. INTRODUCTION
3. SYNTHETIC POSITION-BIAS DEMONSTRATION
3.1 Model identifiability and flexibility
3.2 Generic regularization methods
4. CONCLUSION
5. REFERENCES

Summary

This paper discusses the importance of controlling position bias in neural models for click-through rate (CTR) prediction. It highlights the challenges faced in modeling recommendation outcomes with neural networks and the need for proper regularization methods to address position bias. The study demonstrates the difficulty of attributing outcomes to items and positions in neural networks and proposes approaches to mitigate bias. Various experiments and simulations illustrate the impact of different regularization techniques on model performance and bias reduction. The findings suggest the effectiveness of certain regularization methods in improving model fits and reducing position bias. Overall, the paper contributes to the understanding of position bias in CTR prediction models and provides insights into controlling biases for more accurate predictions.