Neural Learning to Rank Model With Bias Correction and Attention Enhanced Relevance Prediction

K. J. Amala, D. Rajeswari

Abstract: Click data has emerged as an essential tool for training learning-to-rank systems in various sectors, including web search, recommender systems, and digital advertising. This data is fundamentally noisy and biased due to factors such as position bias and contextual influences, including document presentation, title length, and previous click patterns. This paper introduces the Probabilistic Click Prediction Network Based on Attention (APCP), which estimates document relevance and observation bias in a unified, attention-enhanced neural architecture. It examines the impact of contextual features on the likelihood of user observation and click behavior in a series of simulated use cases. In order to parameterize the probability of an observation, our framework integrates both static document attributes and dynamic user behavior signals. Experimental simulations utilizing synthetic click data indicate that the suggested model significantly mitigates the effects of positional and contextual bias, hence enhancing generalization and the accuracy of click-through rate predictions. APCP is trained using the MSLR-WEB30K dataset with simulated clicks and is validated on actual datasets Yahoo Learning to Rank Challenge dataset and Yandex click logs.It is evaluated against six robust baselines.Across five use-case situations, APCP consistently surpasses the most robust baseline, Implicit Intention Network (IIN), attaining AUC improvements ranging from +6.2% to +19.9% (all p <0.001, Cohen’s d >0.89, indicating a substantial impact size). Comparable enhancements are noted in NDCG@10, MAP, and MRR, substantiating APCP’s capacity to improve ranking quality and reduce bias in the MSLR WEB-30k dataset.On Yahoo, APCP enhances NDCG@1, @3, @5, and @10 by + 3.45% to + 5.28% compared to IIN, while on Yandex, it realizes AUC, MAP, and MRR improvements of + 3.55% to + 4.68%. All enhancements are statistically significant (p <0.05). These findings indicate that APCP not only efficiently reduces bias but also generalizes from simulated to actual click data, surpassing robust baselines across many ranking criteria. APCP offers a resilient and scalable methodology for improving search and recommendation systems.

External IDs:dblp:journals/access/AmalaR25