Contextual Dual Learning Algorithm with Listwise Distillation for Unbiased Learning to Rank

Learning to Rank (LTR) has evolved from relying solely on manual relevance labels to leveraging implicit user feedback, such as clicks. Early methods focused on modeling user behavior to infer relevance, but these often suffered from biases like position bias, trust bias, and contextual bias. Counterfactual approaches, such as inverse propensity weighting (IPW), attempted to correct these biases by estimating examination probabilities, yet they primarily considered position effects. The advent of large-scale real-world datasets, notably Baidu ULTR, enabled more realistic validation of ULTR methods. Despite progress, most existing models ignore the complex interactions within document lists, limiting bias correction effectiveness in practical settings.

Current ULTR methods predominantly address position bias, neglecting the influence of surrounding documents and local context on user examination behavior. This leads to inaccurate bias estimation and suboptimal ranking performance in real-world scenarios. Moreover, models trained solely on synthetic or simplified assumptions fail to capture the true user interaction patterns, resulting in poor generalization. The challenge is to develop a model that can simultaneously correct for position and contextual biases, accurately reflect user behavior, and scale to large datasets. Achieving this requires integrating interaction modeling, bias estimation, and knowledge transfer mechanisms within a unified framework.

The paper introduces several key innovations:

1) Transformer encoder-based listwise input model capturing cross-document interactions, enabling local context modeling.

2) Dual learning framework jointly optimizing a propensity model and a ranking model, ensuring unbiased bias estimation.

3) Listwise distillation transferring knowledge from the complex listwise model to a lightweight pointwise model, improving generalization.

These innovations collectively address the limitations of prior approaches, providing a comprehensive bias correction mechanism that aligns well with real user behavior.

• �� Input a list of documents into a Transformer encoder to generate context-aware feature vectors.
• �� Use a dual learning setup where the ranking model and propensity model are trained simultaneously, optimizing their respective loss functions based on inverse propensity weighting.
• �� The Transformer captures local interactions, reconstructing features that encode contextual relevance.
• �� Employ listwise loss functions, normalizing inverse propensity weights, to jointly optimize models.
• �� Implement listwise distillation: train a simpler pointwise model to mimic the listwise model's relevance judgments, using a specialized distillation loss.
• �� Use large-scale search logs for training, tuning hyperparameters such as Transformer layers, heads, and learning rates to maximize performance.

The experiments utilize a Baidu ULTR subset with over 500,000 sessions, covering diverse queries and user behaviors. Baselines include Naive, IPW, DLA, XPA, and UPE. Metrics are nDCG and ERR at top 10 positions. Hyperparameters are tuned via validation, with Transformer layers ranging from 1 to 4, heads from 2 to 8, and learning rates from 2e-6 to 2e-5. The training involves AdamW optimizer, batch size of 30, and list length of 10. Ablation studies compare the impact of listwise modeling and distillation. Results show that CDLA-LD consistently outperforms baselines, with statistically significant improvements in all metrics.

CDLA-LD achieves nDCG@10 of 0.5025, ERR@10 of 0.289, surpassing the best baseline IBOM-DLA by a small but significant margin. Propensity estimation aligns closely with actual user browsing patterns, demonstrating accurate bias modeling. The listwise distillation enhances the model's robustness, leading to better generalization on test data. Ablation results confirm that both the listwise input and distillation components are crucial for optimal performance. The learned propensity estimates reflect real-world user examination probabilities, validating the model's practical relevance.

This framework can be deployed in large-scale search engines and recommendation systems to improve fairness and relevance by correcting biases in user interaction data. It is particularly suitable for platforms with extensive logs of user behavior, enabling more accurate ranking models that better reflect true relevance. The approach can also be adapted for personalized ranking, where understanding individual user biases is critical. Long-term, this work can influence the development of fairer AI systems across various information retrieval domains.

The model's training complexity and computational cost are high, potentially limiting real-time deployment. Its effectiveness in extremely sparse or long-tail query scenarios remains uncertain. Cross-domain transferability needs further validation, and reliance on large annotated logs may restrict application in low-resource settings. Future work should focus on efficiency improvements and broader applicability.