Publications
CONFERENCE (INTERNATIONAL) PR-Rank: A Parameter Regression Approach for Learning-to-Rank Model Adaptation Without Target Domain Data
Takumi Ito (Tsukuba University.), Atsuki Maruta (Tsukuba University.), Makoto P. Kato (Tsukuba University.), Sumio Fujita
The 25th International Web Information Systems Engineering conference (WISE 2024)
November 27, 2024
We propose a novel method that constructs a Learning-torank (LtR) model tailored to the target domain without using any target domain-specific queries, documents, and relevance judgements. Instead, our approach relies solely on domain features, which can be estimated based on domain knowledge, for instance, by search engine engineers. Central to our method is a parameter regression model that learns to predict the optimal weights of the LtR model from the domain features. This eliminates the need for access to data from the target domain, which is often unavailable. The weights prediction model is trained using source domains prepared by dividing a single large dataset into multiple domains with different characteristics. Experimental results suggest that our proposed model outperformed the model trained on a large amount of data without considering domain differences.
Paper :
PR-Rank: A Parameter Regression Approach for Learning-to-Rank Model Adaptation Without Target Domain Data
(external link)