A Study on Predicting Term Weighting Models with Query Performance Predictors to Enhance Information Retrieval Effectiveness

Abstract

Information retrieval systems often exhibit variability in effectiveness across queries and retrieval models. This study investigates the use of pre-retrieval query performance predictors to guide per-query prediction of term weighting models. A prediction-based method (PBM) is employed using a K-Nearest Neighbors (KNN) to predict between BM25 and DFRee term weighting models based on features derived from query performance predictors. Experiments are conducted on standard TREC datasets, which are CW09B, CW12B, and GOV2, and retrieval effectiveness is evaluated using nDCG@20. Results demonstrate that PBM consistently outperforms the baseline models. It achieves statistically significant improvements in average effectiveness and enhanced robustness across queries. Per-query analyses show that PBM accurately predicts the more effective model between term weighting models. These findings indicate that query-level model prediction based on query performance predictors can substantially improve IR performance.