Reversing the Retrieval Engine: Query Performance Prediction as an Inverse Learning Task

Abstract

Query performance prediction (QPP) aims to estimate the effectiveness of search queries in the absence of explicit relevance judgments. Existing approaches typically rely on hand-crafted features or learn forward mappings from retrieval signals to effectiveness metrics. In this work, we propose a novel formulation of QPP as an inverse learning problem , motivated by the hypothesis that observed effectiveness scores arise from latent query characteristics that can be reconstructed by reversing the retrieval process. We introduce an inverse regression framework that maps performance labels (e.g., NDCG@10) back to estimated retrieval features such as clarity, score variance, and entropy, thereby approximating an inverse function \(\hat{x}=F^{-1}(y)\) . This formulation shifts QPP from direct prediction toward reconstructing latent retrieval properties that explain performance outcomes. Experiments on MS MARCO demonstrate the effectiveness of the proposed approach, achieving a Kendall’s \(\tau\) of 0.5402 and a Spearman’s \(\rho\) of 0.742, substantially outperforming a conventional forward model (0.344 and 0.4885, respectively). We further validate robustness on the TREC Robust04 collection, where the inverse model consistently surpasses the forward baseline in rank correlation and achieves lower RMSE values. Overall, the results show that reversing the retrieval process yields more accurate, stable, and generalizable QPP. This inverse perspective enables interpretable, model-agnostic performance prediction and opens new directions for understanding retrieval behavior through latent reconstruction.