PDQUBO: Performance-Driven Feature Selection for Recommender Systems on Quantum Annealers

A new method called PDQUBO (Performance-Driven Quadratic Unconstrained Binary Optimization) has been proposed for feature selection in recommender systems, designed to run directly on quantum annealers. Unlike previous QUBO-based approaches, PDQUBO explicitly quantifies the performance impact of individual features and feature pairs on recommendation models, aligning optimization objectives with model performance. The method uses counterfactual analysis, making it model-agnostic and evaluation-metric-independent. This work, published on arXiv (2410.15272), addresses the challenge of deploying quantum computing for practical recommender systems by ensuring the solution direction is tied to recommendation quality.