Projects

This project advances quantum-enhanced recommender systems by integrating efficient quantum optimization, qubit-aware feature selection, and automated quantum circuit design. Building on our recent work—including the CAQUBO method for performance-driven feature selection using quantum annealing, an extreme-value-theory framework for estimating the quantum sampling cost required to capture near-optimal feature sets, a qubit-efficient quantum semi-random-forest model for recommendation scoring, and a hybrid-action reinforcement-learning approach for discovering compact, high-performing quantum circuit architectures—we are now developing a unified, end-to-end pipeline for quantum-assisted recommendation. The project explores scalable feature compression for large tag spaces, QUBO formulations aligned with real system performance, adaptive sampling strategies for noisy quantum hardware, and RL-driven circuit search tailored to recommender objectives, with the overarching goal of demonstrating practical, NISQ-era viability of quantum-classical hybrid recommenders and providing a roadmap for their continued evolution as quantum devices mature.

This project brings together a body of work on responsible recommendation to create a unified research program that (1) explains and quantifies fairness at the level of individual users and items, (2) integrates explainability into fairness-aware learning, (3) exposes and corrects evaluation and algorithmic biases, and (4) extends these concerns to emerging LLM-based recommenders; specifically, it synthesizes recent contributions on explaining recommendation fairness from a user/item perspective to surface who is affected and why (TOIS 2025), GAN-based fairness-aware training for implicit feedback to mitigate disparate impacts in learning, and multi-task explainable recommendation models that jointly predict ratings and human-readable review summaries to improve transparency and accountability, while also addressing evaluation pathologies such as popularity bias and false-positive metric distortions in offline evaluation, and investigating position and ordering biases that appear when large language models are used for recommendation — the project’s goal is an end-to-end framework combining interpretable explanations, fair representation and training techniques, robust evaluation protocols, and bias-aware deployment practices so that recommender systems can be both effective and demonstrably responsible.

This project brings together a concentrated line of research on rigorous and responsible evaluation of recommender systems, addressing critical flaws in conventional assessment and proposing novel methodologies for more meaningful metrics and simulation. Building on work such as the analysis of popularity bias and the misalignment between true- and false-positive offline metrics (SIGIR 2020 / ACM TIS 2021) DBLP, the team developed a bi-directional Item Response Theory (IRT) framework to better capture user-item interaction dynamics in algorithm evaluation (WWW Companion 2025) DBLP, and created an LLM-enhanced, personality-driven user behavior simulator (SIGIR 2025) DBLP to reflect realistic user heterogeneity. Simultaneously, they conducted a metamorphic evaluation of ChatGPT as a recommender system (CoRR 2024) DBLP and examined position bias in LLM-based recommendation lists (CoRR 2025) DBLP. Through these contributions, the project aims to build an end-to-end evaluation paradigm that incorporates simulation, bias-aware metrics, and human-like behaviors to ensure recommender systems are assessed not just for accuracy, but for fairness, robustness, and real-world validity.