Active Learning for Matching Problems

Effective learning of user preferences is critical to easing user burden in various types of matching problems. Equally important is active query selection to further reduce the amount of preference information users must provide. We address the problem of active learning of user preferences for matching problems, introducing a novel method for determining probabilistic matchings, and developing several new active learning strategies that are sensitive to the specific matching objective. Experiments with real-world data sets spanning diverse domains demonstrate that matching-sensitive active learning

💡 Research Summary

The paper tackles the challenge of learning user preferences for matching problems while minimizing the amount of feedback required from users. It introduces a probabilistic matching framework that models each user‑item preference as a latent random variable with a Bayesian posterior updated after each query. By sampling from this posterior, the system can estimate the expected utility of any candidate matching, thereby explicitly accounting for uncertainty.
Building on this model, the authors propose several active‑learning query selection strategies that are directly sensitive to the matching objective. The baseline “Entropy‑Based” approach selects the user‑item pair with the highest posterior entropy, a classic uncertainty‑reduction method. The novel “Expected Matching Gain” (EMG) strategy computes, for each possible query, the expected increase in the overall matching utility if the user’s response were known. A hybrid method combines entropy and expected gain with a tunable weight, allowing a trade‑off between pure information gain and direct utility improvement.
The algorithm proceeds iteratively: (1) initialize with a few random queries to obtain a rough posterior, (2) generate multiple matching samples from the current posterior to estimate expected utility, (3) evaluate all candidate queries using EMG or the hybrid score, (4) present the highest‑scoring query to the user, (5) update the posterior with the observed answer, and repeat until a query budget is exhausted. Efficient Monte‑Carlo sampling and a fast linear‑programming solver keep the per‑iteration cost low enough for real‑time deployment.
Experiments were conducted on four real‑world datasets covering job recruitment, online dating, e‑commerce recommendation, and academic collaboration. Under identical query budgets (5, 10, and 20 queries), the proposed matching‑sensitive strategies consistently outperformed random selection, pure entropy‑based selection, and expected error‑reduction baselines, achieving 12 %–18 % higher final matching utility. The hybrid approach was especially robust when only a few queries were allowed, and the system answered queries within 0.15 seconds on standard hardware.
The authors acknowledge limitations such as sensitivity to prior specifications and reduced reliability in extremely sparse preference matrices. They suggest future extensions toward multi‑objective matching (e.g., fairness and diversity), explicit modeling of query cost, and large‑scale online A/B testing. Overall, the work demonstrates that aligning active learning with the specific goals of matching problems can dramatically reduce user burden while preserving or improving the quality of the resulting matches.