Cold Start Active Preference Learning in Socio-Economic Domains

Abstract: Active preference learning is a powerful paradigm for efficiently modeling preferences, yet it suffers from the cold-start problem: a significant drop in performance when no initial labeled data is available. This challenge is particularly acute in computational social systems and economic analysis, where labeled data is often scarce, expensive, and subject to expert noise. To address this gap, we propose a novel framework for cold-start active preference learning. Our method initiates the learning process through a self-supervised pre-training phase, utilizing Principal Component Analysis (PCA) to derive initial pseudo-labels from the data's inherent structure, thereby creating a cold-start model without any initial oracle interaction. Subsequently, the model is refined through an active learning loop that strategically queries a simulated noisy oracle for labels. We conduct extensive experiments on diverse datasets from different domains, including financial credibility, career success rate, and socio-economic status. The results demonstrate that our cold-start approach outperforms standard active learning strategies that begin from a blank slate, achieving higher accuracy with substantially fewer labeled pairs. Our framework offers a practical and effective solution to mitigate the cold-start problem, enhancing the sample efficiency and applicability of preference learning in data-constrained environments. We release our code at https://github.com/Dan-A2/cold-start-preference-learning