Abstract
Active preference learning offers an efficient approach to modelingpreferences, but it is hindered by the cold-start problem, which leads to amarked decline in performance when no initial labeled data are available. Whilecold-start solutions have been proposed for domains such as vision and text,the cold-start problem in active preference learning remains largelyunexplored, underscoring the need for practical, effective methods. Drawinginspiration from established practices in social and economic research, theproposed method initiates learning with a self-supervised phase that employsPrincipal Component Analysis (PCA) to generate initial pseudo-labels. Thisprocess produces a \say{warmed-up} model based solely on the data's intrinsicstructure, without requiring expert input. The model is then refined through anactive learning loop that strategically queries a simulated noisy oracle forlabels. Experiments conducted on various socio-economic datasets, includingthose related to financial credibility, career success rate, and socio-economicstatus, consistently show that the PCA-driven approach outperforms standardactive learning strategies that start without prior information. This work thusprovides a computationally efficient and straightforward solution thateffectively addresses the cold-start problem.