Learning Preference from Observed Rankings

Researchers Yu-Chang Chen, Chen Chian Fuh, and Shang En Tsai developed a statistical framework for learning consumer preferences from partial ranking data. It models latent utility with interpretable attributes and a low-rank factor structure. A key innovation corrects for 'exposure bias' where only items in a user's consideration set are observed. The method uses an inverse-probability-weighted, ridge-regularized log-likelihood and a scalable SGD algorithm. In tests with online wine retailer data, it significantly improved recommendations for previously unconsumed products.