Fine-tuning for Inference-efficient Calibrated Recommendations

Abstract

CaliTune is a fine-tuning method for collaborative-filtering recommender systems that improves popularity calibration without relying on computationally expensive post-processing at inference time. Experiments across two backbone models and datasets from the movie and music domains show competitive accuracy–calibration trade-offs, especially when the base model is strongly miscalibrated and recommendation accuracy remains important.

Publication
_Proceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys 2025), pp. 1187-1192, https://doi.org/10.1145/3705328.3759319_