한국형 온디바이스 AI 반도체 및 단안 비전 기반 장애물 회피 및 자율비행 실증 연구

초록

Item cold-start remains a fundamental challenge in recommender systems because newly introduced items lack interaction histories required for collaborative filtering. To address this problem, this study proposes a knowledge distillation-based latent feature transfer framework that utilizes collaborative knowledge learned from existing interactions. Specifically, a teacher model learns item representations from historical interactions, while a student model learns to approximate these representations using item metadata. This design enables the generation of latent representations for new items without requiring their interaction histories. Experiments on MovieLens 1M and 20 Newsgroups demonstrate improvements over popularity-based recommendation in both datasets. In particular, the proposed framework achieves Recall@10 of 0.0648 on MovieLens 1M, representing an 85.7% relative improvement over the baseline. These findings demonstrate the potential of transferring collaborative latent representations through item metadata to alleviate the item cold-start problem.

키워드