Improving Cold-Start User Recommendation via LLM-Judged Signals

We present a lightweight, profile-agnostic framework that improves recommendations for cold users with sparse interaction histories, without relying on side information. Our framework leverages LLMs to enhance user representations. Given an item and the past interactions of two users, the LLM predicts which user is more likely to prefer the item. These signals are incorporated into the training process to enhance user representations and thereby improve recommendation performance. Experiments on the Amazon Review Dataset-specifically the All Beauty and Appliances category subsets-show consistent gains in Recall and Hit Rate for cold users, while maintaining stable performance for warm users. Compared to the nonaugmented baseline, our method achieves substantially better performance. This method is especially relevant for services that continually attract new users, such as e-commerce and streaming platforms. It also offers a practical path toward cold-user recommendation without user profiles.