What question did this study set out to answer?

This work aims to address the cold start problem in recommender systems using a hybrid framework.

June 10, 2026Open Access

Retrieval-Augmented Generation for Mitigating Cold Start Problems in Recommender Systems: A Hybrid Semantic Retrieval Framework

JJJ.Princy J.PrincyCoimbatore Medical College and Hospital

Key Points

This work aims to address the cold start problem in recommender systems using a hybrid framework.
Developed a Retrieval-Augmented Generation (RAG) framework combining semantic vector retrieval and collaborative filtering.
Utilized Sentence-BERT for generating 384-dimensional dense vector embeddings of movies.
Implemented testing on the MovieLens dataset with evaluation metrics including Precision@5, Recall@5, and NDCG@5.
The RAG component achieved NDCG@5 of 0.339 compared to 0.0 for Collaborative Filtering alone.
Collaborative Filtering achieved perfect Precision@5 of 1.0 for warm-start users.
The hybrid framework effectively mitigates cold start challenges while ensuring meaningful recommendations.

Abstract

The cold start problem is one of the most persistent issues in contemporary recommender systems. In operational contexts where users or items being recommended are new, the request volume can reach between 30% and 40%. In other words, it is difficult for collaborative filters to generate recommendations for users or items without any sort of interaction history. Matrix factorization and neighbourhood-based collaborative filtering methods are traditional recommender systems and are essentially dependent on the historical user-item rating scores given by users. Thus, first-time users do not have any rating score and hence, no meaningful recommendation can be generated for this first-time user. This work proposes a Retrieval-Augmented Generation (RAG) framework combining semantic vector retrieval and item-based collaborative filtering to mitigate cold start issues without compromising warm-start user recommendations. The proposed system is tested on the MovieLens dataset which involves 9, 742 movies and user interaction records. The combination of titles and genre metadata makes the textual representations of a movie. It is converted into 384-dimensional dense vectors using the Sentence-BERT model (all-MiniLM-L6-v2). This pre-trained model has been trained on over 215 million pairs of sentences. These embeddings are stored in a Facebook AI Similarity Search (FAISS) vector database that allows sub-linear nearest-neighbor search via L2 Euclidean distance. For warm-start users, we generate recommendations for users from the user-movie rating matrix by item-to-item collaborative filtering, based on cosine similarity. The two strategies are mixed using a configurable interpolation parameter α ∈ 0, 1. recommendations = α × RAGₛcores + (1−α) × CFₛcores. A report with 100 test users using metrics of Precision@5, Recall@5, and NDCG@5 shows that the NDCG@5 of the RAG component is 0. 339 vs the 0. 0 of Collaborative Filtering alone and CF achieves perfect Precision@5 of 1. 0 for warm-starts. The hybrid framework alleviates cold start challenges and provides semantically coherent recommendations from natural language queries even in the complete absence of user-profile data. We validated the strengths of deep semantic retrieval and collaborative pattern learning, and propose a scalable architecture for production deployment in real-time.

Ask AI

Helpful

Bookmark

View Full Paper