Hy-ClustRec: Deep Cluster-Guided Rule Mining for Cold-Start Recommendation

Ava Faghihi, Navid Khaledian, Tegawendé F. Bissyandé

Keywords: Cold-Start Problem, Data Sparsity, Recommender Systems, Clustering, Association Rule Mining, Hybrid Model

TL;DR: We introduce Hy-ClustRec, a framework that uses deep clustering to guide hierarchical association rule mining for effective cold-start recommendation.

Abstract: Recommender systems are critical for navigating vast item catalogs but struggle with the cold-start problem, where a lack of interaction data degrades recommendation quality. While hybrid methods exist, they often fail to effectively structure item content or extract fine-grained preference patterns. In this paper, we introduce \textbf{Hy-ClustRec}, a novel three-stage framework designed to address these challenges. First, we learn dense, non-linear representations of item content using a deep autoencoder. Second, these embeddings are segmented into meaningful communities using HDBSCAN, a density-based clustering algorithm. Third, we employ a hierarchical strategy for Association Rule Mining (ARM) to discover global and specialized co-occurrence patterns. Candidate items are then ranked using a hybrid scoring function that fuses rule confidence, semantic similarity from SBERT embeddings, and a user-cluster affinity score. To further boost performance, we incorporate an item-based KNN recommender into the final score with a weighted sum. Evaluated on a sparse subset of the Million Song Dataset, Hy-ClustRec demonstrates strong performance, proving especially effective in cold-start scenarios. Our work shows that a structured pipeline combining deep clustering with hierarchical rule mining and collaborative signals offers a robust solution to the cold-start problem.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 24684