DemoGCL: Demographic-augmented graph contrastive learning for cold-start recommendation

One of the most persistent problems in e-commerce recommendation is what happens when a new user signs up or a new product goes live the system knows almost nothing about them, and graph-based collaborative filtering, which depends entirely on interaction history, produces embeddings that are essentially random. Existing approaches tend to address cold users and cold items as two separate problems, often at the cost of serving-time language model inference that is far too slow for production systems. In this work, we present DemoGCL, a unified Demographic-Augmented Graph Contrastive Learning framework that resolves both cold-start scenarios within a single lightweight model. At its core, DemoGCL builds on a three-layer LightGCN backbone for warm collaborative signal and augments it with two dedicated branches: a two-layer MLP that projects a five-dimensional behavioral proxy vector derived from publicly observable demographic signals into the shared embedding space for cold users who have no purchase history, and a LoRA-adapted DistilBERT encoder (rank 8, 332K trainable out of 66.7M parameters, or 0.50%) that generates cold-item embeddings from product text alone. A sigmoid-gated fusion layer harmonizes the warm and cold representations, and a cross-view InfoNCE objective aligns them during training so that cold embeddings land in the same semantic neighborhood as their warm counterparts. We evaluate on Amazon Beauty, Amazon Clothing, and Epinions under a sampled ranking protocol and find that DemoGCL achieves warm-split NDCG@10 of 0.0612, 0.0487, and 0.0831, cold-user NDCG@10 of 0.0543, 0.0411, and 0.0762 (retaining 88.7%, 84.4%, and 91.7% of warm performance), and cold-item NDCG@10 of 0.0388, 0.0319, and 0.0594. On Amazon Beauty, based on LLMDiRec’s published numbers under its own evaluation setup, DemoGCL improves over LLMDiRec by 0.18% on cold users and 77.2% on cold items, although the evaluation protocols differ. Ablation studies confirm that the demographic branch is the single most critical component, with its removal causing up to a 41.4% drop in cold-user NDCG@10, while dropping LoRA fine-tuning reduces cold-item performance by up to 37.9%. Because warm embeddings are precomputed and cached as lookup tables, the serving path avoids online large-language-model generation and keeps cold-user inference limited to a small projection module, which supports a lightweight deployment structure for e-commerce pipelines.