DeepCommerceRec: a graph- and sequence-aware deep learning recommender system with cold-start adaptation for large-scale e-commerce platforms

For large-scale e-commerce, modelling long-term collaborative preferences, short-term behavioural intent, and sparse interactions needs to be done simultaneously. Existing graph-based methods can represent structural user–item relationships well, while sequential models capture the temporal evolution of user behaviour; however, many methods cannot adaptively coordinate these complementary signals, and few incorporate cold-start handling within the same ranking framework. To overcome this drawback, the authors propose a graph- and sequence-aware deep learning framework (DeepCommerceRec) that integrates cold-start adaptation (CSA) and context-conditioned fusion (CCF). This structure uses a dual-encoder design: heterogeneous graph representation learning to learn structural user representations and sequential behavioural modelling to learn temporal user representations. These are fused adaptively, depending on interaction availability, sequence length, behavioural recency, and cold-user/cold-item status. In a sparse setting, we build metadata- and neighbourhood-aware representations for cold items and leverage few-shot recent interactions to adapt cold-user representations. Experiments on three public e-commerce datasets show that DeepCommerceRec outperforms traditional, sequential, graph-based, and more recent hybrid baselines on HR@10, NDCG@10, and MRR. Repeated runs and statistical analysis confirm that the observed improvements are robust, and evaluation under mixed warm–cold and cold-start settings shows improved ranking performance in sparse interaction scenarios. Ablation and sensitivity analyses further validate the contributions and stability of the proposed components. These gains are demonstrated via computational evaluation under the tested benchmark conditions with acceptable inference overhead. In summary, DeepCommerceRec offers a single framework for context-adaptive structural–temporal recommendation and incorporates cold-user and cold-item treatment. Not applicable.