Graph‑Enhanced Node Representation for Cold‑Start Recommendation: Neighbour‑Enhanced YouTubeDNN

Cold‑start issues for users and items remain a major challenge in recommendation systems because traditional behavior‑driven models lack sufficient historical signals. To address this, we introduce static attribute graphs and dynamic interaction graphs, and design a graph‑adjacency‑enhanced node representation that better models user and item embeddings.

The overall architecture, called Neighbour‑Enhanced‑YouTubeDNN, extends the classic YouTubeDNN by adding two neighbour‑enhancement modules (one for users, one for items). Each module first applies node‑level attention using a Graph Attention Network (GAT) to aggregate adjacent node embeddings, then applies neighbor‑level attention to fuse static and dynamic graph information, producing enriched embeddings.

Formally, for a target node \(h_i\), we compute attention‑weighted sums over its neighbours \(h_j\) with learnable parameters \(W, q, b\). Softmax normalisation yields attention coefficients \(a_{ij}\), and the final node representation \(Z\) is obtained after the two‑stage attention process.

Experiments on the public MovieLens‑1M dataset, after sparsifying the data to simulate cold‑start scenarios, show that the Neighbour‑Enhanced‑YouTubeDNN consistently improves AUC by about 0.5 % across three sparsified splits, with the most significant gains on the sparsest subset.

Key innovations include (1) the construction of both static and dynamic relational graphs to fully exploit item/user side information, and (2) a multi‑level attention mechanism that adaptively balances static attribute importance and dynamic interaction relevance, leading to superior embedding quality.

In conclusion, leveraging graph‑based relational information markedly enhances user/item representations for cold‑start recommendation, and the proposed framework can be readily integrated into existing large‑scale recommender pipelines. The source code is available at https://github.com/xuanjixiao/onerec.