CONTEXT-GATED GRAPH NEURAL NETWORKS FOR RECOMMENDER SYSTEMS

Recommender systems increasingly operate in scenarios where situational factors such as time, location, and social setting shape user preferences, motivating the development of Context-Aware Recommender Systems (CARS). While Graph Neural Networks (GNNs) have demonstrated strong performance in collaborative filtering by capturing high-order connectivity patterns between users and items, their integration with contextual information remains largely underexplored: existing GNN-based recommenders typically treat context, when available, as an auxiliary feature or a post-hoc scoring adjustment rather than a first-class signal that shapes representation learning itself. This thesis addresses this gap by proposing a context-aware graph-based recommender derived from LightCCF that embeds discrete contextual features directly into interaction edges and uses the resulting representations to gate message passing between users and items. Unlike approaches that inject context only at the scoring stage, the proposed model conditions the propagation process itself: each contextual feature is mapped to a dedicated embedding, projected into a shared latent space, and used as a multiplicative filter that selectively scales neighbour information during aggregation. At inference time, an explicit context-item compatibility term is added to the learned user-item affinity, and the model is trained with a multi-task objective combining Bayesian Personalised Ranking, a Neighbourhood Aggregation loss, and L2 regularisation. This design preserves the efficiency and linearity of LightGCN-style propagation while making the learned embeddings sensitive to the situational conditions under which each interaction occurs. s a further contribution, this work introduces a context-aware negative sampling strategy for the training phase, replacing uniform random negative sampling with a probabilistic scheme that favours, but does not deterministically select, negatives whose context closely resembles that of the positive interaction. Contextual similarity is computed as the complement of the Hamming distance over discrete context features, and candidate negatives are drawn from a temperature-controlled softmax over this similarity, ensuring that contextually-similar negatives are sampled more frequently while still preserving exploration over the full candidate pool. A structural analysis of context density across datasets shows that the effectiveness of this strategy is closely tied to how densely and repeatedly contextual profiles occur in the training data. The model and the sampling strategy are implemented within WarpRec, an open-source recommendation framework, and evaluated on three benchmark datasets spanning distinct domains and context structures: BoardGameGeek, Frappe and Yelp. The proposed model is compared against non-contextual baselines and five factorisation-based context-aware models, using standard ranking metrics as well as a suite of context-sensitive evaluation metrics developed in prior work, employed here solely as evaluation instruments. Experimental results show that the proposed model achieves the best standard ranking performance on Yelp and Frappe, and remains competitive on BoardGameGeek despite a constrained embedding budget imposed by memory limitations. The proposed negative sampling strategy yields consistent gains on datasets with dense, frequently repeated contextual profiles (BoardGameGeek, Frappe), while its effect is comparatively weaker on Yelp, where the extreme sparsity of the context space limits the reliability of contextual similarity as a training signal. Taken together, these findings indicate that embedding context directly into graph propagation, and into the negative sampling process that shapes training signal, is a viable and lightweight strategy for context-aware recommendation, whose benefits are governed by the density and informativeness of the underlying context space rather than by architectural complexity.