Graph Convolutional Metric Learning for Recommender Systems in Smart Cities

Aiming for creating a convenient and intelligent living environment for city dwellers, recommender systems have been used to provide consumers with personalized suggestions when shopping or using public services. Among these systems, the metric learning (MtL)-based method often employs the Euclidean distance between user and item representations to measure the user’s preference for items in a latent semantic space. Compared to the dot product strategy, the MtL strategy better satisfies the triangle inequality principle and can simultaneously learn not only user-item similarity (users’ preference for items), but also user-user and item-item similarities. Although MtL methods have been successful, 1) the indirect high-order relationships between users and items are rarely explored; 2) the pairwise point push-and-pull optimisation strategy alone is not quite efficient in learning data point similarities. To alleviate these issues, we propose a graph convolutional(GC)-based MtL model, named GCCML. Concretely, we first employ a geometric midpoint-oriented GC strategy to generate additional points in the sparse latent space and establish the alignment of neighborhood relationships for semantic information enrichment. We then build a distance-aware neighbor weight assignment for model warm-up training. We conduct necessary experiments on three publicly datasets to validate the effectiveness of our model. The experimental results indicate that our GCCML model can outperform several state-of-the-art models under several metrics.