DINE: Dimensional Interpretability of Node Embeddings

Graph representation learning methods, such as node embeddings, are powerful approaches to map nodes into a latent vector space, allowing their use for various graph learning tasks. Despite their success, these techniques are inherently black-boxes and few studies have focused on investigating local explanations of node embeddings for specific instances. Moreover, explaining the overall behavior of unsupervised embedding models remains an unexplored problem, limiting global interpretability and debugging potentials. We address this gap by developing human-understandable explanations for latent space dimensions in node embeddings. Towards that, we first develop new metrics that measure the global interpretability of embeddings based on the marginal contribution of the latent dimensions to predicting graph structure. We say an embedding dimension is more interpretable if it can faithfully map to an understandable sub-structure in the input graph - like community structure. Having observed that standard node embeddings have low interpretability, we then introduce
Dine
(Dimension-based Interpretable Node Embedding). This novel approach can retrofit existing node embeddings by making them more interpretable without sacrificing their task performance. We conduct extensive experiments on synthetic and real-world graphs and show that we can simultaneously learn highly interpretable node embeddings with effective performance in link prediction and node classification.