New GRaCE Framework Creates Interpretable Graph Embeddings

Organizing and understanding relationships within vast datasets is a critical challenge in today's data-driven world. Graphs are powerful for modeling these connections, but traditional graph-based methods often incur high computational costs. Graph embedding techniques aim to reduce dimensionality while preserving structural information, yet often lack interpretability. Building on the concept of Rank Diffusion Embedding (RaDE), this work proposes GRaCE (Graph and Rank-based Contextual Embeddings). GRaCE is an entirely unsupervised framework designed to produce interpretable embeddings. It achieves this by employing robust rank-based measures to select a representative subset of nodes, which then provides meaning to the embedding dimensions. The framework has been rigorously tested across various datasets, including both textual and image collections. GRaCE consistently demonstrates superior performance compared to RaDE and standard feature descriptors in key tasks such as retrieval, classification, and clustering, even when using state-of-the-art Transformer models for feature extraction and Graph Convolutional Networks for classification.