New Framework Creates Context-Aware Interpretable AI Representations

Despite significant advancements in visual information modeling through CNNs, Transformers, and Foundation Models, challenges persist regarding similarity assessment and model transparency. Specifically, traditional pairwise measures often fail to capture the true intrinsic geometry of datasets (the "Geometric Gap"), and representations frequently lack alignment with human understanding (the "Interpretability Gap"). To address these issues, this research introduces a novel unsupervised framework. It combines Manifold Learning strategies with Rank-based Interpretable Graph Embeddings. The framework first characterizes the contextual information of a dataset through manifold analysis, then generates sparse, self-explainable embeddings. The proposed approach is flexible, allowing for different Manifold Learning and Representation Learning strategies. Extensive evaluations across diverse datasets and features demonstrate that these Context-Aware representations not only reduce dimensionality and provide intrinsic interpretability but also maintain or improve effectiveness in downstream tasks like image retrieval and semi-supervised classification using Graph Convolutional Networks.