Topological Activation Maps for Visual Representation Learning from Tabular Data

This chapter introduces Topological Activation Maps (TAMs), a novel framework designed to transform tabular data into visual representations that can be effectively processed by vision models. The method decouples global topology from sample-specific activations, allowing for a more nuanced and accurate representation of data. The chapter delves into the methodology behind TAMs, including the use of kernel space embedding and Self-Organizing Maps (SOM) to create a robust topological layout. It also explores how each data sample is transformed into a unique activation map, preserving both dataset-level manifold structure and individual sample characteristics. The evaluation section presents a comprehensive comparison of TAMs with existing methods across 11 UCI benchmark datasets, highlighting its superior performance in various metric-dataset combinations. The results demonstrate TAMs' robustness, particularly on imbalanced datasets, and its ability to capture minority class structure effectively. The discussion section provides insights into the theoretical grounding of TAMs and its potential applications beyond classification tasks. Overall, this chapter offers a detailed overview of TAMs, its methodology, evaluation, and potential impact on visual representation learning from tabular data.