PORTRAIT: Holistic Data Visualization using Neural Networks

Abstract

With the exponential growth of complex data across domains, effective visualization has become increasingly crucial for understanding relationships hidden within high-dimensional spaces. However, existing visualization techniques often struggle to effectively capture and represent such high-dimensional data. Motivated by this challenge, we have developed NeuroDAVIS, a neural network model designed to visualize high-dimensional data by extracting meaningful latent representations through deep feature extraction. While NeuroDAVIS has successfully addressed the visualization aspect, we have soon recognized the necessity of identifying the most relevant features that contribute to the visualization and downstream analysis. To address this issue, we have extended our framework and introduced NeuroDAVIS-FS, a feature selection model built upon the NeuroDAVIS architecture. This extension not only visualizes but also selects the most informative features from complex datasets. As data collection techniques have evolved, it has become increasingly common to encounter datasets that capture multiple aspects—or modalities—of the same set of samples. Visualizing and understanding such multi-modal datasets require a more advanced approach that is capable of integrating and producing a joint embedding of these heterogeneous data sources. Inspired by this need, we have enhanced our model to handle multiple modalities, resulting in NeuroMDAVIS, an extended version of NeuroDAVIS, capable of joint visualization across modalities. In addition, we have also developed a feature selection framework that operates on multi-modal data, enabling the identification of key features across different data types. These key features have been found to be effective in predicting the survival of lung cancer patients. Despite these advancements, another significant limitation remained—while the learned embeddings provided meaningful visualization and feature selection, they did not have generalization capabilities. Some regions of the embedding space remained sparse or unused, limiting the model’s ability to generate new, realistic samples. In order to overcome this issue, we have further extended our approach to a generative framework, giving rise to G-NeuroDAVIS. This generative variant produces a smooth, well-distributed latent embedding that is capable of not only capturing the complete data distribution but also generating high-quality.