Data Visualization with Simultaneous Feature Selection

Abstract

Abstract-Data visualization algorithms and feature selection techniques are both widely used in bioinformatics but as distinct analytical approaches. Until now there has been no method of deciding feature saliency while training a data visualization model. We derive a generative topographic mapping (GTM) based data visualization approach which estimates feature saliency simultaneously with the training of the visualization model. The approach not only provides a better projection by modeling irrelevant features with a separate noise model but also gives feature saliency values which help the user assess the significance of each feature. We compare the quality of the projection obtained using the new approach with the projections from traditional GTM and self-organizing maps (SOM) algorithms. The results obtained on a synthetic and a real-life chemoinformatics dataset demonstrate that the proposed approach successfully identifies feature significance and provides coherent (compact) projections.

Recent advances and emerging challenges of feature selection in the context of big data

“…Another is the use of feature selection to improve decision trees -representing agents simulating personnel in an organization so as to model sustainability behaviours-through an expert review of their theoretical consistency [157]. Yet another is a generative topographic mapping-based data visualization approach that estimates feature saliency simultaneously as the visualization model is trained [158]. Krause et al [159] describe a tool in which visualization helps users develop a predictive model of their problem by allowing them to rank features (according to predefined scores), combine features and detect similarities between dimensions.…”

Novel visualization methods for protein data

“…To calculate feature saliency with GTM, it is assumed that features are conditionally independent given the mixture component label [24]. Specifically for a mixture of Gaussians such independence can be achieved using diagonal covariance matrices.…”

“…Instead, we propose to analyse the electrostatic potentials calculated at a fine grid by projecting the dataset onto a low-dimensional space combined with interactive data exploration techniques so that humans can interpret easily a large set of proteins. In the domain of pattern recognition various dimensionality reduction techniques, such as principal component analysis (PCA) [21], projection pursuit [22] and factor analysis [23], have been used in different domains with some success [24]. Dimensionality reduction approaches based on variance, such as PCA, do not provide good clustering or grouping information because certain features with large variance can dominate the actual grouping of the data.…”

“…However, GTM worked better than other dimensionality reduction techniques in the field of bioinformatics as shown in [33] [34] [35]. In [24], an algorithm for GTM with simultaneous feature selection was proposed (GTM-FS), which projects the data and also computes the saliency of each feature to help the user determine the importance of each feature. Both GTM and GTM-FS may fail when applied to high-dimensional datasets (with more than 200 dimensions) partially due to numerical problems.…”

Generalized topographic block model

“…Moreover, generative models for self-organizing maps has been justified [28,29,30]. The Generative Topographic Mapping (GTM) [31] is a probabilistic model of SOM for data visualisation [32,33,34,35]. In GTM the auto-organization of the clusters is directly induced by the parameterization.…”