A 3D playground for t-SNE with explainable classification

Abstract

In this thesis, we build a web-based and interactive dimensionality reduction and visualization tool based on t-SNE. Our tool strives to run in the web browser at the client-side in real-time. In addition, we propose a new feature for linking visualization and explainable classification of a given dataset. Assuming the dataset is subject to classification by a machine learning algorithm, we encode the predictions of the classifier using the color of the datapoints and highlight the explanation of each prediction as given by an agnostic interpretation algorithm. Our contribution is three-folds: (1) We provide an online web playground that offers guidance on how to use t-SNE and depicts the effect of each of its parameters, (2) we present an interactive visualization interface where users can upload their own data, link the data items to their icons and to their datapoints in 3D using ThreeJS, navigate the 3D space equipped with a camera, click on the data points to get instantaneous meta-data information, prediction explanations in particular, and (3) we create a visual link between explainable classification and t-SNE experiment with explanation methods such as LIME. This feature helps in checking the efficiency of the classifier. The user checks whether points that are close together in the visual space have similar explanations for a given classifier. The user can spot outliers and the cause of erroneous decisions. We estimate that our tool will help with the arising challenge of trustworthy and explainable artificial intelligence (XAI) by providing the user with direct access to the cluster-ability patterns of the data in low dimensions and explore the peculiarities of machine learning models, neural networks, and deep learning in particular.