DESIGN AND IMPLEMENTATION OF LEARNING INTERPRETABLE FEATURES OF GRAPHS AND TIME SERIES DATA

DESIGN AND IMPLEMENTATION OF LEARNING INTERPRETABLE FEATURES OF GRAPHS AND TIME SERIES DATA

Abstract

Graphs and time series are two of the most ubiquitous representations of data of modern time. Representation learning of real-world graphs and time-series data is a key component for the downstream supervised and unsupervised machine learning tasks such as classification, clustering, and visualization. Because of the inherent high dimensionality, representation learning, i.e., low dimensional vector-based embedding of graphs and time-series data is very challenging. Learning interpretable features incorporates transparency of the feature roles, and facilitates downstream analytics tasks in addition to maximizing the performance of the downstream machine learning models. In this thesis, we leveraged tensor (multidimensional array) decomposition for generating interpretable and low dimensional feature space of graphs and time-series data found from three domains: social networks, neuroscience, and heliophysics. We present the theoretical models and empirical results on node embedding of social networks, biomarker embedding on fMRI-based brain networks, and prediction and visualization of multivariate time-series-based flaring and non-flaring solar events.

DESIGN AND IMPLEMENTATION OF LEARNING INTERPRETABLE FEATURES OF GRAPHS AND TIME SERIES DATA. GET MORE COMPUTER SCIENCE PROJECT TOPICS AND MATERIALS

Sharing is caring!