LESS IS MORE: IMPROVING THE PERFORMANCE OF LEARNING MODELS WITH FEWER FEATURES OR FEWER PARAMETERS

Abstract

The growing number of features and parameters does not only increase the risk of overfitting, but also hinders the application of learning models in many resource-constrained domains. In this thesis, we address this issue by exploring methods which enable learning models to obtain comparable or better performance with a significantly reduced number of features or parameters. Specifically, we study and propose three methods: feature selection, neural network pruning and time series forecasting, where the first method addresses the issue of growing number of features and the latter two methods address the issue of growing number of model parameters.