Robust Feature Learning on Sequential Data

Abstract

A sequence can be defined as a series of related things: a series of words in natural language, symbols in biological sequences, or a time series representing a stock market or biological signal. Drugs can also be expressed as a series of symbols representing chemical properties. Values in a sequence are not assumed to be independent and identically distributed (i.i.d.), but rather have local or long-distance dependencies and hidden patterns of variable length.<br/> This dissertation proposes three methodologies to understand hidden dependencies and patterns in sequential data. First, motivated by the demand for the accurate, fast, and scalable classification of nucleotide sequences against a group of sequences, we propose a new generative classification method based on the variable order Markov model and universal probability. Second, to learn chemical features automatically without manual curation in chemical-chemical interaction (CCI), we propose an end-to-end convolutional neural network (CNN) based on Siamese architecture. Finally, to achieve robust results in quantitative structure activity relationship (QSAR) prediction, we propose comprehensive ensemble learning incorporating an<br/> individual classifier with CNNs and recurrent neural networks (RNNs).<br/> In summary, this dissertation describes methodologies for solving diverse sequential problems. Information theoretic approaches, and neural network of CNN and RNN approaches were used exquisitely, and we confirmed their robust performance.