A Deep Learning of Time Series for Efficient Analysis

Abstract

The main problem for analyzing time series data with machine learning techniques such as classification and clustering is that a high-dimensional nature of this kind of data can cause computational difficulty in finding optimal solution. Currently, advanced learning strategy such as deep learning has been used extensively and effectively to improve learning performance. In this research, we propose a method to optimize time series analysis by adding a pre-training and fine-tuning process of deep learning based on Deep Belief Networks and Restricted Boltzmann Machines. On evaluating performance of the proposed method, we use electroencephalographic, electrocardiogram, and synthetic time series data to analyze with classification task. The induced classification models are assessed with the four several metrics including cluster evaluation, purity, mean squared error, and processing time. We comparatively compare the three learning schemes: traditional neural networks, deep learning networks, and deep learning networks with added a pre-training and fine-tuning process. The results showed that all three schemes show the same performance on predicting time series data when assessed with mean squared error. For the processing time comparison, neural networks technique is slightly faster than others. But when assessed with cluster formation and purity metrics, we found that deep learning based on the concept of Deep Belief Networks and Restricted Boltzmann Machines that adds a pre-training and fine-tuning process outperforms other learning techniques.