Effect of Data Representation Method for Effective Mining of Time Series Data

摘要

In recent days, generation, collection and storage of large volume of data is increasing due to enhanced sensor and memory technology. Analysis and mining of these big data in various real life problems from health care to disaster prevention is also becoming possible with the ever increasing computational power and maginificant growth of powerful machine learning algorithms. The dynamics of natural or man made disasters can be captured by time series data and their efficient analysis can lead to the development of effective systems to minimize the loss or damage. Efficient techniques and algorithms for time series analysis is a challenging problem. Traditional machine learning algorithms for analysis of static data cannot be directly applied for analyzing dynamic time series data. For analysis of time series, representation of data and the comparison method are very important. Various methods for representation and similarity measurement of time series data have been proposed. In this paper, a comparative study has been done by simulation experiments with 85 bench mark data sets as a first step to study the effect of representation of data for mining task. The main focus of this paper is to study the effect of classification with raw data using different similarity measures with traditional classifiers and convolutional neural network and transformed data in to recurrence plot, classified by convolutional neural network. From the simulation results, it is found that classification accuracy with convolutional neural networks in many data sets are improved with time series represented by recurrence plot.