| 中文摘要 |
This work is an extended version of the paper published by Ito and Chakraborty (2019). Time Series Classification (TSC) is gaining importance in the area of pattern recognition, as the availability of time series data has been increased recently. TSC is a complicated problem because of needs to consider the characteristics of temporal data; periodicity, time correlation, elasticity and unequal lengths of the time series. As all of those characteristics are usually not expressed simultaneously in raw data, design of a unified similarity metric for time series classification or clustering is difficult to achieve. In addition to traditional feature-based, model-based or distance-based algorithms for TSC, ensemble and deep neural network have been proposed recently, and deep neural network model like ResNet is known to be quite effective. However, deep neural network model requires enormous computing resources and computing time as well as large number of training samples. Feature based and distance based approaches till have potential to outperform them in computational time with reasonable classification accuracy. In this work, new temporal data transformation algorithms have been proposed and their combination with nearest neighbor classifier have been compared to existing time series classification methods. From the experimental results, the proposed algorithms with nearest neighbor classifier are found to be inferior to ResNet regarding classification accuracy though comparable to Dynamic Time Warping (DTW) but the computation is much faster than ResNet and DTW, and also the classification accuracy is better in case of small datasets which seems to be important for many real life applications with limited resources. |
本站僅提供期刊文獻檢索。
