Accurate and fast time series classification based on compressed random Shapelet Forest

Achieving accurate, fast, and interpretable time series classification (TSC) has attracted considerable attention from the data mining community over the past decades. In this paper, we propose an efficient algorithm, called Compressed Random Shapelet Forest (CRSF), to tackle this problem. Different from most of the shapelet-based TSC methods, CRSF obtains promising performance by greatly compressing the shapelet features space. In order to achieve the aim of compression, the time series dataset, as well as the shapelets, are represented by Symbolic Aggregate approXimation (SAX) at first. Then, the shapelet-based decision trees are built upon a pool of high-quality shapelet candidates of which the useless shapelets and the self-similar shapelets have been pre-pruned. A new function for measuring the distance between two SAX-represented time series is also introduced. Extensive experiments were conducted on 50 UCR time series datasets. The results show that (1) CRSF can achieve the highest average accuracy on the datasets and it outperforms most of the existing shapelet-based TSC methods; (2) CRSF is slightly superior to gRSF in terms of accuracy and is significantly superior to gRSF in terms of time cost. Specifically, it is on average 41 times faster than gRSF according to the experimental results.