A very concise feature representation for time series classification understanding

摘要

One major problem of time series analysis, particularly of a multivariate time series, is to find their feature representations. Especially, with the emerging of deep recurrent neural networks (RNNs), researchers opt to train the networks with raw signals by using an end-to-end framework to achieve the highest classification accuracy. Their works focus on modifying the network models and fine-tuning millions of hyperparameters; however, they lack the required level of understanding of the intrinsic properties of the data. In our work, we adopted a technique for dimensionality reduction of non-time-series to transform the time series data into small sets of feature representations. Our proposed technique allows the analyst to easily visualize the feature representations of the data and detect an instance which has a potential to cause a test failure. We demonstrated the robustness of our technique by subjecting the extracted features to a conventional classification approach such as Random Forest. The datasets used for the evaluation of this task are from the known benchmarking of 15 multivariate time series datasets and two Motion Caption datasets of 27 and 65 actions. The classification results were compared with the outputs from the Echo State Networks (ESNs) and the deep Bidirectional Neural Networks (BRNNs).