Abstract Neural networks are widely used in machine learning and data mining. Typically, these networks need to be trained, implying the adjustment of weights (parameters) within the network based on the input data. In this work, we propose a novel approach, RandomNet, that employs untrained deep neural networks to cluster time series. RandomNet uses different sets of random weights to extract diverse representations of time series and then ensembles the clustering relationships derived from these different representations to build the final clustering results. By extracting diverse representations, our model can effectively handle time series with different characteristics. Since all parameters are randomly generated, no training is required during the process. We provide a theoretical analysis of the effectiveness of the method. To validate its performance, we conduct extensive experiments on all of the 128 datasets in the well-known UCR time series archive and perform statistical analysis of the results. These datasets have different sizes, sequence lengths, and they are from diverse fields. The experimental results show that the proposed method is competitive compared with existing state-of-the-art methods.

The discovery of motifs (repeated patterns) is an important task in time series data mining. The task can be formulated as finding the most similar non-overlapping pair of subsequences in a given time series. Existing exact motif discovery methods have quadratic time complexities in the length of the time series. In this work, we present an algorithm that can find the exact motif of a given time series in a linear expected time complexity. The algorithm is further modified to find all pairs of subsequences whose distances are below a given threshold value. In practice, if true motifs exist in the data or the threshold is set to a small value, the algorithms are very fast. The proof of correctness and time complexity are detailed and experiments are conducted to verify the analysis. We applied the proposed method to analyze the real-world bird sound and electrical consumption data to demonstrate its effectiveness.

The mining of time series data has attracted much attention in the past two decades due to the ubiquity of time series in our daily lives. In particular, classification is perhaps one of the most well-studied topics for time series data. Many state-of-the-art classification techniques work by identifying and extracting patterns or characteristics from the training data, and then applying these patterns or characteristics to classify unlabeled time series. This paper presents a novel finding that sequences of values that are very different from the patterns in the labeled time series can be used as references to classify time series effectively. We propose an evolution process to generate these sequences of values, which we call separating references, from the training data. The proposed method is robust to over-fitting and is especially suitable for the situation where little labeled data is available. We demonstrate that the proposed approach is highly competitive on the well-known UCR time series classification benchmarks.

Time series classification has attracted much attention due to the ubiquity of time series. With the advance of technologies, the volume of available time series data becomes huge and the content is changing rapidly. This requires time series data mining methods to have low computational complexities. In this paper, we propose a parameter-free time series classification method that has a linear time complexity. The approach is evaluated on all the 85 datasets in the well-known UCR time series classification archive. The results show that the new method achieves better overall classification accuracy performance than the widely used benchmark, i.e. 1-nearest neighbor with dynamic time warping, while consuming orders of magnitude less running time. The proposed method is also applied on a large real-world bird sounds dataset to verify its effectiveness.