A generic method for supervised classification of structured objects is presented. The approach induces a classifier by (i) deriving a surrogate dataset from a pre-classified dataset of structured objects, by segmenting them into pieces, (ii) learning a model relating pieces to object-classes, (iii) classifying structured objects by combining predictions made for their pieces. The segmentation allows to exploit local information and can be adapted to inject invariances into the resulting classifier. The framework is illustrated on practical sequence, time-series and image classification problems. 1.

Sequence classification has a broad range of applications such as genomic analysis, information retrieval, health informatics, finance, and abnormal detection. Different from the classification task on feature vectors, sequences do not have explicit features. Even with sophisticated feature selection techniques, the dimensionality of potential features may still be very high and the sequential nature of features is difficult to capture. This makes sequence classification a more challenging task than classification on feature vectors. In this paper, we present a brief review of the existing work on sequence classification. We summarize the sequence classification in terms of methodologies and application domains. We also provide a review on several extensions of the sequence classification problem, such as early classification on sequences and semi-supervised learning on sequences. 1.

Abstract. Data streams are pervasive in many modern applications, and there is a pressing need to develop techniques for their efficient management. In this paper we consider real-valued streams and deal with the problem of reporting in real-time all the instants in which their distance falls below a given threshold. Current distance measures, such as Euclidean and Dynamic Time Warping (DT W), either are inaccurate or are too time-consuming to be applied in a streaming environment. We propose SDT W, a novel DT W-like distance measure which can be continuously updated in constant time and experimentally show that it improves over DT W by orders of magnitude without sacrificing accuracy. 1

Abstract. Multimedia data is ubiquitous and is involved in almost every aspect of our lives. Likewise, much of the world’s data is in the form of time series, and as will be shown, many other types of data, such as video, image, and handwriting, can be transformed into time series. This fact has fueled enormous interest in time series retrieval in the database and data mining community. However, much of this work’s narrow focus on efficiency and scalability has come at the cost of usability and effectiveness. In this work, we explore the utility of the multimedia data transformation into a much simpler one-dimensional time series representation. With this time series data, we can exploit the capability of Dynamic Time Warping, which results in a more accurate retrieval. We can also use a general framework that learns a distance measure with arbitrary constraints on the warping path of the Dynamic Time Warping calculation for both classification and query retrieval tasks. In addition, incorporating a relevance feedback system and query refinement into the retrieval task can further improve the precision/recall to a great extent. 1

Abstract. Continuously monitoring through time the correlation/distance of multiple data streams is of interest in a variety of applications, including financial analysis, video surveillance, and mining of biological data. However, distance measures commonly adopted for comparing time series, such as Euclidean and Dynamic Time Warping (DT W), either are known to be inaccurate or are too time-consuming to be applied in a streaming environment. In this paper we propose a novel DT W-like distance measure, called SDT W, which, unlike DT W, can be efficiently updated at each time step and experimentally show that it improves over DT W by orders of magnitude without sacrificing accuracy. For instance, with a sliding window of 512 samples, SDT W is 400 times faster than DT W. 1.

Abstract — In this work we introduce the new problem of finding time series discords. Time series discords are subsequences of longer time series that are maximally different to all the rest of the time series subsequences. They thus capture the sense of the most unusual subsequence within a time series. While discords have many uses for data mining, they are particularly attractive as anomaly detectors because they only require one intuitive parameter (the length of the subsequence) unlike most anomaly detection algorithms that typically require many parameters. While the brute force algorithm to discover time series discords is quadratic in the length of the time series, we show a simple algorithm that is 3 to 4 orders of magnitude faster than brute force, while guaranteed to produce identical results. We evaluate our work with a comprehensive set of experiments on electrocardiograms and other medical datasets.

I hereby declare that I am the sole author of this thesis. I authorize the University of Waterloo to lend this thesis to other institutions or individuals for the purpose of scholarly research.

Summary. Classification of time series is an important task with many challenging applications like brain wave (EEG) analysis, signature verification or speech recognition. In this paper we show how characteristic local patterns (motifs) can improve the classification accuracy. We introduce a new motif class, generalized semi-continuous motifs. To allow flexibility and noise robustness, these motifs may include gaps of various lengths, generic and more specific wildcards. We propose an efficient algorithm for mining generalized sequential motifs. In experiments on real medical data, we show how generalized semi-continuous motifs improve the accuracy of SVMs and Bayesian Networks for time series classificiation.

In previous works, a time series classification system has been presented. It is based on boosting very simple classifiers, formed only by one literal. The used literals are based on temporal intervals.

Abstract not found