s of the Keynotes Health Data Management and Analytics with Privacy and Confidentiality

Wind turbines are typically organised as a fleet in a wind park, subject to similar, but varying, environmental conditions. This makes it possible to assess and benchmark a turbine’s output performance by comparing it to the other assets in the fleet. However, such a comparison cannot be performed straightforwardly on time series production data since the performance of a wind turbine is affected by a diverse set of factors (e.g., weather conditions). All these factors also produce a continuous stream of data, which, if discretised in an appropriate fashion, might allow us to uncover relevant insights into the turbine’s operations and behaviour. In this paper, we exploit the outcome of two inherently different discretisation approaches by statistical and visual analytics. As the first discretisation method, a complex layered integration approach is used. The DNA-like outcome allows us to apply advanced visual analytics, facilitating insightful operating mode monitoring. The second discretisation approach is applying a novel circular binning approach, capitalising on the circular nature of the angular variables. The resulting bins are then used to construct circular power maps and extract prototypical profiles via non-negative matrix factorisation, enabling us to detect anomalies and perform production forecasts.

We present PSEUDo, an adaptive feature learning technique for exploring visual patterns in multi-track sequential data. Our approach is designed with the primary focus to overcome the uneconomic retraining requirements and inflexible representation learning in current deep learning-based systems. Multi-track time series data are generated on an unprecedented scale due to increased sensors and data storage. These datasets hold valuable patterns, like in neuromarketing, where researchers try to link patterns in multivariate sequential data from physiological sensors to the purchase behavior of products and services. But a lack of ground truth and high variance make automatic pattern detection unreliable. Our advancements are based on a novel query-aware Locality-Sensitive Hashing technique to create a feature-based representation of multivariate time series windows. Most importantly, our algorithm features sub-linear training and inference time. We can even accomplish both the modeling and comparison of 10,000 different 64-track time series, each with 100 time steps (a typical EEG dataset) under 0.8 seconds. This performance gain allows for a rapid relevance feedback-driven adaption of the underlying pattern similarity model and enables the user to modify the speed-vs-accuracy trade-off gradually. We demonstrate PSEUDo’s superiority in terms of efficiency, accuracy, and steerability through a quantitative performance comparison and a qualitative visual quality comparison to the state-of-the-art algorithms in the field. Moreover, we showcase PSEUDo’s usability through a case study demonstrating our visual pattern retrieval concepts in a large meteorological dataset. We find that our adaptive models can accurately capture the user’s notion of similarity and allow for an understandable exploratory visual pattern retrieval in large multivariate time series datasets.

Time series visualization is one of the most fundamental tasks, which is often used to discover patterns by a user interface. Many increasing interests in time series visualization in the last decade have resulted in the development of various state-of-the-art visualization techniques. In most cases, time series visualization is based on one of the representation frameworks which not only reduce the dimensionality, but sufficiently reflect the shapes of the raw time series as well. In this paper, a new time series visualization based on shape features is proposed to discover surprising patterns and mine frequent trends (motifs discovery). Since the shape features validly summarize both the global and local structures of time series and often use the slopes (or the angles) of the changeable values over time to describe the trends of time series, part of the work is to extract the important shape features and transform them into symbol string. After dimensionality reduction and symbol representation, a circle plotted by the visualization technique is split into many small sectors which simultaneously represent substring patterns of unequal length by multi-resolution function. Since the overall shape features of time series are based on data point importance, the surprising patterns and frequent trends can be accurately discovered even under a high compress ratio. The refined results of patterns discovery and frequent trends mining on financial and other time series datasets indicate that the proposed approach is an effective visualization tool for time series mining.

Time series is a sequential collection of values with respect to time obtained from various applications. The time series data have basic features like huge data size, high dimensionality with characteristics like trend, cyclical, seasonal, and irregular. The cumulative use of time series data has initiated a great deal of research and development attempts in the field of data mining. Dimensionality of time series is directly proportional to the efficiency of various data mining algorithms used for time series analysis. In this paper, a widespread review on the existing time series dimensionality reduction methods is given. The chief objective of this paper is to aid interested researchers to have a general idea about the current investigation in time series dimensionality reduction methods and identify their potential research direction to advance investigation in the same. The papers also discuss about the possibilities of using automata model for time series dimensionality reduction.

Time series discretization is a technique commonly used to tackle time series classification problems. This manuscript presents an enhanced multi-objective approach for the symbolic discretization of time series called eMODiTS. The method proposed uses a different breakpoints vector, defined per each word segment, to increase the search space of the discretization schemes. eMODiTS’ search mechanism is the well-known evolutionary multi-objective algorithm NSGA-II, which finds a set of possible solutions according to entropy, complexity, and information loss estimations. Final solutions were appraised depending on the misclassification rate computed through the decision tree classifier. The trees obtained also produce graphical and significant information from the regions, relationships, or patterns in each database. Our proposal was compared against ten state-of-the-art time symbolic discretization algorithms. The results suggest that our proposal finds a suitable discretization scheme regarding classification, dimensionality, cardinality reduction, and information loss.

Time series data are generated in abundance of information technology applications. Classification of time series is one of the significant areas of interest in Time Series Data Mining. Over the last decade, different approaches have been developed to achieve a superior effectiveness and efficiency in classification of time series. We propose a new model for classification of time series which first uses SAX (symbolic aggregate approximation) method to transform the time series to symbols and then applies a probabilistic methodology in dealing with classification of symbolic sequences. This model is developed based on Bayesian rule and Probability Chain rule on symbolic representation of time series along with considering a penalty for mismatching points. We call this model Bayesian and Chain rule Model (BCM). The performance of BCM has been evaluated using 43 time series data sets. We performed an extensive experiment to compare the performance of BCM to Euclidean Distance, Dynamic Time Warping (DTW) with best warping window, and DTW with no warping window. This comparison was done through three stages. Stage 1 was predicting the domains and problems, in which BCM works relatively well. Stage 2 included measuring the prediction accuracy of BCM to those of the other methods. Stage 3 involved constructing the confusion matrixes for predictions in Stage 1 as well as measuring and comparing the predictability of the methods. The results show that, compared to the other popular time series classification methods, BCM is more accurate where it is predicted to perform better.

Time series classification is an important problem in data mining with several applications in different domains. Because time series data are usually high dimensional, dimensionality reduction techniques have been proposed as an efficient approach to lower their dimensionality. One of the most popular dimensionality reduction techniques of time series data is the Symbolic Aggregate Approximation (SAX), which is inspired by algorithms from text mining and bioinformatics. SAX is simple and efficient because it uses precomputed distances. The disadvantage of SAX is its inability to accurately represent important points in the time series. In this paper we present Extreme-SAX (E-SAX), which uses only the extreme points of each segment to represent the time series. E-SAX has exactly the same simplicity and efficiency of the original SAX, yet it gives better results in time series classification than the original SAX, as we show in extensive experiments on a variety of time series datasets.

Time series analysis is widely used in the fields of finance, medical, and climate monitoring. However, the high dimension characteristic of time series brings a lot of inconvenience to its application. In order to solve the high dimensionality problem of time series, symbolic representation, a method of time series feature representation is proposed, which plays an important role in time series classification and clustering, pattern matching, anomaly detection and others. In this paper, existing symbolization representation methods of time series were reviewed and compared. Firstly, the classical symbolic aggregate approximation (SAX) principle and its deficiencies were analyzed. Then, several SAX improvement methods, including aSAX, SMSAX, ESAX and some others, were introduced and classified; Meanwhile, an experiment evaluation of the existing SAX methods was given. Finally, some unresolved issues of existing SAX methods were summed up for future work.

Throughout recent years, dynamic time warping (DTW) has remained as a robust similarity measure in time series classification (TSC). 1-nearest neighbor (1-NN) algorithm with DTW is the most widely used classification method on time series serving as a benchmark. With the increasing demand for TSC on low-resource devices and the widespread of wearable devices, the need for a efficient and accurate time series classifier has never been higher. Although 1-NN DTW attains accurate results, it highly falls back on efficiency due to its quadratic complexity in the length of time series. In this paper, we propose a new approximation method for reducing the length of the time series as the input of DTW. We call it control chart approximation (CCA), after a similar concept used in statistical quality control processing. CCA representation approximates raw time series by transforming them into a set of segments with aggregated values and durations forming a reduced 3-D vector. We also propose an adaptation of DTW in 3-D space as a distance measure for 1-NN classifier, and denote the method as 1-NN 3-D DTW. Our experiments on 85 datasets from UCR archive—including 28 long-length (>500 points) time series datasets—show up to two orders of magnitude performance gain in running time compared to the state-of-the-art 1-NN DTW implementation. Moreover, it shows similar or better accuracy on the long time series in the experiment.

This thesis concerns the field of computational intelligence (CI), an important area of computer science that predominantly endeavours to model complex systems with heuristic algorithms. Heuristic algorithms from CI are generally nature or biologically inspired programs that iteratively learn from experience. More specifically, the research focuses on the overlap between the fields of CI and financial analysis, where the financial markets (in the form of financial time series) provide the complex system of interest. Therefore the objective of the thesis can be summarized as a exercise in improving the abilities of CI algorithms for modelling financial time series. CI has been applied to a whole spectrum of domains where techniques are developed at a more general level and then applied to a particular application area or complex system. The financial markets are somewhat unique. Unlike other complex systems in nature, the financial markets are of our own creation and their evolution is a by product of human nature, where the beliefs and bias of the participants (humans) in the complex system (financial markets) govern how the system behaves. This is in contrast to many complex systems where, for example, the opinions of experts have no effect on the outcome. Thus, the motivation of this work is to quantify meaningful characteristics (behaviour) of the financial markets as a means to improve and understand how heuristic algorithms respond to them. This process of applying more scrutiny to the analysis of the application area, yields an approach to algorithm development that takes into account the unique characteristics of the market. To achieve this goal the thesis is structured into three sections that comprise four contribution chapters. The contribution chapters are labelled: validity, implications and innovations and each is motivated by a separate research question. The validity chapter is based on determining a reasonable characterization of the financial markets. This includes a detailed literature review of the popular competing market theories as well as some new innovative tests. There is not a general consensus as to which theory is correct but from a computational intelligence perspective the adaptive market hypothesis (AMH) is revealed as a reasonable characterization of the financial markets and one that provides quantifiable characteristics to be utilized in following chapters. The implications chapter concerns testing the effect of implications of the AMH, if any, on the robustness of models derived from CI. Specifically three implications are examined, i.e., variable stationarity, variable efficiency and the waxing and waning of investment strategies. The experiments concerned six algorithms from four of the major paradigms in supervised learning. The results from each of the studies demonstrated that the implications of the AMH affect CI derived models. This conclusion reveals that the unique properties of the financial markets should be taken into account when applying CI algorithms for modelling and forecasting. The two chapters concerning innovations explore how CI techniques can be improved based on the results from the validity and implications chapters. The first chapter (chapter 6) concerns the development of a meta learner based on the implication of the waxing and waning of investment strategies, the meta learning algorithm called LATIS (Learning Adaptive Technical Indicator System) is a blend of micro and macro modelling perspectives and allows for online adaptive learning with an interpretable white box framework. The second innovations chapter (chapter 7) concerns the discretization of financial time series data into a finite alphabet. A discretization algorithm is developed, which extends an existing state-of-the-art algorithm to handle the characteristics of financial time series. The proposed algorithm, called alSAX (adaptive local Symbolic Aggregate approXimation), is demonstrated to be superior in terms of its symbolic mappings, in relation to a gold standard, and in the popular time series subsequence analysis task. Additionally, an invalid theoretical assumption of the existing algorithm is revealed. The flaw in the algorithm is discussed and its impact is determined based on the characteristics of the time series and the parameters of the algorithm. From the analysis, the thesis offers viable fixes to compensate for the flaw, where the suitability of the fixes are dependent on the problem domain and objective of the data mining task.

This chapter introduces a method that discovers characteristic sequential patterns from sequential data based on background knowledge. The sequential data is composed of rows of items. This chapter focuses on the sequential data based on the tabular structured data. That is, each item is composed of an attribute and an attribute value. Also, this chapter focuses on item constraints in order to describe the background knowledge. The constraints describe the combination of items included in sequential patterns. They can represent the interests of analysts. Therefore, they can easily discover sequential patterns coinciding to the interests of the analysts as characteristic sequential patterns. In addition, this chapter focuses on the special case of the item constraints. It is constrained at the last item of the sequential patterns. The discovered patterns are used to the analysis of cause, and reason and can predict the last item in the case that the sub-sequence is given. This chapter introduces the property of the item constraints for the last item.

There are many factors that can contribute to corrosion in the pipeline. Therefore, it is important for decision makers to analyze and identify the main factor of corrosion in order to take appropriate actions. The factor of corrosion can be analyzed using data mining based on historical datasets collected from monitoring sensors. The purpose of this study is to analyze the trends of corroding agents for pipeline corrosion based on symbolic representation of time series corrosion dataset using Symbolic Aggregation Approximation (SAX). The paper presents the analysis and evaluation of the patterns using N-gram model. Text mining using N-gram model is proposed to mine trend changes from corrosion time series dataset that are transformed as symbolic representation. N-gram was applied for the analysis in order to find significant symbolic patterns that are represented as text. Pattern analysis is performed and the results are discussed according to each environmental factor of pipeline corrosion.

The recording and analyzing human motor control movements are fundamental parts of both behavioral biometrics and biomedical research studies. The dynamics of human motor functions of fingers, hand and wrist movements can be studied while handwriting, drawing or hand gestures. The major difficulties are the acquisition of high quality data and the characterization of the acquired data accurately and efficiently. A multisensoric Biometric Smart Pen BiSP device has been developed which has the ability to measure dynamics of fingers, hand and wrist movements while handwriting drawing and hand gesturing during writing on paper pad or free in space. Hence, biometric features are recorded in terms of pen refill pressures, finger grip pressures holding the pen, and pen inclination, acceleration and tilts in three dimensions. The grip sensing of BiSP is unique and provides excellent dynamics of the fine motor skills of a writer. The features recorded by BiSP have not only been analyzed to recognize person or handwritten object in biometric applications but also to characterize motor dysfunctions of the writer due to Parkinson’s disease for diagnostics in medical applications. The major problems in data processing are automatic feature selection and classification of data accurately especially for short sequence of input (e.g., single characters). For above applications, research has been conducted that involves improved procedures in diverse sensing techniques for data acquisition, signal processing and pattern recognition for person or object recognition based on signals (time series) obtained from novel digital pen. Many data processing techniques including enhanced data pre-processing and dimension reduction methods as well as feature extraction and classification techniques have been implemented. For this, dynamic time warping (DTW) together with its variants and support vector machine (SVM) techniques are established and evaluated for data analysis. In biometrics, the improved performances of person authentication and of handwritten object recognition by using the new designed BiSP system have been obtained. For medical application, personal neuro-motoric features obtained from handwriting, drawing or gesturing movements have been analyzed in order to characterize Parkinson’s disease or to control medication. In brief, to evaluate the newly developed procedures (sensor techniques and software methods) several experiments have been performed to determine the feasibility of the BiSP system in order to register and to analyze human fine motor features in multiple applications.

The objective of this thesis is considering the safety analysis of the simplified cooling system of a single module of the ITER Central Solenoid in a cold test facility to classify the different abnormal transients and identify the components failure that may drive the system into a Loss-Of-Flow Accident. The IDPSA framework is applied for the first time to a fusion system to study the dynamic response of the system. Two different operating condition are considered: a first simplified case, in which it is assumed a constant current flowing in the module. In the second part, the real ITER current scenario is considered to study the consequence of the abnormal transients both for the safety of the facility and the consequence for ITER. In fact, the hot-spot temperature of the strands computed in this second case can be assumed close to the values reached in ITER. A set of 100 different accidental scenarios is generated according to the Multiple Value Logic and it is simulated using the validate 4C thermal-hydraulic code. To post-process the scenarios, it is applied the spectral clustering algorithm to identify the most similar scenarios grouped using the Fuzzy-C-Means. The results are compared with the ones obtained using the ESAX algorithm. The Silhouette and the DaviesBouldin indexes are used to evaluate the clusters obtained and to choose the spectral clustering as best algorithm. Its results are used to identify the prototypical components failure that may drive the system to evolve towards the different clusters obtained.

The massive collection of data via emerging technologies like the Internet of Things (IoT) requires finding optimal ways to reduce the observations in the time series analysis domain. The IoT time series require aggregation methods that can preserve and represent the key characteristics of the data. In this paper, we propose a segmentation algorithm that adapts to unannounced mutations of the data (i.e., data drifts). The algorithm splits the data streams into blocks and groups them in square matrices, computes the Discrete Cosine Transform (DCT), and quantizes them. The key information is contained in the upper-left part of the resulting matrix. We extract this sub-matrix, compute the modulus of its eigenvalues, and remove duplicates. The algorithm, called BEATS, is designed to tackle dynamic IoT streams, whose distribution changes over time. We implement experiments with six datasets combining real, synthetic, real-world data, and data with drifts. Compared to other segmentation methods like Symbolic Aggregate approXimation (SAX), BEATS shows significant improvements. Trying it with classification and clustering algorithms it provides efficient results. BEATS is an effective mechanism to work with dynamic and multi-variate data, making it suitable for IoT data sources. The datasets, code of the algorithm and the analysis results can be accessed publicly at: https://github.com/auroragonzalez/BEATS.

The ever-increasing volume and complexity of time series data, emerging in various application domains, necessitate efficient dimensionality reduction for facilitating data mining tasks. Symbolic representations, among them symbolic aggregate approximation (SAX), have proven very effective in compacting the information content of time series while exploiting the wealth of search algorithms used in bioinformatics and text mining communities. However, typical SAX-based techniques rely on a Gaussian assumption for the underlying data statistics, which often deteriorates their performance in practical scenarios. To overcome this limitation, this work introduces a method that negates any assumption on the probability distribution of time series. Specifically, a data-driven kernel density estimator is first applied on the data, followed by Lloyd-Max quantization to determine the optimal horizontal segmentation breakpoints. Experimental evaluation on distinct datasets demonstrates the superiority of our method, in terms of reconstruction accuracy and tightness of lower bound, when compared against the conventional and a modified SAX method.

The detection of outliers in time series data is a core component of many data-mining applications and broadly applied in industrial applications. In large data sets algorithms that are efficient in both time and space are required. One area where speed and storage costs can be reduced is via symbolization as a pre-processing step, additionally opening up the use of an array of discrete algorithms. With this common pre-processing step in mind, this work highlights that (1) existing symbolization approaches are designed to address problems other than outlier detection and are hence sub-optimal and (2) use of off-the-shelf symbolization techniques can therefore lead to significant unnecessary data corruption and potential performance loss when outlier detection is a key aspect of the data mining task at hand. Addressing this a novel symbolization method is motivated specifically targeting the end use application of outlier detection. The method is empirically shown to outperform existing approaches.

The abundance and value of mining large time series data sets has long been acknowledged. Ubiquitous in fields ranging from astronomy, biology and web science the size and number of these datasets continues to increase, a situation exacerbated by the exponential growth of our digital footprints. The prevalence and potential utility of this data has led to a vast number of time-series data mining techniques, many of which require symbolization of the raw time series as a pre-processing step for which a number of well used, pre-existing approaches from the literature are typically employed. In this work we note that these standard approaches are sub-optimal in (at least) the broad application area of time series comparison leading to unnecessary data corruption and potential performance loss before any real data mining takes place. Addressing this we present a novel quantizer based upon optimization of comparison fidelity and a computationally tractable algorithm for its implementation on big datasets. We demonstrate empirically that our new approach provides a statistically significant reduction in the amount of error introduced by the symbolization process compared to current state-of-the-art. The approach therefore provides a more accurate input for the vast number of data mining techniques in the literature, providing the potential of increased real world performance across a wide range of existing data mining algorithms and applications.

The Piecewise Aggregate Approximation (PAA) is widely used in time series data mining because it allows to discretize, to reduce the length of time series and it is used as a subroutine by algorithms for patterns discovery, indexing, and classification of time series. However, it requires setting one parameter: the number of segments to consider during the discretization. The optimal parameter value is highly data dependent in particular on large time series. This paper presents a heuristic for time series compression with PAA which minimizes the loss of information. The heuristic is built upon the well known metaheuristic GRASP and strengthened with an inclusion of specific global search component. An extensive experimental evaluation on several time series datasets demonstrated its efficiency and effectiveness in terms of compression ratio, compression interpretability and classification.