Shapelets are interclass discriminative subsequences that can be used to characterize target classes. Learning shapelets by continuous optimization has recently been studied to improve classification accuracy. However, there are two issues in previous studies. First, since the locations where shapelets appear in the time series are determined by only their shapes, shapelets may appear at incorrect and non-discriminative locations in the time series, degrading the accuracy and interpretability. Second, the theoretical interpretation of learned shapelets has been limited to binary classification. To tackle the first issue, we propose a continuous optimization that learns not only shapelets but also their probable locations in a time series, and we show theoretically that this enhances feature discriminability. To tackle the second issue, we provide a theoretical interpretation of shapelet closeness to the time series for target / off-target classes when learning with softmax loss, which allows for multi-class classification. We demonstrate the effectiveness of the proposed method in terms of accuracy, runtime, and interpretability on the UCR archive.

The need for explainability in time-series classification models has been increasing. Counterfactual explanations recommend how to modify the features of an original instance so that the prediction by a given classifier flips to the desired class. Since features in the time series are temporally dependent, interpretability is improved by considering intervals where the counterfactual can deviate from the original instance. In this study, we propose a model-agnostic counterfactual generation method (CEI) that jointly learns these intervals and the counterfactual. Furthermore, CEI can generate a counterfactual tailored to the directly specified limited number of intervals. We mathematically formulate CEI as a continuous optimization and demonstrate its effectiveness on the UCR datasets.

Shapelets are subsequences that are effective for classifying time-series instances. Learning shapelets by a continuous optimization has recently been studied to improve computational efficiency and classification performance. However, existing methods have employed predefined and fixed shapelet lengths during the continuous optimization, despite the fact that shapelets and their lengths are inherently interdependent and thus should be jointly optimized. To efficiently explore shapelets of high quality in terms of interpretability and inter-class separability, this study makes the shapelet lengths continuous and learnable. The proposed formulation jointly optimizes not only a binary classifier and shapelets but also shapelet lengths. The derived SGD optimization can be theoretically interpreted as improving the quality of shapelets in terms of shapelet closeness to the time series for target / off-target classes. We demonstrate improvements in area under the curve, total training time, and shapelet interpretability on UCR binary datasets.

Shapelets are subsequences that are effective for classifying time-series instances. In this study, we consider when each time-series instance is obtained as progress, and formulate the problem of learning shapelet evolution over progress. For example, shapelets can change their shapes according to progress with human habituation, seasonal effects, and system degradation. When given time-series instances, progress values, and binary class labels, the proposed optimization formulation can jointly learn not only the shapelets and a classifier but also regression models for predicting shapelet evolution. The derived optimization solution method allows regression models to be learned by using off-the-shelf regression solvers, and scales linearly with time-series length. We demonstrate its effectiveness in industrial case studies.

Shapelets are subsequences that are effective for classifying time-series instances. Joint learning of both classifiers and shapelets has recently been studied because this approach improves algorithmic complexity and classification performance. However, the existing methods lack the power of feature discrimination due to using traditional sigmoid cross-entropy loss functions. To enhance feature discriminability, we propose self-adaptive scaling of the loss functions, inspired by the recent discriminative loss in computer vision. In addition, we propose a theoretically sound regularization that enhances feature discriminability and maintains shapelet interpretability by shrinking appropriate features. Using UCR datasets, we demonstrate improved area under the curve and interpretability of shapelets with a small number of shapelets.

It is common to forecast time-series data in a cloud server environment by building a forecasting model after collecting all the time-series data at the server-side. However, this may not be efficient in time-critical forecasting, control, and decision-making due to high latency, bandwidth, and network connectivity issues. Hence, edge devices can be employed to make quick forecasting on a real-time basis. However, due to limited computing resources and processing power, edge devices cannot handle a huge volume of multivariate time-series data. Therefore, it is desirable to develop an algorithm that trains and updates a forecasting model incrementally. This can be done by using a small chunk of multivariate time-series data without sacrificing the forecasting accuracy, while training and inference can be executed in the edge device itself. In this context, we propose a new forecasting method called Incremental Learning Vector Auto Regression (ILVAR). It works by minimizing the variance difference between actual and forecasted values as a new chunk of time-series data arrives sequentially and thereby it updates the forecasting model incrementally. To show the effectiveness of the proposed method, experiments were performed on 11 publicly available datasets from diverse domains using Raspberry Pi-2 as an edge device and evaluated using five metrics such as MAPE, RMSE, $\mathrm{R}^{2}$ score, Computational time, and Memory consumption for 1-step and 24-step ahead forecasting tasks. The performance was compared with the state-of-the-art methods such as Vector Auto Regression (VAR), Incremental Learning Extreme Learning Machine (ILELM), and Incremental Learning Long Short-Term Memory (ILLSTM). These experimental results suggest that our proposed method performs better than existing methods and is able to achieve the desired performance for forecasting with edge devices.

Shapelets are time-series segments effective for classifying time-series instances.Joint learning of both classifiers and shapelets has been studied in recent years because this approach provides both superior classification performance and interpretable results.However, the optimization formulation is nonconvex, so bad local minima must be avoided.Very recently, this issue has been tackled by introducing Self-Paced Learning (SPL) into these methods.With the aim of intelligently discovering initial shapelets, we introduce two steps into this binary classification method so as to consistently optimize the same loss function of interest: optimizing discovery of discriminative initial shapelets from many time-series segments by using supervised feature selection, and jointly optimizing shapelets, model parameters, and latent instance weights in SPL.Using UCR datasets, we demonstrate improved area under the curve, and interpretability of shapelets where the number of shapelets is small.

Nowadays a lot of data are being continuously or incrementally collected at various fields through IoT (Internet of Things) devices and sensors. To extract useful patterns from these huge volumes of data, pattern recognition and machine learning techniques are applied, which build models by extracting patterns from data at once, and the models are not updated until the performances of the models deteriorate significantly. This traditional approach of learning a model using all the data at once may not be feasible in many applications because it may require a huge communication cost and storage to collect the data and take a very long time to build a model, and the onetime built model may not be able to learn the changed patterns automatically over time. To overcome some of the limitations of the traditional approach, we propose a new method called Robust Incremental Logistic Regression (RILR), which learns and updates model parameters as new batches of training data arrive. We show the effectiveness of the proposed method by performing experiments with 10 publicly available data sets and evaluating it in terms of AUC (Area Under the receiver operating characteristic Curve) on test data, robustness, execution time, and storage requirement. Experimental results suggest that our proposed method is able to achieve the desired performance on most of the data sets.

Large amounts of time-series data have become accessible due to the rapid development of Internet-of-Things technologies and the demand for extracting useful knowledge from these data is increasing. Toward this goal, time-series segmentation — dividing data into similar segments — is a promising method for understanding the mechanisms in time-series data. In this paper, we focus on time-lag that appears in real datasets. Time lag — a typical phenomenon in time-series data — occurs when the speed of information diffusion differs between variables. However, conventional methods cannot distinguish differences in segmentation positions. In response, we propose Lag-Aware Multivariate Time-Series Segmentation (LAMTSS), an algorithm capturing time-lag across variables to determine segmentation positions for each variable. LAMTSS utilizes dynamic time warping without hyperparameter tuning. We confirm the accuracy of LAMTSS using artificial datasets and demonstrate the discovery of useful knowledge in real datasets.

Shapelets are time-series segments effective for classifying time-series instances. Joint learning of both classifiers and shapelets has been studied in recent years, because such methods provide both interpretable results and superior accuracy. Partial Area Under the ROC curve (pAUC) for a low range of False Positive Rates (FPR) is an important performance measure for practical cases in industries such as medicine, manufacturing, and maintenance. In this study, we propose a method that jointly learns both shapelets and a classifier for pAUC optimization in any FPR range, including the full AUC. We demonstrate superiority of pAUC on UCR time-series datasets and its effectiveness in industrial case studies.

There is a need for forecasting of short-range future values in data streams such as traffic flows, stock prices, and electricity consumption. However, concept drift in non-stationary data streams is an important problem. We propose an online prediction method called OPOSSAM for such data streams. OPOSSAM manages time-series segments in short-term memory and long-term memory, and forecasts future values by local regression based on the similarity of time-series segments. In particular, OPOSSAM keeps long-term memory consistent by reducing redundant samples with large prediction errors, and automatically adjusts the prediction model based on short-term memory from the prior model learned from the entire memory in order to deal with concept drift. Experimental results show accuracy superior to that of baseline methods on real-world datasets of traffic flow, stock prices, and electricity consumption.

An ultra-low power CMOS voltage reference circuit has been fabricated in a 0.35-μm standard CMOS process. The circuit generates a reference voltage based on threshold voltage of a MOSFET at absolute zero temperature. Theoretical analyses and experimental results showed that the circuit generates a quite stable reference voltage of 745 mV on average. The temperature coefficient and line sensitivity of the circuit were 7 ppm/°C and 20 ppm/V, respectively. The power supply rejection ratio (PSRR) was −45 dB at 100 Hz. The circuit consists of subthreshold MOSFETs with a low-power dissipation of 0.3 μW or less and a 1.5-V power supply. Because the circuit generates a reference voltage based on threshold voltage of a MOSFET in an LSI chip, it can be used as an on-chip process monitoring circuit and as a part of the on-chip process compensation circuit systems.

An on-chip process, supply voltage, and temperature (PVT) compensation technique for a low-voltage CMOS digital circuit is proposed. Because the degradation of circuit performance originates from the variation of the saturation current, a compensation technique that uses a reference current that is independent of PVT variations was developed. The operations of the circuit were confirmed by SPICE simulation with a set of 0.35-mum standard CMOS parameters. Moreover, Monte Carlo simulations assuming process spread and device mismatch in all MOSFETs showed the effectiveness of the proposed technique and achieved performance improvement of 74%. The circuit is useful for on-chip compensation to mitigate the degradation of circuit performance with PVT variation in low-voltage digital circuits.

A low power on-chip reference clock generator consisting of subthreshold MOSFET circuits is proposed. It uses a simple frequency-locked loop technique with no inductor, quartz resonator, or MEMS oscillator. Theoretical analyses and a SPICE simulation with 0.35-mum CMOS parameters showed that the clock frequency could be controlled in the frequency range of 10-1000 kHz. When operated at 170 kHz, the generator showed a temperature coefficient of 100 ppm/degC, a line sensitivity of 3%/V, and a power consumption of 20 muW. Our clock generator can be used as a reference clock for intermittent operation in power aware LSIs.

Recently, Bayesian network classifiers (BNCs) have attracted many researchers because they can produce classification models with dependencies among attributes. From the application viewpoint, however, BNCs sometimes produce models too complicated to interpret easily. In this paper, we propose k-Bayesian network classifier (k-BNC), which is a new method to reconstruct the attribute-dependency relationship from data for health promotion planning. From the health promotion viewpoint, it would be highly advantageous if occupational physicians could make effective plans for employees, and if employees could carry out the plans easily. Therefore, we focus on the attribute dependencies in classification models represented as a directed acyclic graph (DAG), and find the effective attributes by measuring the standardized Kullback-Leibler divergence from parent attributes to their children. In experimental evaluation, we firstly compare the accuracy of k-BNC with that of Naive Bayes Classifiers, and other wellknown Bayesian Networks and structure learning methods (k2 algorithm etc.) on some public datasets. We show that our proposed k-BNC method successfully produces classification models for the prioritization of health promotion plans on our health checkup data.

A floating millivolt reference circuit to generate a PTAT current was developed by using MOSFETs operated in the subthreshold region. The circuit generates a floating voltage of about 10 mV. The variations in the reference are ±2.7 % in a temperature range from -20 to 100 °C. The accuracy of the reference circuit can be improved to ±0.3 % with a correction technique using a curvature-correction circuit. The total power consumption of the circuit was 4.6μW at 100 °C.

For many real world problems we must perform classification under widely varying amounts of computational resources. For example, if asked to classify an instance taken from a bursty stream, we may have from milliseconds to minutes to return a class prediction. For such problems an anytime algorithm may be especially useful. In this work we show how we can convert the ubiquitous nearest neighbor classifier into an anytime algorithm that can produce an instant classification, or if given the luxury of additional time, can utilize the extra time to increase classification accuracy. We demonstrate the utility of our approach with a comprehensive set of experiments on data from diverse domains.

An enterprise service-oriented architecture is typically realized on a messaging infrastructure called an enterprise service bus (ESB). An ESB is a bus which delivers messages from service requesters to service providers. Since it sits between the service requesters and providers, it is not appropriate to use any existing capacity planning methodology for servers, such as modeling to estimate an ESB's capacity. There are programs which run on an ESB called mediation modules. Their functionalities vary and depend on how people use the ESB. This creates difficulties for capacity planning and performance evaluation. This paper proposes a performance evaluation methodology and techniques for ESBs. We actually run the ESB on a real machine while providing a pseudo-environment around it. In order to ease setting up the environment we provide ultra-light service requestors and service providers for the ESB under test. We show that the proposed mock environment can be set up with practical hardware resources available at the time of hardware resource assessment. Our experimental results showed that the testing results with our mock environment are equivalent to the results in the real environment