Mikhail Zymbler

Currently, the mining of streaming time series has become increasingly important in a wide range of applications. The matrix profile (MP) is a simple yet powerful exploratory time series mining tool, which provides the discovery of numerous time series mining primitives without requiring prior domain knowledge. The research community has introduced a substantial number of MP computation algorithms. However, to the best of our knowledge, no existing method is both GPU-oriented and stream-oriented. In this article, we introduce the StreaMP (Streaming Matrix Profile) algorithm for GPUs, which overcomes the above limitation. Our algorithm processes incoming data in a segment-wise manner, accumulating the time series arriving from a sensor in RAM. StreaMP merges the MPs of the segment and the time series accumulated so far using a proposed multistep scheme. To compute MPs, StreaMP is able to apply any GPU-based algorithm, which encapsulates all details of parallel processing. The MP built by StreaMP makes it possible to discover repeated and anomalous patterns in the streaming time series in linear time. Experimental evaluation shows that StreaMP outperforms SCAMP, which is currently the fastest GPU-based MP computation algorithm for batch processing.

Currently, in a wide spectrum of subject domains, time series data mining requires the efficient subsequence anomaly discovery in a very long time series, which cannot be entirely placed in RAM. At present, one of the best approaches to solving such a problem is to formalize the anomaly as a discord, a given-length subsequence that is maximally far away from its non-overlapping nearest neighbor. In the article, we introduce a novel parallel algorithm called PADDi (PALMAD-based anomaly discovery on distributed GPUs), which discovers arbitrary-length discords in a very long time series on a high-performance cluster with nodes, each of which is equipped with multiple GPUs. The algorithm exploits two-level parallelism: first, when the time series is divided into equal-length fragments stored on disks associated with the cluster nodes, and second, when a fragment is split into equal-length segments to be processed by GPUs of the respective node. To implement data exchanges between nodes and calculations on GPUs within a node, we employ MPI and CUDA technologies, respectively. The algorithm performs as follows. Firstly, in each segment processed by one GPU, the algorithm selects potential discords and then discards false positives, resulting in the local candidate set. Next, local candidate sets are sent among cluster nodes in an ‘‘all-to-all’’ manner, resulting in a global candidate set. Then, each cluster node refines the global candidates within its fragment, obtaining the local resulting set of true positive discords. Finally, each cluster node sends the local resulting sets to a master node, which outputs the end result as the intersection of the received local resulting sets. Extensive experiments over real-world and synthetic million-length time series on various configurations of two high-performance clusters with different models of GPU onboard (from 48 to 64 GPUs in total) showed that our algorithm’s scalability remains linear without stagnation or degradation.

Currently, in a wide spectrum of applications, to avoid the processing and analysis of incomplete time series, end-users need efficient and accurate approaches to online imputation of missing values. In the article, we introduce a novel method called SANNI (snippet and artificial neural network-based imputation) for the recovery of missing values in multivariate time series coming online. SANNI leverages behavioral patterns (called snippets) that are subsequences representing an actor’s typical activities, which are reflected by the time series. Preprocessing is performed for each series of a representative fragment of the input data, where we normalize all the subsequences with non-NaN values and discover snippets. To impute, our method employs two deep learning models: Recognizer and Reconstructor. Given a multivariate subsequence ended by a missing value, Recognizer outputs a snippet to which the subsequence of the series is the most similar. Reconstructor, for each series, imputes missing values using the snippet discovered previously and results taken from Recognizer. In the extensive experiments, SANNI on average outperforms state-of-the-art competitors over time series from diverse subject domains related to an actor with predefined activities as well as under the blackout scenario.

In medical practice, the primary diagnosis of diseases should be carried out quickly and, if possible, automatically. The processing of multimodal data in medicine has become a ubiquitous technique in the classification, prediction and detection of diseases. Pneumonia is one of the most common lung diseases. In our study, we used chest X-ray images as the first modality and the results of laboratory studies on a patient as the second modality to detect pneumonia. The architecture of the multimodal deep learning model was based on intermediate fusion. The model was trained on balanced and imbalanced data when the presence of pneumonia was determined in 50% and 9% of the total number of cases, respectively. For a more objective evaluation of the results, we compared our model performance with several other open-source models on our data. The experiments demonstrate the high performance of the proposed model for pneumonia detection based on two modalities even in cases of imbalanced classes (up to 96.6%) compared to single-modality models’ results (up to 93.5%). We made several integral estimates of the performance of the proposed model to cover and investigate all aspects of multimodal data and architecture features. There were accuracy, ROC AUC, PR AUC, F1 score, and the Matthews correlation coefficient metrics. Using various metrics, we proved the possibility and meaningfulness of the usage of the proposed model, aiming to properly classify the disease. Experiments showed that the performance of the model trained on imbalanced data was even slightly higher than other models considered.

Currently, discovering subsequence anomalies in time series remains one of the most topical research problems. A subsequence anomaly refers to successive points in time that are collectively abnormal, although each point is not necessarily an outlier. Among numerous approaches to discovering subsequence anomalies, the discord concept is considered one of the best. A time series discord is intuitively defined as a subsequence of a given length that is maximally far away from its non-overlapping nearest neighbor. Recently introduced, the MERLIN algorithm discovers time series discords of every possible length in a specified range, thereby eliminating the need to set even that sole parameter to discover discords in a time series. However, MERLIN is serial, and its parallelization could increase the performance of discord discovery. In this article, we introduce a novel parallelization scheme for GPUs called PALMAD, parallel arbitrary length MERLIN-based anomaly discovery. As opposed to its serial predecessor, PALMAD employs recurrent formulas we have derived to avoid redundant calculations, and advanced data structures for the efficient implementation of parallel processing. Experimental evaluation over real-world and synthetic time series shows that our algorithm outperforms parallel analogs. We also apply PALMAD to discover anomalies in a real-world time series, employing our proposed discord heatmap technique to illustrate the results.

The detection of anomalous subsequences in a time series is required today in a wide range of computationally intensive applications such as digital industry, the Internet of Things, personal medicine, etc. One of the most attractive approaches to solving this problem is the concept of discord—the subsequence of a time series with the greatest distance to its nearest neighbor—because it requires the analyst to set only one intuitive parameter (subsequence length). The previously proposed DRAG (discord range aware gathering) algorithm for finding discords is exact, but its parallel versions are absent for any hardware architecture. The paper proposes a new approach to parallelizing this algorithm for a graphics processing unit, which is called PD3 (parallel DRAG-based discord discovery). A data preprocessing phase was added to the PD3 to compute mean values and standard deviations of all subsequences of the time series. The results are used further at the candidate selection and refinement phases to calculate the distances between the subsequences of the time series. Each algorithm phase is parallelized separately on the basis of the data parallelism concept and using vector data structures. Experiments show that the PD3 is much ahead of competing parallel algorithms in terms of the average time taken to find one discord.

This paper proposed a short-term two-stage hybrid algorithmic framework for trade and trend analysis of the Forex market by augmenting the currency pair datasets with transformed attributes using a few technical indicators and statistical measures. In the first phase, an optimized deep predictive coding network (DPCN) based on a meta-heuristic reptile search algorithm (RSA) inspired by the intelligent hunting activities of the crocodiles is exploited to develop this RSA-DPCN predictive model. The proposed model has been compared with optimized versions of extreme learning machine (ELM) and functional link artificial neural network (FLANN) with genetic algorithm (GA), particle swarm optimization (PSO), and differential evolution (DE) along with the RSA optimizers. The performance of this model has been evaluated and validated through several statistical tests. In the second phase, the up and down trends are analyzed using the Higher Highs Higher Lows, and Lower Highs Lower Lows (HHs/HLs and LHs/LLs) trend analysis tool. Further, the observed trends are compared with the actual trends observed on the exchange price of real datasets. This study shows that the proposed RSA-DPCN model accurately predicts the exchange price. At the same time, it provides a well-structured platform to discern the directions of the market trends and thereby guides in finding the entry and exit points of the Forex market.

Summarization of a long time series often occurs in analytical applications related to decision-making, modeling, planning, and so on. Informally, summarization aims at discovering a small-sized set of typical patterns (subsequences) to briefly represent the long time series. Apparent approaches to summarization like motifs, shapelets, cluster centroids, and so on, either require training data or do not provide an analyst with information regarding the fraction of the time series that a typical subsequence found corresponds to. Recently introduced, the time series snippet concept overcomes the above-mentioned limitations. A snippet is a subsequence that is similar to many other subsequences of the time series with respect to a specially defined similarity measure based on the Euclidean distance. However, the original Snippet-Finder algorithm has cubic time complexity concerning the lengths of the time series and the snippet. In this article, we propose the PSF (Parallel Snippet-Finder) algorithm that accelerates the original snippet discovery schema with GPU and ensures acceptable performance over very long time series. As opposed to the original algorithm, PSF splits the calculation of the similarity of all the time series subsequences to a snippet into several steps, each of which is performed in parallel. Experimental evaluation over real-world time series shows that PSF outruns both the original algorithm and a straightforward parallelization.

Botanical plants suffer from several types of diseases that must be identified early to improve the production of fruits and vegetables. Mango fruit is one of the most popular and desirable fruits worldwide due to its taste and richness in vitamins. However, plant diseases also affect these plants’ production and quality. This study proposes a convolutional neural network (CNN)-based metaheuristic approach for disease diagnosis and detection. The proposed approach involves preprocessing, image segmentation, feature extraction, and disease classification. First, the image of mango leaves is enhanced using histogram equalization and contrast enhancement. Then, a geometric mean-based neutrosophic with a fuzzy c-means method is used for segmentation. Next, the essential features are retrieved from the segmented images, including the Upgraded Local Binary Pattern (ULBP), color, and pixel features. Finally, these features are given into the disease detection phase, which is modeled using a Convolutional Neural Network (CNN) (deep learning model). Furthermore, to enhance the classification accuracy of CNN, the weights are fine-tuned using a new hybrid optimization model referred to as Cat Swarm Updated Black Widow Model (CSUBW). The new hybrid optimization model is developed by hybridizing the standard Cat Swarm Optimization Algorithm (CSO) and Black Widow Optimization Algorithm (BWO). Finally, a performance evaluation is undergone to validate the efficiency of the projected model.