Symbolic Time Intervals (STIs) represent events having a non-zero time duration, which are common in various application domains. In this article, we focus on the challenge of STIs series classification (STIC). While in the related problem of time series classification (TSC) Rocket is well-known for its exceptionally fast runtime while achieving accuracy comparable to state-of-the-art, it has only recently been studied in the field of STIC. However, since Rocket as well as its enhanced variants for TSC (e.g., MiniRocket and MultiRocket) solely rely on global features, they might not always fit best for the classification of thousands of time-units long STI series out-of-the-box, which are rather common in STIC. We introduce STORM—a novel, generic MapReduce framework for STIC, which (1) converts raw input STIs series into multivariate time series (MTS) representation; (2) partitions the converted MTS into fixed-sized blocks, each transformed independently into a uniform latent space via a common, desired Rocket variant used as a base transformation in STORM; and (3) performs sequence classification of the blocks’ transformed feature vectors via a deep, lightweight, bidirectional LSTM network. The evaluation demonstrates that STORM significantly improves accuracy over eight state-of-the-art methods for STIC either when applied with MiniRocket and MultiRocket as base transformations, as well as over the baselines of applying the respective Rocket variants directly to the converted MTS representation, that is, while also reporting overall comparable training times, on a benchmark of eight real-world STIC datasets including both extremely long and short STIs series.

Predicting an event of interest based on heterogeneous multivariate temporal data is challenging but desirable as it allows the utilization of all types of temporal variables. In various domains, symbolic time intervals (STIs) can be used to represent real-life events that vary in duration, such as the period a traffic light remains green, or the time a patient undergoes treatment or is on medication. Further, heterogeneous multivariate temporal data may be composed of STIs along with event-driven or continuous temporal variables, such as traffic collisions or blood test values. Temporal abstraction can be used to uniformly represent heterogeneous multivariate temporal variables with STIs, from which frequent time intervals related patterns (TIRPs) can be discovered. We extend earlier work on continuous completion prediction of a single TIRP that ends with an event of interest, introducing a continuous prediction method based on multiple different instances of multiple TIRPs that end with the event of interest, for which we propose and evaluate several weighted aggregation functions. The proposed method overall performed better on real-life, medical, and non-medical datasets, than the use of a single TIRP, and in comparison to the baseline models (XGBoost, ResNet, LSTM-FCN, and ROCKET).

Many time series classification algorithms have been proposed, including deep neural networks based, which so far focused mainly on improving model architectures rather than on data pre-processing. Generalization is crucial in time series classification and it can be achieved by abstracting the data. Data abstraction may also be useful to avoid handling challenges with error measurements, missing values, and irregular sampling. We propose transforming the raw time series into a symbolic time series representation, using a method known as temporal abstraction, before feeding it to the deep neural networks. This transformation can greatly enhance generalization and may potentially improve classification performance. In particular, we investigate the effectiveness of temporal abstraction when combined with convolution-based sequence models or recurrent neural networks. The methods were evaluated on 128 univariate datasets. Our evaluation shows that even when using equal frequency discretization, a relatively simple method, outperforms most state-of-the-art deep neural networks’ performance for univariate time series classification when fed by raw time series.

The increasing availability of multivariate temporal data in many domains, such as biomedical, security and more, provides exceptional opportunities for temporal knowledge discovery, classification and prediction, but also challenges. Temporal variables are often sparse and in many domains, such as in biomedical data, they have huge number of variables. In recent decades in the biomedical domain events, such as conditions, drugs and procedures, are stored as time intervals, which enables to discover Time Intervals Related Patterns (TIRPs) and use for classification or prediction. In this study we present a framework for outcome events prediction, called Maitreya, which includes an algorithm for TIRPs discovery called KarmaLegoD, designed to handle huge number of symbols. Three indexing strategies for pairs of symbolic time intervals are proposed and compared, showing that the use of FullyHashed indexing is only slightly slower but consumes minimal memory. We evaluated Maitreya on eight real datasets for the prediction of clinical procedures as outcome events. The use of TIRPs outperform the use of symbols, especially with horizontal support (number of instances) as TIRPs feature representation.

Detecting unknown worms is a challenging task. Extant solutions, such as anti-virus tools, rely mainly on prior explicit knowledge of specific worm signatures. As a result, after the appearance of a new worm on the Web there is a significant delay until an update carrying the worm's signature is distributed to anti-virus tools. During this time interval a new worm can infect many computers and cause significant damage. We propose an innovative technique for detecting the presence of an unknown worm, not necessarily by recognizing specific instances of the worm, but rather based on the computer measurements. We designed an experiment to test the new technique employing several computer configurations and background applications activity. During the experiments 323 computer features were monitored. Four feature selection techniques were used to reduce the amount of features and four classification algorithms were applied on the resulting feature subsets. Our results indicate that using this approach resulted in exceeding 90% mean accuracy, and for specific unknown worms accuracy reached above 99%, using just 20 features while maintaining a low level of false positive rate.