status: Published

In many time series classification tasks, each instance is described by multiple channels (i.e., signals). This introduces an additional computational burden as the time and resources to train a classifier increase as more channels become available. This is problematic as some tasks can be described by a huge number of channels. However, not all channels may be necessary as some could be irrelevant or redundant, and including these can (dramatically) increase run time while yielding no benefit in terms of predictive performance. Therefore, it can be useful to automatically select a subset of the channels to include in the analysis. We propose TSelect, a novel scalable and classifier-agnostic approach that automatically selects a relevant and non-redundant subset of the channels for multivariate time series classification (MTSC). Experimentally, we show on a large benchmark suite that TSelect (1) eliminates on average 62% of the channels, (2) significantly improves a classifier’s run time without sacrificing predictive performance, and (3) outperforms the two state-of-the-art channel selectors (ECS and ECP) on the majority of the experiments.

Anomaly detection algorithms attempt to find instances that deviate from the expected behavior. Because this is often tackled as an unsupervised task, anomaly detection models rely on exploiting intuitions about what constitutes anomalous behavior. These typically take the form of data-driven heuristics that measure the anomalousness of each instance. However, the effectiveness of unsupervised detectors are limited by the validity of their intuition. Because these are not universally true, one can improve the detectors' performance by using a semi-supervised approach that exploits a few labeled instances. This paper proposes a novel semi-supervised tree ensemble based anomaly detection framework. We compare our proposed approach to several baselines and show that it achieves comparable performance to state-of-the-art neural networks on six real-world and 14 benchmark datasets.

In the multi-instance learning (MIL) setting instances are grouped together into bags. Labels are provided only for the bags and not on the level of individual instances. A positive bag label means that at least one instance inside the bag is positive, while a negative bag label restricts all the instances in the bag to be negative. MIL data naturally arises in many contexts, such as anomaly detection, where labels are rare and costly, and one often ends up annotating the label for sets of instances. Moreover, in many real-world anomaly detection problems, only positive labels are collected because they usually represent critical events. Such a setting, where only positive labels are provided along with unlabeled data, is called Positive and Unlabeled (PU) learning. Despite being useful for several use cases, there is no work dedicated to learning from positive and unlabeled data in a multi-instance setting for anomaly detection. Therefore, we propose the first method that learns from PU bags in anomaly detection. Our method uses an autoencoder as an underlying anomaly detector. We alter the autoencoder's objective function and propose a new loss that allows it to learn from positive and unlabeled bags of instances. We theoretically analyze this method. Experimentally, we evaluate our method on 30 datasets and show that it performs better than multiple baselines adapted to work in our setting.