Foundation models have recently emerged as a promising approach for time series analysis, adapting transformer architectures originally designed for natural language processing to handle continuous temporal data. While these models demonstrate strong performance across various time series tasks, their handling of multivariate time series, particularly inter-channel dependencies, remains underexplored. In this paper, we present a comprehensive analysis of current foundation models for time series, including tokenization-based, patch-based, and shape-based approaches, focusing on their mechanisms and data representations for capturing relationships between channels. Our analysis shows that even though these models have advanced architectures, they mostly process channels independently, which may prevent them from fully capturing cross-channel patterns. We examine this limitation across different model families and discuss its implications for multivariate time series analysis. Our empirical evaluation shows that foundation models perform well on simpler tasks but exhibit diminished effectiveness as channel dependencies increase, with specialized time series methods consistently outperforming them on complex datasets. These findings highlight the critical need for channel-aware architectures and more effective strategies for modeling inter-channel relationships in foundation models.

Attribution methods are essential for interpreting time series predictive models by quantifying the relevance of each time step for the prediction. State-of-the-art methods are often based on SHAP, an attribution method developed for tabular data. However, this has several challenges. First, SHAP is expensive to compute, especially for long time series, hence to speed it up it is usually approximated. Second, the impact of the background selection for emulating data ‘missingness’, essential to compute SHAP, remains understudied. Third, SHAP and more generally attribution methods for time series regression are notably lacking. In this paper, we address these limitations and propose TSHAP, a novel SHAP-based attribution method for time series classification and regression. TSHAP leverages a sliding window to group temporal data, enabling the efficient computation of exact SHAP values for each group. We further develop a methodology for the principled selection of background data. We evaluate TSHAP’s performance and robustness using comprehensive experiments on synthetic and real-world time series datasets.

We report on a trans-Atlantic PhD student mobility program that connects two graduate research training initiatives in the US and Ireland, centered on developing future researchers in artificial intelligence (AI) and machine learning (ML). We discuss both the structure of the student exchange experiences and share key learnings from this international collaboration. The most important lesson learned is that providing a structured mobility program and matched visiting pairs is a highly effective way to improve learning outcomes compared to more typical ad-hoc individual visits.

Explanation for Multivariate Time Series Classification (MTSC) is an important topic that is under explored. There are very few quantitative evaluation methodologies and even fewer examples of actionable explanation, where the explanation methods are shown to objectively improve specific computational tasks on time series data. In this paper we focus on analyzing InterpretTime, a recent evaluation methodology for attribution methods applied to MTSC. We showcase some significant weaknesses of the original methodology and propose ideas to improve both its accuracy and efficiency. Unlike related work, we go beyond evaluation and also showcase the actionability of the produced explainer ranking, by using the best attribution methods for the task of channel selection in MTSC. We find that perturbation-based methods such as SHAP and Feature Ablation work well across a set of datasets, classifiers and tasks and outperform gradient-based methods. We apply the best ranked explainers to channel selection for MTSC and show significant data size reduction and improved classifier accuracy.

Abstract Time series classification is a task which deals with temporal sequences, a prevalent data type common in domains such as human activity recognition, sports analytics and general sensing. In this area, interest in explanability has been growing as explanation is key to understand the data and the model better. Recently, a great variety of techniques (e.g., LIME, SHAP, CAM) have been proposed and adapted for time series to provide explanation in the form of saliency maps, where the importance of each data point in the time series is quantified with a numerical value. However, the saliency maps can and often disagree, so it is unclear which one to use. This paper provides a novel framework to quantitatively evaluate and rank explanation methods for time series classification. We show how to robustly evaluate the informativeness of a given explanation method (i.e., relevance for the classification task), and how to compare explanations side-by-side. The goal is to recommend the best explainer for a given time series classification dataset. We propose AMEE, a Model-Agnostic Explanation Evaluation framework, for recommending saliency-based explanations for time series classification. In this approach, data perturbation is added to the input time series guided by each explanation. Our results show that perturbing discriminative parts of the time series leads to significant changes in classification accuracy, which can be used to evaluate each explanation. To be robust to different types of perturbations and different types of classifiers, we aggregate the accuracy loss across perturbations and classifiers. This novel approach allows us to recommend the best explainer among a set of different explainers, including random and oracle explainers. We provide a quantitative and qualitative analysis for synthetic datasets, a variety of time-series datasets, as well as a real-world case study with known expert ground truth.

Accuracy is a key focus of current work in time series classification. However, speed and data reduction are equally important in many applications, especially when the data scale and storage requirements rapidly increase. Current multivariate time series classification (MTSC) algorithms need hundreds of compute hours to complete training and prediction. This is due to the nature of multivariate time series data which grows with the number of time series, their length and the number of channels. In many applications, not all the channels are useful for the classification task, hence we require methods that can efficiently select useful channels and thus save computational resources. We propose and evaluate two methods for channel selection. Our techniques work by representing each class by a prototype time series and performing channel selection based on the prototype distance between classes. The main hypothesis is that useful channels enable better separation between classes; hence, channels with a larger distance between class prototypes are more useful. On the UEA MTSC benchmark, we show that these techniques achieve significant data reduction and classifier speedup for similar levels of classification accuracy. Channel selection is applied as a pre-processing step before training state-of-the-art MTSC algorithms and saves about 70% of computation time and data storage with preserved accuracy. Furthermore, our methods enable efficient classifiers, such as ROCKET, to achieve better accuracy than using no selection or greedy forward channel selection. To further study the impact of our techniques, we present experiments on classifying synthetic multivariate time series datasets with more than 100 channels, as well as a real-world case study on a dataset with 50 channels. In both cases, our channel selection methods result in significant data reduction with preserved or improved accuracy.

Sentence compression reduces the length of text by removing non-essential content while preserving important facts and grammaticality.Unsupervised objective driven methods for sentence compression can be used to create customized models without the need for ground-truth training data, while allowing flexibility in the objective function(s) that are used for learning and inference.Recent unsupervised sentence compression approaches use custom objectives to guide discrete search; however, guided search is expensive at inference time.In this work, we explore the use of reinforcement learning to train effective sentence compression models that are also fast when generating predictions.In particular, we cast the task as binary sequence labelling and fine-tune a pre-trained transformer using a simple policy gradient approach.Our approach outperforms other unsupervised models while also being more efficient at inference time.

Multi-document summarization (MDS) aims to compress the content in large document collections into short summaries and has important applications in story clustering for newsfeeds, presentation of search results, and timeline generation.However, there is a lack of datasets that realistically address such use cases at a scale large enough for training supervised models for this task.This work presents a new dataset for MDS that is large both in the total number of document clusters and in the size of individual clusters.We build this dataset by leveraging the Wikipedia Current Events Portal (WCEP), which provides concise and neutral human-written summaries of news events, with links to external source articles.We also automatically extend these source articles by looking for related articles in the Common Crawl archive.We provide a quantitative analysis of the dataset and empirical results for several state-of-the-art MDS techniques.The dataset is available at

Previous work on automatic news timeline summarization (TLS) leaves an unclear picture about how this task can generally be approached and how well it is currently solved.This is mostly due to the focus on individual subtasks, such as date selection and date summarization, and to the previous lack of appropriate evaluation metrics for the full TLS task.In this paper, we compare different TLS strategies using appropriate evaluation frameworks, and propose a simple and effective combination of methods that improves over the stateof-the-art on all tested benchmarks.For a more robust evaluation, we also present a new TLS dataset, which is larger and spans longer time periods than previous datasets.

News and social media now play a synergistic role and neither domain can be grasped in isolation. On one hand, platforms such as Twitter have taken a central role in the dissemination and consumption of news. On the other hand, news editors rely on social media for following their audience's attention and for crowd-sourcing news stories. Twitter hashtags function as a key connection between Twitter crowds and the news media, by naturally naming and contextualizing stories, grouping the discussion of news and marking topic trends. In this work, we propose Hashtagger+, an efficient learning-to-rank framework for merging news and social streams in real-time, by recommending Twitter hashtags to news articles. We provide an extensive study of different approaches for streaming hashtag recommendation, and show that pointwise learning-to-rank is more effective than multi-class classification as well as more complex learning-to-rank approaches. We improve the efficiency and coverage of a state-of-the-art hashtag recommendation model by proposing new techniques for data collection and feature computation. In our comprehensive evaluation on real-data, we show that we drastically outperform the accuracy and efficiency of prior methods. Our prototype system delivers recommendations in under 1 minute, with a Precision@1 of 94 percent and article coverage of 80 percent. This is an order of magnitude faster than prior approaches, and brings improvements of 5 percent in precision and 20 percent in coverage. By effectively linking the news stream to the social stream via the recommended hashtags, we open the door to solving many challenging problems related to story detection and tracking. To showcase this potential, we present an application of our recommendations to automated news story tracking via social tags. Our recommendation framework is implemented in a real-time Web system available from insight4news.ucd.ie.

Existing approaches to time series classification can be grouped into shape-based (numeric) and structure-based (symbolic). Shape-based techniques use the raw numeric time series with Euclidean or Dynamic Time Warping distance and a 1-Nearest Neighbor classifier. They are accurate, but computationally intensive. Structure-based methods discretize the raw data into symbolic representations, then extract features for classifiers. Recent symbolic methods have outperformed numeric ones regarding both accuracy and efficiency. Most approaches employ a bag-of-symbolic-words representation, but typically the word-length is fixed across all time series, an issue identified as a major weakness in the literature. Also, there are no prior attempts to use efficient sequence learning techniques to go beyond single words, to features based on variable-length sequences of words or symbols. We study an efficient linear classification approach, SEQL, originally designed for classification of symbolic sequences. SEQL learns discriminative subsequences from training data by exploiting the all-subsequence space using greedy gradient descent. We explore different discretization approaches, from none at all to increasing smoothing of the original data, and study the effect of these transformations on the accuracy of SEQL classifiers. We propose two adaptations of SEQL for time series data, SAX-VSEQL, can deal with X-axis offsets by learning variable-length symbolic words, and SAX-VFSEQL, can deal with X-axis and Y-axis offsets, by learning fuzzy variable-length symbolic words. Our models are linear classifiers in rich feature spaces. Their predictions are based on the most discriminative subsequences learned during training, and can be investigated for interpreting the classification decision.

We address the problem of real-time recommendation of streaming Twitter hashtags to an incoming stream of news articles. The technical challenge can be framed as large scale topic classification where the set of topics (i.e., hashtags) is huge and highly dynamic. Our main applications come from digital journalism, e.g., promoting original content to Twitter communities and social indexing of news to enable better retrieval and story tracking. In contrast to the state-of-the-art that focuses on topic modelling approaches, we propose a learning-to-rank approach for modelling hashtag relevance. This enables us to deal with the dynamic nature of the problem, since a relevance model is stable over time, while a topic model needs to be continuously retrained. We present the data collection and processing pipeline, as well as our methodology for achieving low latency, high precision recommendations. Our empirical results show that our method outperforms the state-of-the-art, delivering more than 80% precision. Our techniques are implemented in a real-time system that is currently under user trial with a big news organisation.

We present a framework for discriminative sequence classification where linear classifiers work directly in the explicit high-dimensional predictor space of all subsequences in the training set (as opposed to kernel-induced spaces). This is made feasible by employing a gradient-bounded coordinate-descent algorithm for efficiently selecting discriminative subsequences without having to expand the whole space. Our framework can be applied to a wide range of loss functions, including binomial log-likelihood loss of logistic regression and squared hinge loss of support vector machines. When applied to protein remote homology detection and remote fold recognition, our framework achieves comparable performance to the state-of-the-art (e.g., kernel support vector machines). In contrast to state-of-the-art sequence classifiers, our models are simply lists of weighted discriminative subsequences and can thus be interpreted and related to the biological problem -- a crucial requirement for the bioinformatics and medical communities.

The Web has the potential to become the world's largest knowledge base. In order to unleash this potential, the wealth of information available on the Web needs to be extracted and organized. There is a need for new querying techniques that are simple and yet more expressive than those provided by standard keyword-based search engines. Searching for knowledge rather than Web pages needs to consider inherent semantic structures like entities (person, organization, etc.) and relationships (isA, located In, etc.). In this paper, we propose NAGA, a new semantic search engine. NAGA builds on a knowledge base, which is organized as a graph with typed edges, and consists of millions of entities and relationships extracted from Web-based corpora. A graph-based query language enables the formulation of queries with additional semantic information. We introduce a novel scoring model, based on the principles of generative language models, which formalizes several notions such as confidence, informativeness and compactness and uses them to rank query results. We demonstrate NAGA's superior result quality over state-of-the-art search engines and question answering systems.

A common representation used in text categorization is the bag of words model (aka. unigram model). Learning with this particular representation involves typically some preprocessing, e.g. stopwords-removal, stemming. This results in one explicit tokenization of the corpus. In this work, we introduce a logistic regression approach where learning involves automatic tokenization. This allows us to weaken the a-priori required knowledge about the corpus and results in a tokenization with variable-length (word or character) n-grams as basic tokens. We accomplish this by solving logistic regression using gradient ascent in the space of all ngrams. We show that this can be done very efficiently using a branch and bound approach which chooses the maximum gradient ascent direction projected onto a single dimension (i.e., candidate feature). Although the space is very large, our method allows us to investigate variable-length n-gram learning. We demonstrate the efficiency of our approach compared to state-of-the-art classifiers used for text categorization such as cyclic coordinate descent logistic regression and support vector machines.

The presence of encyclopedic Web sources, such as Wikipedia, the Internet Movie \nDatabase (IMDB), World Factbook, etc. calls for new querying techniques that \nare simple and yet more expressive than those provided by standard \nkeyword-based search engines. Searching for explicit knowledge needs to \nconsider inherent semantic structures involving entities and relationships.\n\nIn this demonstration proposal, we describe a semantic search system named \nNAGA. NAGA operates on a knowledge graph, which contains millions of entities \nand relationships derived from various encyclopedic Web sources, such as the \nones above. NAGA's graph-based query language is geared towards expressing \nqueries with additional semantic information. Its scoring model is based on the \nprinciples of generative language models, and formalizes several desiderata \nsuch as confidence, informativeness and compactness of answers.\n\nWe propose a demonstration of NAGA which will allow users to browse the \nknowledge base through a user interface, enter queries in NAGA's query language \nand tune the ranking parameters to test various ranking aspects.

The World Wide Web provides a nearly endless source of knowledge, which is mostly given in natural language. A first step towards exploiting this data automatically could be to extract pairs of a given semantic relation from text documents - for example all pairs of a person and her birthdate. One strategy for this task is to find text patterns that express the semantic relation, to generalize these patterns, and to apply them to a corpus to find new pairs. In this paper, we show that this approach profits significantly when deep linguistic structures are used instead of surface text patterns. We demonstrate how linguistic structures can be represented for machine learning, and we provide a theoretical analysis of the pattern matching approach. We show the benefits of our approach by extensive experiments with our prototype system LEILA.