Category Archives: Machine Learning

New Research Paper – A Novel Multi-View Perspective on XAI for Time Series Classification 🤖🔍

Posted on by

The “black box” nature of many AI models causes hesitation in trusting their predictions, especially in sensitive fields like medicine and finance, where interpretability and accountability are essential. Overcoming this challenge has become a central focus in AI development. Time series classification (TSC) is a critical task with numerous real-world applications, such as detecting anomalies in ECG data for healthcare and identifying fault patterns in electronic signals for industrial manufacturing.

That’s why we are so excited to introduce our paper “MIX: A Multi-view Time-Frequency Interactive Explanation Framework for Time Series Classification” just accepted at NeurIPS 2025! 🎉 It’s our novel approach, designed to finally open up the black box and empower humans to truly understand and interpret complex time series models.

Current explanation way focus to one perspective of the data. 🧩

Most current explanation methods suffer from tunnel vision. They almost exclusively focus on a single perspective: the timeline. By analysing which time steps or segments are important, they often completely ignore crucial patterns hidden in the data’s frequency.

While a recent method, SpectralX, made progress by including frequency, it is still limited to a single, fixed configuration. Ultimately, relying on just one view of the data provides an incomplete and less reliable explanation of how a model truly makes its decisions.

An Overview of MIX: Multi-view, Interaction & Traversal 🚀

Our MIX Framework introduces a totally new concept to the AI research community: multi-view explanations. It’s built on three powerful, interactive ideas.

🖼️ Multi-view Explanations: Seeing the Full Picture

We have introduced a new problem in XAI: how do you explain a model’s decision from many different perspectives at once? In time series, we create these “views” using a powerful signal processing tool (Haar DWT) that acts like a set of different camera lenses. Some lenses capture the broad, long-term trends, while others zoom in on high-frequency details. MIX intelligently gathers the most important features from all views, giving you explanations at various levels of granularity.A diagram of a graph AI-generated content may be incorrect.

🤝 View-Interaction: Making the Views Work Together

This is where MIX truly shines. We’ve created the first-ever interactive mechanism where the views don’t just exist in isolation, but they talk to each other! The strongest, most confident explanation from one view helps to refine and improve the explanations in all the others. This connection ensures that every individual explanation becomes more faithful and robust.

🗺️ View-Traversal: Finding the Most Important Features

Finally, MIX goes on a treasure hunt. Instead of just giving you a ranked list of features from a single perspective (like traditional methods), it travels across all the views in a smart, greedy search. It identifies the absolute most critical features overall, no matter which view they came from. The result is a comprehensive “greatest hits” list that gives users a complete, holistic understanding of the model’s decision.

How Does This New Perspective Perform? 📊

Our research shows significant improvements in both faithfulness (how true the explanation is) and robustness (how stable it is). MIX proved its effectiveness across 11 diverse datasets and 3 different deep learning architectures compared to 4 state-of-the-art methods. The results are consistently clear. ✅

Why This is a Game-Changer 🚀

This is not just a fascinating research paper; it’s a practical tool with a huge real-world impact. For every Data Scientist and AI Engineer out there, MIX provides a golden opportunity to finally understand time series classifiers from multiple angles. ✨

PaCoBi: Scaling Parallelism and Convexity Hurdles in Bi-Level Machine Learning

Posted on by

The accuracy of artificial neural network (ANN) models depends on the ability to effectively and efficiently incorporate adequately large data sets into the model training. The training of these models may include not only the tuning of network weights and biases for minimising inference error, but also the tuning of network topologies and other hyperparameters that have impact on generalisation quality, robustness, and inference efficiency. These include features such as the sparse utilisation of network edges and neurons, and binarised and/or quantised restrictions on parameter and activation values. Thus, we pose the research objective of improving upon the ability to incorporate additional features into optimisation modelling of ANN training while retaining the scale at which training data may be incorporated.  These objectives require improvement in both the optimisation methodologies and in the parallelisation of the methods that underly the training of machine learning (ML) models under these multiple objectives.  

We explore various ML model formulations, including variants of SGD approaches, and also alternative optimisation models based on convexified and discretised models including mixed-integer linear reformulations of binarised/quantised models.  For these models, we consider the decompositions that are likely to yield effective application of parallelisation approaches. We address the challenges presented by the non-convexity, the combinatorial, and large-scale qualities of the problems as they arise in parallel algorithmic paradigms such as the alternating direction method of multipliers (ADMM), which recent literature results provide at least limited guarantees for various nonconvexity structures. Importantly, we address the mitigation of nonconvexity as it 1) impedes the coordination of loosely coupled subproblems in approaches such as ADMM and 2) as it leads to suboptimal locally optimal solutions. For this, we apply a variety of approaches based on combinatorial reformulations, applications of convexifications, and embeddings in combinatorial frameworks for tightening the convexification. The relationship between the different ANN training objectives intersects with the challenging areas of mixed-integer/combinatorial optimisation, multiobjective optimisation, and bilevel optimisation. Developments in these areas would enable the improved capacity to solve increasingly complicated ML models for improved versatility and robustness. 

PaCoBi is Brian Dandurand’s individual Marie Skłodowska Curie Fellowship (Horizon Europe project number 101153359), supported by UKRI EPSRC project grant EP/Z001110/1.

Sweeping AAAI’25 success

Posted on by

We have been fortunate to have 3 papers accepted at AAAI’25.

Hung and colleagues will present their work on explainability of time series classification. InteDisUX aims to create explanations that are accessible and meaningful to users (real people) by identifying subsequences of the time series that provide positive or negative influence on a prediction. It uses a segment-level integrated gradient to merge successive segments into variable-length segments with high faithfulness and robustness. Follow the paper here: https://pure.qub.ac.uk/en/publications/intedisux-intepretation-guided-discriminative-user-centric-explan or come visit Hung at poster #8580. This work is funded by the MSCA-DN network RELAX.

Zichi and colleagues will present their work on WaveletMixer, a new time series forecasting method that leverages wavelets to create a latent representation at multiple levels of resolution and phases. It creates distinct forecasting models for each resolution, where the relationships between different frequency domains are exploited to update each of the models. Zichi also introduces a new MLP model for timeseries forecasting that works well in this setting. Follow the paper here: https://pure.qub.ac.uk/en/publications/waveletmixer-a-multi-resolution-wavelets-based-mlp-mixer-for-mult or come visit Zich at poster #10198. Zichi is supported by a scholarship from the China Scholarship Council.

Kazi Hasan Ibn Arif is a PhD student at Virginia Tech who we collaborate with through the US-Ireland project ‘SWEET’ (USI-226). Kazi has developed a new technique to improve the computational efficiency of high-resolution Vision-Language Models. A VLM combines two models, one to generate language tokens from the image, followed by a large language model. The technique uses attention in the token generation model to selectively drop tokens according to predefined budgets. The paper is on arxiv: https://arxiv.org/abs/2408.10945. Come visit Kazi at poster #7547.

Accelerating Scientific Discovery Using Domain Adaptive Language Modeling

Posted on by

Scientific corpora, such as papers and patents, are great source of information. Incorporating this information into scientific discovery pipelines is a great challenge that could reduce the discovery costs and speed-up the process. Motivating by this fact and leveraging the recent advances of the Natural Language Processing (NLP) domain, we provide domain adaptive NLP methods that are able to understand the scientific domain and its specific characteristics and facilitate necessary tasks for the discovery process.