About

Nowadays, many practical decision tasks require to build models based on data which included serious difficulties, as imbalanced class distributions, a high number of classes, high-dimensional features, a small or extremely high number of learning examples, limited access to ground truth, data incompleteness, or data in motion, to enumerate only a few. Such characteristics may strongly deteriorate the final model performances. Therefore, the proposition of the new learning methods that can combat the aforementioned difficulties should focus on intense research. The main aim of this workshop is to discuss the problems of data difficulties, identify new issues, and shape future directions for research.

Program

LDD, as a half-day event, will consist of two 90-minute sessions, lasting from 9:00 to 10:30 am and from 11:00 to 12:30 pm separated by a 30-minute coffee break.

Session 1 (9:00–10:30)

09:00–09:05 LDD Organizing Committee; Opening
09:05–09:50 TBA; Keynote
09:50–10:10 Lynn Houthuys; Multi-view mid fusion: a universal approach for learning in an HDLSS setting
10:10–10:30 Rubén González Barriada, David Masip; Adversarial Training for Suppressing Eye Disease Damage Bias in Retinal Image Quality Assessment

Session 2 (11:00–12:30)

11:00–11:20 Caoilfhionn Ní Dheoráin, Ellen Rushe, Kathryn Dane, Will Connors, Caithríona Yeomans, Anthony Ventresque; Try and Try Again: A Case Study on Tackle Detection with Limited Repurposed Data
11:20–11:40 Nicolas Salvatici, Billy Peralta, Alvaro Soto; Visual Prompting for Object Counting: A Hybrid Approach for Low-Data Scenarios
11:40–12:00 Julis Neumann, Robert Lange, Yuni Susanti, Michael Färber; Optimizing BERT-based Models for Multi-Label Sentiment Analysis in Short Texts
12:00–12:20 Abdul-Kazeem Shamba, Kerstin Bach, Gavin Taylor; ReMargin: Revisiting Contrastive Learning with Margin-Based Separation
12:20–12:30 LDD Organizing Committee; Conclusion of the workshop

Keynote

Towards Holistic Continual Learning: Bridging Knowledge Retention and Adaptation to Changes

Continual learning has recently been regarded as a cornerstone of artificial intelligence, promising systems that can accumulate knowledge over time while adapting to ever-changing environments. Yet most existing methods approach the problem in fragments — focusing either on mitigating catastrophic forgetting, or on enabling adaptation under concept drift, but rarely on their integration. In this keynote, I will argue for a holistic perspective on continual learning, one that unifies memory retention, adaptive generalization, and robustness to real-world change. I will discuss how knowledge can be dynamically reorganized to balance stability with flexibility, allowing models to evolve rather than collapse under novelty. By highlighting both the challenges and opportunities of holistic continual learning, I hope to spark cross-disciplinary discussion on how we can move from isolated techniques towards cohesive, real-world continual learning systems that retain, adapt, and thrive in dynamic environments.

Bartosz Krawczyk

Bartosz Krawczyk is an assistant professor in the Chester F. Carlson Center for Imaging Science at the Rochester Institute of Technology, where he heads MLVision Lab. Dr. Krawczyk's current research interests include machine learning, continual and lifelong learning, data streams and concept drift, class imbalance, and explainable artificial intelligence. He has authored more than 70 journal papers and over 100 contributions to conferences. Dr. Krawczyk coauthored the book Learning from Imbalanced Data Sets (Springer 2018). He was a recipient of prestigious awards for his scientific achievements, such as the IEEE Richard Merwin Scholarship, IEEE Outstanding Leadership Award, and Amazon Machine Learning Award and Best Paper Award at CLVISION CVPR Workshop.

Key dates

All deadlines are at the end of the day specified, anywhere on Earth (UTC-12).

Full paper submission deadline: ~~2 June 2025~~ ~~16 June 2025~~ 23 June 2025 (final extension)
Requests for consideration of papers rejected from the main conference 14 July 2025
Author notification date: ~~10 July 2025~~ 15 July 2025
Publication of the final workshop schedule: 8 August 2025
Early registration deadline: 3 September 2025
Workshop: 25 October 2025, Morning

Submission instructions and conference proceedings

Workshop LDD follows all requirements of the ECAI 2025 main conference. Papers must be written in English, be prepared for double-blind review using the ECAI LaTeX template, and not exceed 7 pages (plus at most 1 extra page for references). Excessive use of typesetting tricks to make things fit is not permitted. Please do not modify the style files or layout parameters. Last year's Workshop Proceedings were published in Proceedings of Machine Learning Research.

In addition to regular paper submissions, the LDD Workshop may accept papers rejected from the main conference purely based on the main track reviews, which must be submitted by athors as a separate PDF file. We invite authors to submit a request for their rejected paper to be considered by 14 July 2024.

Go to the Submission System

Topics

Learning from imbalanced data

You try to build a model, but it is biased towards the class that is better represented in the dataset.

Learning from imbalanced data streams, including concept drift management

The situation turns out to be even more difficult than in the previous case, because the data arrives (potentially) forever.

Learning from multi-view/multimodal data

You solve the problem of the curse of dimensionality through space decomposition and ensemble methods.

Automated machine learning

As in meta-learning, you try to give the method full control over the learning process.

Life-long machine learning

You already have a working model, but it turns out that it should solve a new task. And you really don't want to train it from the ground up.

Learning with limited ground truth access

You have experts to label the data, but there are a million objects and only three experts.

Learning in a open set

You're training your model to tell dogs from cats, but you also want to know what happens when you show it a raccoon.

Learning from high dimensional data

In the general case, you have a very large number of features in the set, but you don't want to solve this problem with multi-view approaches.

Learning with a high number of classes

Sometimes there are more classes than objects in a set.

Learning from massive data, including instance and prototype selection

You are trying to manage the problem of a very large dataset by initially sorting it out and finding the most valuable instances.

Learning based on limited data sets, including one-shot learning

It turns out that your data set is not massive. On the contrary, it covers only a few cases. What are you doing?

Learning from incomplete data

Or maybe the data set is not too small, but it turns out to be extremely leaky?

Case studies and real-world applications

Share your struggles with the real datasets!

Organizers

We’re researchers from Department of Systems and Computer Networks, which since 25 years conducts fundamendal research on Machine Learning models in difficult scenarios. We are from Poland.

Paweł Zyblewski

Assistant Professor at the Department of Systems and Computer Networks, Wroclaw University of Science and Technology, Poland

Paweł Ksieniewicz

Associate Professor of Computer Science at the Department of Systems and Computer Networks, Wroclaw University of Science and Technology, Poland.

Michał Woźniak

Professor of Computer Science at the Department of Systems and Computer Networks, Wroclaw University of Science and Technology, Poland.