Machine Learning and Knowledge Discovery in Databases - European Conference, ECML PKDD 2022, Grenoble, France, September 19-23, 2022, Proceedings, Part IV

Tias Guns, Massih-Réza Amini, Stéphane Canu, Petra Kralj Novak, Grigorios TsoumakasEcml/pkdd 2023 · DOI · DBLP

Abstract

This article summarizes the proceedings of Part IV of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2022). Held in Grenoble, France, this volume captures a pivotal moment in the field, focusing on Reinforcement Learning, Multi-agent systems, and Privacy-preserving AI. With a selective acceptance rate and a strong emphasis on applied data science, the proceedings highlight the community's shift toward bridging the gap between theoretical research and real-world utility in a post-pandemic hybrid environment.

Introduction

The ECML PKDD 2022 conference represented a significant return to hybrid engagement for the international machine learning and knowledge discovery community. This volume, Part IV of the proceedings, serves as a comprehensive snapshot of the state-of-the-art as of late 2022. It documents 242 accepted papers that address foundational challenges and practical applications across a diverse range of domains.

Core Research Themes

The papers in this volume are primarily organized around three pillars: decision-making in dynamic environments, multi-agent coordination, and the preservation of data privacy.

Reinforcement Learning and Multi-Agent Systems

A significant portion of the volume is dedicated to Reinforcement Learning (RL) and its application to multi-agent systems. Technical innovations include:

  • Imitation Learning: The application of Sinkhorn distances to improve imitation learning performance.
  • Representation Learning: Utilizing style transfer techniques for more robust state representations.
  • Real-World Decision Problems: The use of multi-objective actor-critics for complex tasks like real-time bidding and deep RL for graph-structured data.
  • Interpretability: A growing trend toward "Interpretable Multi-agent RL," employing decision tree policies to move beyond "black box" models in high-stakes scenarios.

Privacy-Preserving AI and Federated Learning

Addressing the inherent tension between data utility and user privacy, several papers focus on Differential Privacy and Federated Learning. Key research areas include:

  • Risk Assessment: Investigating model inversion risks to prevent the leakage of sensitive training data.
  • Non-IID Data Challenges: Developing optimal mixture weights for federated learning systems where data is not independently and identically distributed across nodes.

Methodological Innovations

The proceedings also highlight broader methodological shifts in the field:

  • Causal Discovery: The "Machine Learning Best Paper Award" was given to research on Local Causal Discovery, signaling a move toward understanding cause-and-effect relationships rather than mere statistical correlations.
  • Geometric Deep Learning: The "Data-Mining Best Paper Award" explored infinite-depth Graph Neural Networks, enhancing the expressiveness of models dealing with geometric and structured data.
  • Scientific Integration: Active learning techniques were showcased in physical science applications, such as the detection of planetary magnetopause crossings, demonstrating the expanding reach of ML into specialized scientific domains.

Conference Impact and Results

The 2022 conference maintained rigorous standards for publication. The Research track saw an acceptance rate of 27.3%, while the Applied Data Science track was even more selective at 22.2%. This selectivity ensured that the 242 accepted papers represented high-quality contributions to the field. The event drew over 1,080 participants from 37 countries, reflecting the global scale and collaborative nature of modern AI research.

Conclusion

ECML PKDD 2022 (Part IV) underscores the field's ongoing evolution toward more applied, interpretable, and privacy-conscious machine learning. While the "Applied Data Science Track" successfully bridges some gaps between theory and practice, the editors acknowledge that fully resolving these complexities remains an ongoing mission. Future work will likely continue to focus on the trade-offs between model accuracy and privacy, as well as the deployment of complex RL agents in unpredictable real-world environments.

Summary updated

BibTeX
@inproceedings{guns2023machine,
  author = {Tias Guns and Massih-Réza Amini and Stéphane Canu and Petra Kralj Novak and Grigorios Tsoumakas},
  title = {{Machine Learning and Knowledge Discovery in Databases - European Conference, ECML PKDD 2022, Grenoble, France, September 19-23, 2022, Proceedings, Part IV}},
  booktitle = {Ecml/pkdd},
  year = {2023},
  doi = {10.1007/978-3-031-26412-2},
}
Figure