icdm2025-conference-notes
🌟 Day 2 – Takeaways from ICDM 2025, Washington, DC 🌟
Another intense and inspiring day at ICDM 2025. A few highlights and reflections from today:
9:00–10:00 – Keynote by Dr. John Quackenbush (Harvard University)
Title: “Why Networks Matter: Embracing Biological Complexity”
My takeaway:
- Biology is not easy, and there is no free lunch theorem here.
- Heavy reliance on simulated data can be misleading for biology—Dr. Quackenbush clearly prefers real, messy data over clean simulations.
10:00–10:30 – Meeting Dr. Jilles Vreeken
I met Dr. Jilles Vreeken in person, the Program Chair of ICDM 2025. He and Dr. Wei Ding (UMass Boston) have introduced several innovations into this year’s conference format. The presented stats are interesting and speak for themselves.
We had a great half-hour conversation about causality, especially:
- His seminal work on MDL for causal inference, and
- His current projects on causal invariants and domain adaptation.
Very aligned with my own research interests.
10:30–12:00 – Panel A: “The Future of Probabilistic Modeling in Data Mining and AI”
The panel focused on challenges such as:
- Scalability
- Explainability (the “why” questions)
- The mathematical foundations of LLMs and other advanced models
In the Q&A, I raised the point that many of these concerns are exactly what causality is designed to address—yet causality was barely mentioned.
My takeaways from the responses:
- Research often follows funding and mainstream trends.
- Formal axiomatizations of causality are still relatively young (only a few decades old), and there is uncertainty about what a full “causal turn” in AI would imply.
15:00–16:00 – Main Track S16: Interpretability
I attended three talks in the interpretability session. While technically interesting, I left wanting more:
- Little clarity on how the proposed models are interpretable.
- Almost no discussion of how interpretability was quantified.
A reminder that interpretability is often claimed but rarely operationalized.
16:00–17:00 – Main Track S15: Time Series I
One talk on few-shot domain adaptation for time series really caught my attention—a novel framework designed for few-shot unsupervised domain adaptation.
I’ll definitely be reading that paper more carefully for new ideas related to my work on causal and robust TS methods.
Poster Session
The poster display was, unfortunately, the least engaging part of the day:
- Many authors were absent due to scheduling conflicts.
With better time allocation and fewer overlaps, this could be a much richer venue for deep, one-on-one scientific conversations—far more than the few minutes available in Q&A at the end of talks.
Closing Thought
Overall, Day 2 reinforced for me how much room there is to bring causal thinking, interpretability, and careful evaluation into mainstream data mining and AI, especially in biology and time series.
More reflections coming soon. Stay tuned! 🚀📊
