icdm2025-conference-notes

🌟 Day 3 – Takeaways from ICDM 2025, Washington, DC 🌟

---

8:30–9:00 – ICDM Award Announcements

Two awards especially caught my eye:

🏆 Best Paper:
“Attributed Hypergraph Generation with Realistic Interplay Between Structure and Attributes”
Authors: Jaewan Chun, Seokbum Yoon, Minyoung Choe, Geon Lee, Kijung Shin (KAIST)

🏆 Best Student Paper:
“HyHG: A Temporal Hypergraph Contrastive Learning Framework for Biomedical Hypothesis Generation”
Authors: Amir Shariatmadari, Sikun Guo, Nathan Sheffield, Aidong Zhang (University of Virginia),
Kishlay Jha (University of Iowa)

Both works rely on hypergraphs to tackle modern challenges in data mining and knowledge discovery—another reminder of how deep graph theory and mathematics remain central to computer science.

---

A Fun Personal Story

Earlier in the conference, during a morning coffee break, I decided to be more proactive and just sit down at a random table and start a conversation.

I ended up talking with a young researcher from South Korea—turns out he was Jaewan, a first-year PhD student working on hypergraph generators and the best paper winner!

I mentioned my own earlier work on hypergraphs with Marco Valtorta, Lincoln Lu, and Zhiyu Wang from my PhD days. It was a bit of an awakening moment that I should revisit this topic more seriously in my current proposal work.

---

Keynote – Dr. Aidong Zhang (University of Virginia)

Title: “Optimizing External and Internal Knowledge of Large Foundation Models for Scientific Discovery”

My takeaway:

• LLMs can be more innovative in hypothesis generation than individual humans because they tap into a vast and diverse pool of scientific knowledge.
• But their ideas are often less feasible in practice.

Human judgment is still crucial to:

• Filter
• Refine
• Ground these hypotheses in realistic constraints.

---

Plenary Panel – Where Is Data Mining Going?

Three big questions structured the discussion:

1️⃣ Data mining vs. foundational models
2️⃣ Can LLMs defend the hypotheses they generate?
3️⃣ What will data mining look like in the next decade?

Jilles Vreeken made a provocative statement:

“Data mining is dead” — but its core objective, knowledge discovery, is very much alive.

Another key message:

• Science is more than LLMs.
• We shouldn’t put all our eggs in one basket or spend all our talent only tweaking foundation models.

There was even some discussion about whether the conference itself should eventually rebrand around “knowledge discovery.”

---

Evening Banquet – International Spy Museum 🕵️‍♂️

The day ended with the banquet and award ceremony at the International Spy Museum.

I had a long and insightful conversation with Dr. Mohamed Nafea (Missouri S&T) about:

• Causality
• Fairness

I also shared my struggles with securing external funding. He told me about a colleague who finally succeeded after 21 submissions.

A powerful reminder for all early-career researchers:

• 🔁 Be resilient.
• 💡 Stay confident in your ideas.
• ⏳ Persistence often matters as much as brilliance.

More reflections and photos from ICDM 2025 coming soon—stay tuned! 🚀📊

Opening session of the morning plenary — the hall begins to fill for the main award and panel segments.

“Major ICDM Activities in the Past 25 Years” — a retrospective on milestones such as double-blind reviewing and the KAIS journal partnership.

Announcement of the 2025 ICDM Research Contributions Award — honoring Prof. Zhi-Hua Zhou (Nanjing University) for seminal ensemble learning work.

10-Year Highest-Impact Paper Award — *“Fusing Similarity Models with Markov Chains for Sparse Sequential Recommendation.”*

“Thank you for this incredible honor!” — a moment of appreciation from the awardees to the ICDM community.

Ruining He reflecting on a decade of progress — from early recommender systems research to industrial-scale ML at Google DeepMind.

Slide introducing the ICDM 2025 Tao Li Award Presentation — honoring outstanding contributions in data mining.

Jundong Li presenting the ICDM 2025 Tao Li Award.

Research Theme — *Learning from Structured Data*: an integrative view of algorithmic foundations, trustworthiness, reasoning, and applications.

Detailed structure of the *Learning from Structured Data* theme — linking fairness, robustness, causal reasoning, and LLM reasoning layers.

Round 2 panel – “LLMs can propose hypothesis—but can they defend them?”

“Learning from Structured Data – Foundations (T1)” overview timeline.

“Learning from Structured Data – Trustworthiness (T2)” highlights.

“Learning from Structured Data – Reasoning (T3)” with causal and LLM reasoning.

Best Paper Award poster and authors at the poster session.

Talk – “HyHG: A Temporal Hypergraph Contrastive Learning Framework for Biomedical Hypothesis Generation.”

Best Paper Award talk – “Attributed Hypergraph Generation with Realistic Interplay Between Structure and Attributes.”

Panel slide – “LLMs — End or Evolution of Data Mining?”

ICDM panel on LLMs and the future of data mining.

Round 1 panel – “Data Mining vs. Foundation Models: A Disciplinary Crossroads.”

ICDM 2025 panelists and organizers in front of the IEEE ICDM 2025 conference screen.

Roundtable panel on stage discussing the question “LLMs—End or Evolution of Data Mining?”

Round 1: “Data Mining vs. Foundation Models: A Disciplinary Crossroads” – how LLMs blur the boundary between model and data.

Round 2: “Scientific Discovery and Data Integrity” – asking whether LLMs can propose hypotheses and also defend them.

Round 3: “The Future: Rebuilding Data Mining for the LLM Era” – envisioning the next decade if data mining is reborn.

Panel discussion slide: “What should data miners start doing before the next wave eats us too?”