icdm2025-conference-notes

🌟 Day 1 Takeaways from AI4TS @ ICDM 2025 in Washington, DC! 🌟

I spent an incredibly stimulating day at the AI for Time Series Analysis: Theory, Algorithms, and Applications (AI4TS) workshop (link), diving deep into the cutting edge of time series (TS) modeling.

A central theme that resonated with me across several talks was the growing, yet still evolving, role of causality in TS analysis.

Key Highlights & Insights

1️⃣ “Four Pillars for PTS Analysis” & “TSFMs for Perception” – Agnes F. Liu (University of Maryland)

A good and concise summary of the state-of-the-art in TS models, offering a fresh perspective on integrating these models for complex systems.
The concept of Temporal Foundation Models (TSFMs) as the pillar for Perception (as shown in the slides) is a nice one and gives a modular view of TS pipelines.

2️⃣ “TS-CausalNN: Learning Temporal Causal Relations from Non-linear Non-stationary Time Series Data” – Omar Faruque (University of Maryland)

This talk sat right at the intersection of:

Causal discovery
Time series
Continuous optimization

Very much in the spirit of NOTEARS (Zheng et al., 2018) but tailored to temporal structure. This is one of the kinds of direction that connects my own interests in causal modeling and TS.

3️⃣ Keynotes by Dr. Panagiotis Papapetrou (Stockholm University) & Dr. Rex Ying (Yale University)

Dr. Papapetrou touched on counterfactuals in time series, but interestingly, causality itself mostly appeared as “future work,” even though counterfactuals are inherently causal concepts by nature.
Dr. Ying discussed benchmarking TS models, including issues like poor alignment due to missing causal links. From my Pearl-style causality lens, the way causality entered these benchmarks still feels non-causal—or at least not yet explicitly framed in that language.

Overall Impression

There are huge opportunities at the intersection of time series modeling and causal reasoning, both in terms of:

Theory (causal structure in temporal models, counterfactuals for TS), and
Practice (benchmarks, evaluation, robustness across domains).

Stay tuned for more updates from ICDM 2025!

Tags:
#ICDM2025 #AI4TS #TimeSeries #Causality #DeepLearning #DataScience #FoundationModels

Slide: Time Series Foundation Models overview — *Agnes F. Liu’s slide on Time Series Foundation Models and multimodal FMs.*

Slide: Pillar 3 TSFMs for Perception — *“Pillar 3: TSFMs for Perception” — positioning TSFMs as the perception layer in physical time-series analysis.*

Slide: Related works on temporal causal discovery — *Related works slide for TS-CausalNN, comparing assumptions and shortcomings of existing temporal causal methods.*

Slide: Key contributions of TS-CausalNN — *Key contributions of TS-CausalNN — learning lagged and contemporaneous causal links in non-stationary time series.*

Hotel breakfast at ICDM 2025 — *Day 1 breakfast fuel: granola, fruit, muffin, yogurt, coffee, and orange juice before AI4TS.*

Lunch plate at ICDM 2025 — *Day 1 lunch — colorful buffet plate between workshop sessions.*

Ali Javidian at the ICDM 2025 backdrop — *Formal shot at the IEEE ICDM 2025 backdrop.*

Another photo at the ICDM 2025 backdrop — *Another ICDM 2025 backdrop photo to remember Day 1.*

Floppy-disk style notebook, front view — *Retro floppy-disk style notebook — conference swag with a CS flavor.*

Floppy-disk notebook with pen — *Back view of the floppy-disk notebook with matching pen.*

Time-Series Foundation Models — *Time Series Foundation Models.*