CDANs
pip install cdans
Constraint-based causal discovery for autocorrelated and non-stationary multivariate time series, with optimized conditioning sets and changing-module detection.
Stable on PyPI
Muhammad
Hasan Ferdous
Ph.D., Information Systems · UMBC · 2026
Causal discovery assumes time series are stationary, regularly sampled, and free of seasonal structure. Real data is none of these. I build methods that work anyway.
I completed my Ph.D. in Information Systems at the University of Maryland, Baltimore County, defending on July 22, 2026 under Dr. Md Osman Gani in the UMBC Causal AI Lab. My work has appeared at KDD, MLHC, AAAI, ICMLA, and IEEE PerCom Workshops, with further manuscripts under review at ICDM, NeurIPS, and AAAI.
A lagged causal graph. Edge colour encodes the lag τ at which the cause acts. This palette runs through the whole site.
The research program
Four assumptions that break, and what to do about each
Constraint-based causal discovery rests on assumptions about time that observational data rarely satisfies. My dissertation takes them one at a time. The first three are completed contributions; the fourth is where the work goes next.
-
Pathology 01
Autocorrelation
Conditioning on the full history is expensive and statistically weak. CDANs and eCDANs condition on identified lagged parents instead, cutting the search space and raising detection power.
-
Pathology 02
Non-stationarity
Causal structure that shifts over time defeats a single global graph. DCD separates trend, seasonal, and residual components, tests each with the right procedure, and recombines them.
-
Pathology 03
Seasonality
Removing seasonality before search discards information. SPC-CD conditions on a deterministic phase basis instead, which blocks seasonal confounding without deseasonalizing.
-
Pathology 04
Irregular sampling
Discrete lags assume a fixed grid. Ongoing work treats the process in continuous time so that gaps carry information rather than requiring interpolation.
Selected work
Publications
-
TimeGraph: Synthetic Benchmark Datasets for Robust Time-Series Causal Discovery
KDD 2025 · Proc. 31st ACM SIGKDD Conf. on Knowledge Discovery and Data Mining, pp. 5425–5435
@inproceedings{ferdous2025timegraph, author = {Ferdous, Muhammad Hasan and Hossain, Emam and Gani, Md Osman}, title = {TimeGraph: Synthetic Benchmark Datasets for Robust Time-Series Causal Discovery}, booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '25)}, pages = {5425--5435}, year = {2025}, doi = {10.1145/3711896.3737439} } -
CDANs: Temporal Causal Discovery from Autocorrelated and Non-Stationary Time Series Data
MLHC 2023 · Proc. 8th Machine Learning for Healthcare Conf., pp. 186–207
@inproceedings{ferdous2023cdans, author = {Ferdous, Muhammad Hasan and Hasan, Uzma and Gani, Md Osman}, title = {CDANs: Temporal Causal Discovery from Autocorrelated and Non-Stationary Time Series Data}, booktitle = {Proceedings of the 8th Machine Learning for Healthcare Conference (MLHC '23)}, pages = {186--207}, year = {2023}, url = {https://proceedings.mlr.press/v219/ferdous23a.html} } -
PreprintDCD: Decomposition-based Causal Discovery from Autocorrelated and Non-Stationary Temporal Data
arXiv:2602.01433 · Under review at IEEE ICDM 2026
News
Recent
- August 2026Ph.D. degree conferred by UMBC. On the academic job market for Spring 2027 and Fall 2027 starts.
- July 2026Defended my dissertation, Bridging Theory and Practice: Robust Causal Discovery from Autocorrelated, Non-Stationary, and Seasonal Time Series Data, at UMBC on July 22.
- July 2026SPC-CD: Seasonal-Phase Conditioned Causal Discovery for Multi-Period Time Series submitted to AAAI-27.
- 2026Composable Causality, a toolkit pairing time-series causal discovery with treatment-effect benchmarking, submitted to the NeurIPS 2026 Evaluations and Datasets Track.
- Spring 2026COEIT Research Day Student Award, UMBC, for the G-DCD poster.
- February 2026DCD preprint released (arXiv:2602.01433). The paper is now under review at IEEE ICDM 2026.
- 2026CDANs released as an open-source Python package:
pip install cdans(GitHub). - December 2025Causal Time Series Modeling of Supraglacial Lake Evolution in Greenland under Distribution Shift accepted at IEEE ICMLA 2025.
- August 2025TimeGraph published at KDD 2025, and the COEIT Summer Student Project Award ($5,000) from UMBC.
Software
Methods you can run
Every method I publish ships as code. A result nobody can reproduce is not much of a result.
TimeGraph
A synthetic benchmark suite for temporal causal discovery. Controlled causal structures, calibrated autocorrelation, non-stationarity, multiple noise families, and seasonality.
Published at KDD 2025