Causal structure in time series that misbehave
Machine learning is good at finding statistical regularities and poor at explaining them. That gap matters wherever a decision depends on knowing why something happened, not just that it tends to co-occur. My work develops causal discovery methods for complex time series, and the open-source software that makes them usable by other people.
The work sits between statistics, machine learning, and information systems. My undergraduate training in statistics at the University of Dhaka supplies the inferential foundation, my graduate training at UMBC supplies the orientation toward deployable research, and the Causal AI Lab under Dr. Md Osman Gani supplied the algorithmic depth.
My dissertation, Bridging Theory and Practice: Robust Causal Discovery from Autocorrelated, Non-Stationary, and Seasonal Time Series Data, was defended in July 2026. It is organized around a single question: what has to change in constraint-based causal discovery when the temporal assumptions it rests on do not hold.
Three methods, three broken assumptions
Autocorrelation CDANs and eCDANs
Constraint-based algorithms condition on the full history of a process. Under autocorrelation that is both computationally expensive and statistically inefficient, because the conditioning set is padded with variables that carry no new information. CDANs (MLHC 2023) identifies lagged parents explicitly and conditions on those, which shrinks the search space and raises detection power. It also detects changing modules, components whose causal structure shifts as the distribution shifts, and considers contemporaneous and lagged parents together when assessing those shifts.
eCDANs (AAAI 2023) refines the lagged-adjacency search and keeps
the accuracy of CDANs at lower computational cost, which makes it
usable on larger datasets and in constrained settings such as
bedside monitoring and edge devices. CDANs is on PyPI:
pip install cdans
(GitHub).
Non-stationarity DCD
Time series in climate, finance, and operations carry overlapping seasonal cycles and long-run trends, and those structures confound standard discovery algorithms. DCD (arXiv:2602.01433, under review at IEEE ICDM 2026) decomposes each series into trend, seasonal, and residual parts and applies a different inference procedure to each: stationarity testing for trend, kernel-based dependence measures for the seasonal part, constraint-based discovery for the residual. The component-level findings recombine into one multi-scale causal graph. On synthetic benchmarks and real climate data it improves consistently on state-of-the-art baselines under strong non-stationarity and autocorrelation.
Seasonality SPC-CD
The usual treatment for seasonality is to remove it before the causal search runs. That discards information and can distort the dependence structure left behind. SPC-CD (under review) conditions on seasonality instead. It detects the periods present in the data, builds a deterministic multi-harmonic phase basis from them, and includes that basis in every conditional independence test, which blocks seasonal confounding without deseasonalizing the series.
Two components support it: a structured lag set combining short lags with integer multiples of the detected periods, and a per-edge routing rule that sends each candidate edge to a linear or a nonlinear independence test based on a residual nonlinearity diagnostic. Validation runs over five synthetic generators, each isolating one failure mode, plus a real air-quality study in which the method recovers all eight pre-registered chemistry and physics hypotheses where no baseline recovers more than six.
Benchmarks TimeGraph and Composable Causality
Progress in causal discovery has been held back by inconsistent evaluation. Methods are tested on different datasets under simplified conditions, which makes comparison unreliable and hides the conditions under which each approach fails. TimeGraph (KDD 2025) generates synthetic datasets with controlled causal structure, calibrated autocorrelation, non-stationarity, several noise families, and seasonality. Composable Causality (under review) extends the program from structure recovery to decisions, pairing discovery with treatment-effect estimation so a graph can be scored by the quality of the intervention estimates it supports.
Applications and collaborations
Methods earn their keep on other people's problems. Most of the applied work below is collaborative and my contribution sits on the causal modeling side.
Climate. I am a co-author on two projects run through the UMBC Causal AI Lab with collaborators at the University of Colorado Boulder: a causal deep-learning framework for Arctic sea ice prediction (IEEE PerCom Workshops 2025) and a causal time series model of Greenland supraglacial lake evolution under distribution shift (IEEE ICMLA 2025). Both were led by Emam Hossain; I contributed the causal discovery and time series components.
Wearable sensing. With collaborators at Miami University in Ohio I contributed to ClassyGlass, a multimodal benchmark for activity and mobility analysis from smart eyewear (repository).
Plant biology. A collaboration now getting under way between the UMBC Causal AI Lab, plant biologists at UMBC, and a systems biology group at the Pontificia Universidad Católica de Chile, applying machine learning to RNA sequencing data and to a large corpus of the plant science literature.
Healthcare. I have applied causal discovery to physiological time series and to clinical datasets, looking for parameters that are actual intervention targets rather than correlational risk markers.
Where the program goes
-
Thrust 01
Streaming and non-stationary systems
Extending CDANs, DCD, and SPC-CD to streaming and concept-drifting data. The first direction, already under development, treats irregularly sampled series in continuous time rather than forcing them onto a regular grid, which is the main open problem my dissertation leaves. Also: online discovery under bounded memory, causal change-point detection, missingness from sensor failure rather than missing-at-random, and causal reasoning joined to anomaly detection.
-
Thrust 02
Causal inference inside deep learning
Deep models capture complex structure and explain little. Causal models are explicit and scale poorly. This thrust builds hybrid neuro-causal frameworks that embed discovery in representation learning, keeping predictive performance while gaining the ability to reason under distribution shift. Good territory for undergraduate projects and senior capstones.
-
Thrust 03
An open-source ecosystem
Three public artifacts already exist: the CDANs package, TimeGraph, and the Composable Causality toolkit. Next come semi-synthetic benchmarks from finance, climate, healthcare, and security data, reference implementations of competing baselines, and workshop organization at KDD, AAAI, and adjacent venues. Software gives a department visibility out of proportion to its cost and is sustainable work for undergraduate assistants.
-
Thrust 04
Interdisciplinary collaboration
Methods earn their keep on real problems. Three areas transfer directly: healthcare and digital health, where the target is actionable intervention rather than risk scoring; climate and environmental science, extending existing collaborations on ice sheets, sea ice, and teleconnections; and security analytics, where analysts need root-cause reasoning rather than anomaly scores.
Proposal strategy
| Source | Program | Focus | Years |
|---|---|---|---|
| NSF | CRII, then CAREER | Foundational causal AI methods and open-source infrastructure | 1–3, 3–5 |
| NSF | SaTC, CICI | Causal and adversarially robust ML for security analytics | 2–5 |
| NIH | R03, R21, R-series | Trustworthy clinical decision support and physiological time series | 2–5 |
| DoE, NOAA | Data science programs | Causal discovery for climate and environmental modeling | 2–5 |
| Industry | Sponsored research | Finance, enterprise IT, regional technology partners | Ongoing |
Students in the program
The program is built so students can contribute something real. Benchmarks and software packages produce work a well-supervised undergraduate can own, and the artifacts that result, published code, datasets, reproducibility audits, carry weight on graduate applications and industry resumes. My targets are at least one peer-reviewed publication per year with a student co-author by year three, and at least one student per year placed into an industry internship or a graduate program.