As shown, with unbounded outcome benchmark dataset (IHDP), our approach provides a valid "INDIVIDAULIZED" bound!
📄 Paper: arxiv.org/abs/2601.17160
💻 GitHub: github.com/yonghanjung/...
📦 Install: pip install itbound
Causal inference needs strong assumptions 😔
However, BOUNDING CAUSAL EFFECTS should not need strong assumptions. 😃
`itbound` gives data-driven causal bounds under assumption-lean settings: unmeasured confounding, unbounded outcomes, no sensitivity parameters, etc.
📦 Install: pip install itbound
Specifically, our result shows that the proposed CATE estimator for the front-door (FD-DR, FD-R) outperforms the plug-in front-door estimator, which empirically evidenced the sample efficiency and robust behavior against bias.
Front-door (FD) enables causal effect identification through an observed mediator even when treatment-outcome confounding is unobserved.
Our work provides estimators that achieves sample efficiency and allowing personalized treatment effects when treatment-outcome confounding is unobserved.
Our paper “Debiased Front-Door Learners for Heterogeneous Effects” was accepted to ICLR 2026.
- Paper (arXiv): arxiv.org/abs/2509.22531
- Reproducible code: github.com/yonghanjung/...
Quick start:
pip install fd-cate
fdcate demo --outdir ./fdcate-demo
#ICLR2026 #CausalInference #MachineLearning
Yonghan Jung: Debiased Front-Door Learners for Heterogeneous Effects https://arxiv.org/abs/2509.22531 https://arxiv.org/pdf/2509.22531 https://arxiv.org/html/2509.22531
Thrilled to share our new paper!
📄 Paper: arxiv.org/abs/2509.22531
💻 Code: github.com/yonghanjung/...
We develop the first orthogonal ML estimators for heterogeneous treatment effects (HTE) under front-door adjustment, enabling HTE identification even with unmeasured confounders.
If you're interested in working with me, feel free to reach out at yhansjung@gmail.com.
I'm excited to share that I'll be joining the School of Information Sciences at UIUC as an Assistant Professor this Fall (ischool.illinois.edu/news-events/...). If you're interested in causal inference and its applications to trustworthy AI and healthcare, join me & let's work together!
PhDone 🎓 I’ve successfully defended my thesis!
Huge thanks to my amazing advisor Elias Bareinboim and committee—Jennifer Neville, Jin Tian, Yexiang Xue, and @idiaz.bsky.social.
Grateful to collaborators, colleagues, lab mates, friends, neighbors—and above all, my wife, kid, and family!
New paper alert (hey, I can't doom scroll all the time): This one's on doing causal inference with "microlevel data" where we suspect that the treatment has spatial spillover & temporal carryover effects. We illustrate our new approach + package w/ application to US counterinsurgency efforts in Iraq
📌Interesting way of using copula method for the sensitivity analysis in causal inference.
Reinforcement learning has led to amazing breakthroughs in reasoning (e.g., R1), but can it discover truly new behaviors not already present in the base model?
A new paper with Zak Mhammedi and Dhruv Rohatgi:
The Computational Role of the Base Model in Exploration
arxiv.org/abs/2503.07453
It looks interesting!
I really enjoy reading this paper. On a perspective of causal inference researcher, I agree that ML's real-world impact relies on science theory, because understanding causal mechanisms requires domain knowledge or theoretical assumptions. ML without theory simply leads us nowhere.
link 📈🤖
Adaptive Experimentation When You Can't Experiment () arXiv:2406.10738v1 Announce Type: cross
Abstract: This paper introduces the \emph{confounded pure exploration transductive linear bandit} (\texttt{CPET-LB}) problem. As a motivating example, often online services cannot directly assig
👉 Join our #CIIG seminar next month for an Introduction to Mechanism Learning
👉 Mechanism learning proposes using front-door causal bootstrapping such that ML models learn causal rather than "associational" (or spurious) relationships
See abstract and register: turing-uk.zoom.us/meeting/regi...
@pedrosantanna.bsky.social onlinelibrary.wiley.com/doi/10.1111/... biostatistics literature will use PO notation to describe the relevant objects. Just treat RL as MDP with unknown transitions (it's true RL doesn't use PO notation - it gets cumbersome and many key objects relate to the Bellman eqn)
I've decided to collect my DiD materials in a single place.
psantanna.com/did-resources
There, you will find
- 14 lectures of my comprehensive DiD course
- Shorter lectures/talks I have given on DiD
- My DiD R/Stata/Python packages
- Some DiD checklists
- DiD materials from my friends
Enjoy!
Merry Christmas, friends and colleagues! Hope you all have wonderful days with joys! 🎄
Looking ahead, my future direction will explore:
1️⃣ High-dimensional, online streaming datasets.
2️⃣ Multi-modal data (e.g., text, images).
3️⃣ Robust causal inference with uncertainty quantification.
My past work focuses on estimating causal effects identifiable from graphs, with applications in xAI and healthcare. This includes advancing methods to handle multi-domain experimental data, distributional treatment effects, and designing computationally efficient estimators.
Excited to share that I’m on the academic job market! I’ve been fortunate to work with Elias Bareinboim on causal inference, developing causal effect estimators using modern ML methods. Published in ICML, NeurIPS, AAAI, & more. Details: www.yonghanjung.me
In sum, our work provides a computationally efficient and statistically robust estimator for various covariate adjustment estimands, including cases where no such estimators previously existed.
Come see our poster and let us chat more!
Next, we developed Double-machine learning (DML)-based estimators for the UCA-class and provided finite sample guarantees, showing that it achieves doubly robustness and scalability (i.e., computational efficiency).
The UCA class incorporates a functional in a form of a product of various conditional probabilities. It includes the front-door adjustment, Verma’s equation, S-admissibility, Effect of treatment on the treated, soft-intervention, and many other practical causal estimands.
In this work,
1. We define a function class called "Unified Covariate Adjustment (UCA)" that incorporates various covariate adjustments; and
2. We developed a double machine learning (DML)-based estimator for the UCA-classes and provided finite-sample learning guarantees.
We will present our work "Unified Covariate Adjustment for Causal Inference” (joint work with Jin Tian &
Elias Bareinboim) at #NeurIPS2024!
- Wed (12/11) from 11am - 2pm
- Poster Session 1 (East Hall A-C) #4901
- Link: openreview.net/pdf?id=aX9z2...
Come and see us!
I am attending NeurIPS 2024 this Tuesday through Sunday. I am also in the academic job market this year (www.yonghanjung.me). Happy to discuss potential opportunities! Get in touch if you’d like to chat! #NeurIPS2024
