Haoyue Dai    

Hi! I am a PhD student at CMU Philosophy, fortunately advised by Prof. Kun Zhang. I am in the CMU-CLeaR (Causal Learning and Reasoning) group.

My research interests are centered around causality. On the causal discovery side, I care about identifiability with latent variables, with relaxed assumptions, and with heterogeneity. Specifically, I am curious about what we can really get from just observational data. On the machine learning side, I aim to develop methods for causal representation learning and causality-inspired explainability. On the application side, I thrive to solve real-world problems from domains including physics, biology, and social science. My ultimate goal is to enable computers to reason about the world.

Prior to CMU, I completed my undergraduate at IEEE Honor Class and Zhiyuan Honor Program, Shanghai Jiao Tong University (2017-2021).

hyda [AT] cmu.edu  /  Google Scholar  /  GitHub

The first to deal with dropouts fully nonparametrically: conditional independence (CI) relations in the data with dropouts, after deleting the samples with zero values for conditioned variables, are identical to the CI relations in the original data.

The first to tackle local causal discovery in cyclic models. By independent subspace analysis, all the local causal structures and coefficients in the equivalence class are identified (intersecting cycles allowed). A regression variant is given for acyclic cases.

Transformed Independent Noise (TIN) condition: use linear transformation of variables to pursue independence! It entails graphical criteria for causal discovery with latent vairables, and specifically, in the presence of measurement error.

The first supervised causal discovery approach on discrete observational data. The identifiability is guaranteed, and empirically, it remarkably outperforms other sota algorithms in terms of accuracy, reliability, robustness and tolerance.

Skeleton (undirected causal graph) learning via supervision. Vicinal graphs are proposed to address the domain shift.

Using Perturb-seq data from 67 gene knockouts, we discover the causal relationships among different genes and cell states, and accordingly propose novel gene knockouts (in the remaining ~20,000 genes) to shift as many T cells as possible into cancer-fighting states. The proposed knockouts are then validated through real-lab experiments.

From students' learning history at an online learning platform (time series, with e.g., changing mechanism, data missingness), we discover the causal relationships among knowledge constructs, and estimate the conditional average treatment effects.

• causal-learn
  I do code quality control and speedup (e.g., probably the fastest Chi-sq test, the fastest PC, 1/3 faster KCI).


• I listen to...
• I read...
• I record...

thanks jon!