I am passionate about applying trustworthy machine learning, through the lens of optimization, to solve real-world decision support problems, and am actively looking for research/job opportunities in both fundamental and applied machine learning.
Bio Sketch
I am currently pursuing a Ph.D. degree at the University of British Columbia in Vancouver, Canada. Prior to this, I received my bachelor’s and master’s degrees from China University of Petroleum in Qingdao, China, in 2019 and 2022.
My research focuses on machine learning optimization (optimal lightweight and interpretable decision tree) and its decision support applications in diverse industrial scenarios including both traditional and renewable energy systems, and control problems.
I mainly develop a set of tree-based algorithms (decision tree, random forest and reinforcement learning with decision tree policy) and address the following four questions, especially for real-world industrial applications: (1) Can a single tree outperform an entire forest? (2) Do we really need hundreds of machine learning models for tabular data-driven applications? (3) Do we have to rely on heavy black-box convolutional neural networks for image classification? (4) Can reinforcement learning and learning-based control be interpretable as decision tree?
Beyond these machine learning optimization studies, recently working on test time training and adaptation for large language model (LLM and MoE) using sparse learning.
Recent News
-
Received the "Research Award" (only 15 recipients).
-
Received the "NeurIPS Scholar Award" for NeurIPS 2025.
-
A paper on learning linear model predictive control, which successfully applies our proposed differentiable decision tree algorithm, was accepted by the IEEE Conference on Decision and Control (leading conference in control).
-
It took me 10 years, but I married my college sweetheart, Xiao Yang — the greatest achievement of my Ph.D. and my life.
-
Two papers detailing successful engineering applications of our proposed decision tree algorithm were accepted by SPE ATCE 2025 (leading conference in petroleum engineering).
-
Received the scholarship of "Dabrowski & Shepherd Award in Environmental Engineering".
-
A paper on cognitive computing for drilling target decision recommendation was accepted by Journal of Petroleum Science and Engineering.
Research Highlights
Fundamental Machine Learning
Differentiable Decision Tree via "ReLU+Argmin" Reformulation
Advances in Neural Information Processing Systems 2025. Spotlight Paper.
This study proposes RADDT, a novel differentiable decision tree to address two key challenges: high accuracy and differentiability. Its enables large-scale, gradient-based optimization, making it effective even for million-scale datasets through distributed multi-GPUs acceleration.
Large Language Models
UReFTA: Breaking the Weight Reconstruction Bottleneck in Tensorized Parameter-efficient Fine-tuning.
IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.
This study presents a tensor-based PEFT approach with a single tensor-rank parameter for large models that achieves extremely low trainable parameter counts by leveraging the low-rank structure of weight tensors.
Decision Support
A Decision Support Engine for Infill Drilling Attractiveness Evaluation Using Rule-based Cognitive Computing Under Expert Uncertainties
Journal of Petroleum Science and Engineering. 2022.
This study developes a rules set-based reasoning engine for drilling target recommendation and well placement optimization.
Do We Really Need Hundreds of Machine Learning Models in Industry?
SPE Annual Technical Conference and Exhibition. 2025.
This study challenges recent practice in tabular data-driven petroleum applications by proposing a decision tree-based paradigm with 26 applications. It emphasizes practical field demands of accuracy, interpretability and lightweight structure.
Process Monitoring & Fault Detection
Visual Process Monitoring of Biomass Conversion Reactors Using Transfer Learning and Generative AI.
Computer and Chemical Engineering. 2025.
This study proposes a GenAI-based data augmentation and tailored transfer learning strategy to address data scarcity, specifically for the faulty smoke events during biomass-to-biochar conversion processes.
Stop Using CNN: Knowledge Distillation for An Interpretable and Lightweight Decision Tree in Rod Pump Working Condition Diagnosis
SPE Annual Technical Conference and Exhibition. 2025.
This study challenges recent reliance on black-box, heavy models in dynamometer card-based fault diagnosis by proposing an interpretable and lightweight decision tree alternative. The proposed paradigm is also applied to other image classification tasks.
Control
Exact Learning of Model Predictive Control Laws Using Oblique Decision Trees with Linear Predictions
IEEE Conference on Decision and Control. 2025.
This study presents a successful application of our proposed differentiable decision tree (tree with linear predictions) in explicit model predictive control.
Contact Me
-
maoq@student.ubc.ca
-
2360 East Mall
Vancouver, BC, Canada, V6T 1Z3