Hi there, I'm Arman Behnam!

I'm a Computer Science PhD candidate at Illinois Institute of Technology, advised by Dr. Binghui Wang, and Co-Founder of Quis, Inc. My research is on causal explanation, representation, and reasoning in graphs and language models, published at NeurIPS (2025, 2026) and ECCV (2024). At Quis, I work on memory for AI agents: how an assistant can come to understand a person over long-running conversations, and how to measure whether it does.

Previously, I was an AI Research Scientist Intern at Mayo Clinic, where I used graph neural networks to explain the progression of mild cognitive impairment toward Alzheimer's disease, in work co-authored with Dr. Ronald C. Petersen and published at IEEE BIBM 2024. Before my PhD, I was the Founding Head of Data Science at Golrang Industrial Group, one of Iran's largest industrial groups with 100+ subsidiaries, where I built and led its machine learning division.

I started teaching AI in 2018 as a lecturer at Jahad Daneshgahi Sharif, where I founded one of Iran's first deep learning and neural network training programs.

I serve on the Editorial Board of the American Journal of Lifestyle Medicine and review for NeurIPS (Silver Reviewer, 2026), ICML, and AAAI, as well as for JAMA Network, the Journal of General Internal Medicine, and The Journal of Primary Prevention. See my Web of Science profile for my full publication record.

I'm a Founding Researcher at a Quis.Inc startup pioneering GNN-powered search technologies. My research focuses on pushing the boundaries of Graph Neural Networks, with breakthrough methods that have outperformed all state-of-the-art approaches in GNN explanation, published at top-tier venues including ECCV and NeurIPS. I'm a CS PhD at Illinois Institute of Technology (IIT). I am fortunate to be advised by Dr. Binghui Wang, as a CS research scientist at IIT.

Previously, I was an AI Research Scientist at Mayo Clinic, where my GNN research contributed to breakthrough methods in Alzheimer's disease research, co-authored with world-renowned expert Dr. Ronald C. Petersen. I also served as Founding Head of Data Science at Golrang Industrial Group (Iran's largest industrial group with 100+ subsidiaries), where I established and led their machine learning-driven data science division.

My journey in AI began as one of the youngest college lecturers at Jahad Daneshgahi Sharif, where I founded Iran's first and most prominent Deep Learning/Neural Network teaching platform in 2018 — years before such educational resources became mainstream. This early commitment to democratizing AI education reflects my belief in the transformative power of knowledge sharing.

I serve on the Editorial Board of the American Journal of Lifestyle Medicine and regularly review for leading publications including Journal of General Internal Medicine and The Journal of Primary Prevention. View my Web of Science author profile for complete publication record.

Email Scholar ResearchGate LinkedIn

I am broadly interested in redefining causal Machine Learning (The knowledge of cause and effect for learning) to artificial intelligence as a backbone for theoretical computer science. I am also applying my new algorithms and frameworks on real-world datasets to explore the information within images, graphs, social networks, and protein chains. More technically, I am now working on causal representation learning for language models.

• Description: A measure-theoretic framework for anti-causal representation learning where labels cause observations (Y → X ← E), using two-level representations: low-level φ_L captures causal dynamics from data while high-level φ_H learns environment-invariant features. It handles both perfect and imperfect interventions through interventional kernels and enforces invariance via environment independence (R₁) and causal structure consistency (R₂) regularizers.
• Outcome: First measure-theoretic approach providing theoretical guarantees for anti-causal learning with OOD generalization bounds and environmental robustness proofs. Results demonstrate 99%+ accuracy on synthetic datasets (CMNIST, RMNIST, Ball Agent) and 84.4% on real-world medical data (Camelyon17), significantly outperforming state-of-the-art methods across all invariance metrics while successfully removing spurious environment-specific correlations.

• Description: Explore potential causal explanation of MCI progression by temporal patient data, including chronic diseases, biomarkers, and genetic information, into a graph structure to capture causal effects within variables.
• Outcome: Identified a causal subgraph with informative variables including hypertension, arrhythmia, congestive heart failure, coronary artery disease, stroke, lipid-related issues, and sex.

• Description:A GNN causal explainer based on a graph's causal structure and its corresponding neural causal model. It correctly does the graph classification via causal inference (based on interventional data) by understanding and quantifying cause-and-effect relations between observable variables.
• Outcome: it's the first GNN causal explainer. We leverage the neural-causal connection, design the GNN neural causal models, and train them to identify the causal explanatory subgraph. Our results show the effectiveness of CXGNN and its superiority over the state-of-the-art association-based and causality-inspired GNN explainers

• Description: From the beginning of the outbreak until April 30, 2020, over 90,000 confirmed cases of COVID-19 have been reported in Iran. Due to socio-economic problems of this disease, it is required to predict the trend of the outbreak and propose a beneficial method to find out the correct trend.
• Outcome: A dataset including the number of confirmed cases, the daily number of death cases, and the number of recovered cases is compiled. New rates such as weekly death rate, life rate, and new approaches to mortality rate and recovery rate are created by Gaussian functions.

• Description: Breast cancer is the second reason for cancer mortality. Approximately 30%- 40% of patients suffering from breast cancer will experience recurrence and 10%-15% of them were reported to die of cancer metastasis. Early diagnosis or prediction of metastasis will reduce mortality rate and treatment cost.
• Outcome: Multi-Layer Perceptron Outperforms other methods to predict metastasis.

• Description: Different tumor features are available in various datasets for breast cancer detection. Filtering those to obtain an accurate diagnosis is time-consuming and challenging. Machine learning algorithms are beneficial for finding a significant relationship between various features and malignant tumors.
• Outcome: The suggested framework’s performance works perfectly due to the selection of more appropriate features by the Extra Trees algorithm. Using this framework, breast cancer recovery and therapy will be more successful. Moreover, to evaluate the performance of the proposed framework, it has been implemented on three other medical datasets.

• A great need is felt to analyze the applicability of AI-enabled IoT technologies in energy systems in response to the call for smart and autonomous architecture for multicarrier energy grids. Due to this, the current chapter gives an informative description of new AI-based IoT frameworks concentrating on architecture, elements, variables, applications, and also its challenges. At first, an insight into IoT basics is provided that helps readers to go through the next sections. The required infrastructure for deploying the aforementioned technologies is discussed, explaining electronic and IoT components. Energy internet as the IoT in energy grid’s output has tremendous features that are appropriate for the grid’s performance evaluation, which is explained in this chapter.

• With the development of new artificial intelligence and data science (DS) technologies, their applications for implementing analysis in the Internet of Things (IoT) energy systems are becoming more important for smart grids (SGs). DS approaches integrated with IoT data collection protocols in energy sectors can help in improving efficiency in such areas. Data gathering by sensors is an important step in IoT-based energy grids when it comes to a large amount of data. In the real-time data collection process, the frequency of data rises to become big data and the analysis needs new methods to manage and evaluate this data. The outcome of these analytics comes up as SG intelligence so the system becomes smart, which is depicted as demographics, figures, and informative dashboards.