Bio & Experience
I'm a PhD candidate in Computer Science at the University of Oklahoma, advised by Dr. Lan and Dr. Hougen. My research focuses on machine learning, artificial intelligence, and optimization, with particular emphasis on federated learning, recommender systems, randomized algorithms, low-rank matrix factorization, and learning under missing data. I develop scalable ML systems that are both theoretically grounded and effective in practice. My work spans feature engineering, dimensionality reduction, and model optimization (PCA, autoencoders, learned feature selection), including neural network design choices such as activation functions, with pipelines achieving up to 10× runtime speedups while maintaining accuracy. I've validated these approaches on real-world datasets including MovieLens, FilmTrust, and BookCross. My expertise covers supervised and unsupervised learning, deep learning, and privacy-preserving distributed systems including differential privacy and large-scale matrix recovery. Previously, I completed an M.Sc. in Artificial Intelligence, publishing peer-reviewed research on community detection and recommender systems using multi-objective evolutionary algorithms.
News
- Paper accepted at ACM GECCO 2026 (San José, Costa Rica): "Evolving Multi-Channel Confidence-Aware Activation Functions for Missing Data with Channel Propagation"
- Paper accepted at SCRS ITAI 2026 (Boston, USA): "One-Shot Federated Low-Rank Matrix Factorization"
Research Interests
Active Researches
Randomized Least Square for Efficient Low-Rank Matrix Recovery
Under Review (2026)
Low-rank matrix recovery is a fundamental problem. Popular recovery techniques such as alternate least square (ALS) or nuclear norm minimization all require iterative optimization of the recovered matrix, which limits their efficiency. In this paper, we propose an extremely fast recovery technique based on randomized least square (RLSMR). It does not require iterative optimization, for it randomly assigns one low-rank factor and only optimizes another low-rank factor once (which are then used to recover the matrix). In experiment, we show RLSMR runs substantially faster than other techniques while maintaining competitive performance.
A Dragonfly and Ants at Work: A Hybrid Bio-Inspired Framework for Active Learning in Community Detection
Paper in Progress
We introduce a bio-inspired hybrid approach that combines Ant Colony Optimization and the Dragonfly algorithm for feature selection and community detection in complex networks. By integrating an active learning component, the framework reduces labeling costs while enhancing model robustness. Preliminary results on benchmark datasets (e.g., Zachary's Karate Club) demonstrate improved modularity, labeling efficiency, and stability compared to baseline methods.
Projects
Featured Project
A self-correcting RAG system built on a multi-agent architecture that autonomously handles academic research tasks. Routes queries intelligently across local papers, arXiv, and Semantic Scholar, generates citation-grounded responses with page-level attribution, and validates outputs through hallucination detection with automatic retry.
Python, LangGraph, LangChain, ChromaDB, MCP, Streamlit, Docker
An uncertainty-aware Random Forest implementation that handles measurement uncertainties in both input features and labels, making it significantly more robust in real-world scenarios where data inherently contains noise.
Python
Full-stack manufacturing tracker using Java + Azure SQL that takes you from customer order to cost accounting and job close-out.
Java, Azure SQL
Selected Publications
N. Shahabi Sani, Ferial Najiantabriz, Shayan Shafaei, Dean Frederick Hougen
Recently Accepted (2026)
We propose a method to handle missing data in neural networks by developing custom activation functions. Rather than using standard functions like ReLU that only consider feature values, we employ genetic programming to create multivariate functions incorporating the feature value, missingness indicators, and confidence scores. We introduce ChannelProp, an algorithm that tracks these reliability signals throughout the network. Testing on datasets with various types of missingness shows our approach improves classification performance compared to conventional methods that don't account for data quality indicators.
One-Shot Federated Low-Rank Matrix Factorization
N. Shahabi Sani, Shayan Shafaei, Shangqing Zhao, Qi Li, Dean F. Hougen, Chao Lan
Recently Accepted (2026)
Low-rank matrix factorization (LRMF) is a fundamental problem with broad applications. In recent years, federated LRMF has gained intensive interests due to the increasing ethical concerns on data privacy and advancement of IoT and edge computing infrastructure. However, current federated LRMF algorithms are based on gradient descent, which requires repeated communication between distributed clients and a center. Such repeated communication is extremely inefficient and has many downsides from the energy and security perspectives. In this paper, we propose the first one-shot federated LRMF algorithm called OFMF. It only requires one round of client-center communication, by leveraging the idea of random matrix factorization. Efficiency analysis suggests OFMF is more efficient than current algorithms in both communication and local computation. Through experiments on three real-world data sets, we show OFMF achieves similar or better factorization performance than prior algorithm, while substantially improves the overall factorization efficiency (e.g., 40 times faster). We also evaluate a differentially private version of OFMF and show it performs similarly, which means OFMF additionally gains differential privacy for free.
N. Shahabi Sani, Mohammad Manthouri, Faezeh Farivar
Studying the structure of the evolutionary communities in complex networks is essential for detecting the relationships between their structures and functions. Recent community detection algorithms often use the single-objective optimization criterion. One such criterion is modularity which has fundamental problems and disadvantages and does not illustrate complex networks' structures. In this study, a novel multi-objective optimization algorithm based on ant colony algorithm (ACO) is recommended to solve the community detection problem in complex networks. In the proposed method, a Pareto archive is considered to store non-dominated solutions found during the algorithm's process. The proposed method maximizes both goals of community fitness and community score in a trade-off manner to solve community detection problem. In the proposed approach, updating the pheromone in ACO has been changed through Pareto concept and Pareto Archive. So, only non-dominated solutions that have entered the Pareto archive after each iteration are updated and strengthened through global updating. In contrast, the dominated solutions are weakened and forgotten through local updating. This method of updating the Pheromone will improve algorithm exploration space, and therefore, the algorithm will search and find new solutions in the optimal space. In comparison to other algorithms, the results of the experiments show that this algorithm successfully detects network structures and is competitive with the popular state-of-the-art approaches.
N. Shahabi Sani, Ferial Najian Tabriz
Recommender systems are among the most important parts of online systems, including online stores such as Amazon, Netflix that have become very popular in the recent years. These systems lead users to finding desired information and goods in electronic environments. Recommender systems are one of the main tools to overcome the problem of information overload. Collaborative filtering (CF) is one of the best approaches for recommender systems and are spreading as a dominant approach. However, they have the problem of coldstart and data sparsity. Trust-based approaches try to create a neighborhood and network of trusted users that demonstrate users' trust in each other's opinions. As such, these systems recommend items based on users' relationships. In the proposed method, we try to resolve the problems of low coverage rate and high RMSE rate in trust-based recommender systems using k-means clustering and ant colony algorithm (TBRSK). For clustering data, the k-means method has been used on MovieLens and Epinion datasets and the rating matrix is calculated to have the least overlapping.
Teaching
Graduate Teaching Assistant, University of Oklahoma, School of Computer Science
- Mentored students through office hours and targeted help sessions, emphasizing rigorous reasoning and clear problem decomposition.
- Provided algorithm walkthroughs (e.g., how/why an approach works, edge cases, and complexity reasoning) to strengthen students' conceptual understanding.
- Evaluated assignments and exams with standardized rubrics; delivered constructive feedback to improve future performance.
- Supported course logistics by answering questions, resolving confusion early, and ensuring consistent expectations across sections.
Cryptography
Spring 2026
Discrete Structures
Fall 2025
Algorithm Analysis
Spring 2025
Discrete Structures
Fall 2024
Machine Learning
Spring 2024
Theory of Computation
Fall 2023
Skills & Tools
Languages
ML / AI
Tools
Contact
For inquiries, feel free to email me at shahabi@ou.edu or use the form below.