Sanup S Araballi

Explore my research, interests, and life activities through this personal website showcasing my work.

Discover

“There is an art, it says, or rather, a knack to flying.The knack lies in learning how to throw yourself at the ground and miss.” - Douglas Adams, Hitchhiker’s Guide to the Galaxy.

The website is under construction. More details will be added soon. Please check back in some time.

Research Focus

My research focuses on developing a unified framework for modelling multi-source information diffusion in dynamic, evolving networks where node adoption thresholds adapt over time to internal diffusion outcomes and external environmental shocks. This includes understanding and modelling intrinsic node properties affecting cascade propagation, with planned extensions into Graph Neural Networks (GNNs) and Explainable AI (XAI).

Education

§ Syracuse University, Syracuse, NY, USA
Expected Graduation: May 2026
§ Dissertation Topic: "Multi-Source Diffusion in Evolving Adaptive Networks"
§ Advisors: Dr. Chilukuri K. Mohan and Dr. Sucheta Soundarajan
§Courses Taken: Artificial Neural Networks, Reinforcement Learning for AI, Ubiquitous Learning, Machine Learning with Graphs, Network Science, Ethics of Machine Learning

Life Activities

Discover my life activities that reflect my diverse interests, showcasing how I balance research, academia, and personal pursuits to create a fulfilling and enriching lifestyle.

Ph.D. in Computer Science & Engineering 2020 – Present

Research Experience

PhD Research Intern, Air Force Research Lab, Rome, NY Summer 2024

Developed a novel fuzzy rule-based framework to enhance the interpretability of Reinforcement Learning (RL) agents in continuous action spaces, enabling transparency for 'black box' neural network models.
Successfully translated the complex decision-making processes of a Soft Actor-Critic (SAC) RL agent into a transparent Fuzzy Knowledge Base (FuzzyKB) composed of human-readable IF-THEN rules, effectively creating a "digital twin" for explainability.
Engineered the framework to capture state-action-reward trajectories from RL agent training (e.g., in LunarLanderContinuous-v2) and employed Takagi-Sugeno-Kang (TSK) fuzzy partitioning for robust state-space representation and rule extraction.
Demonstrated the framework's efficacy by reducing over 26,000 captured experiences from SAC training into a concise set of 74 human-readable FuzzyKB rules, showcasing the potential to explain complex neural network behaviour with significantly fewer rules.
Addressed the challenging problem of explainability in continuous action spaces, a key distinction from approaches focused solely on discrete actions, thereby broadening the applicability of XAI in RL.
The resulting framework and findings were documented in a research paper (related work "Crisp Distance Metrics for Fuzzy Rules," co-authored with C. Mohan, is being revised for new submission) and presented as a poster ("Fuzzy Rule-Based Framework for Explainable Reinforcement Learning"), with support from AFRL's Griffis Institute.

Signage on a wall reads 'Cybercrime Analysis & Research Alliance @ NTU (CARA)' in bold, white letters against a dark, textured surface. A corridor with partial ceiling panels is visible, and there's some greenery in the background.

PhD Research Assistant, Dr. Venkata Gandikota’s Lab, Syracuse, NY. Summer 2025

Working with Dr. GV and his team to develop frameworks on several fronts. This work still is in its nascent stage.
We are working on Graph Reconstruction from Cascades, Bi-level Optimization Using Evolutionary Methods, Louvain Communities in Deep Neural Networks.

Led the computer science aspects of an interdisciplinary research project with the S.I. Newhouse School of Public Communication and ComScore, focusing on TV ad viewing retention and effective ad curation. My role encompassed data analysis, feature engineering, model development, and experimental design.
Designed and implemented machine learning models, including Random Forests and Neural Networks (using error back-propagation), to analyze extensive second-by-second ComScore TV Essentials viewing data for various programs and ad placements.
Developed predictive models achieving over 80% accuracy in identifying key attributes (e.g., ad pod number, program viewed percentage, originality, ad duration) that contribute to TV ad viewing retention.
This research contributed to the development of an ad curation model aimed at maximizing TV ad audiences and resulted in three peer-reviewed conference publications, including an award-winning paper at the Broadcast Education Association (BEA) Conference.
Contributed to identifying optimal program and ad attributes for maintaining ad viewership, such as focusing on the first ad pod, higher program viewing percentage, and fewer ads per break.
Currently spearheading PhD dissertation research focused on developing a novel, adaptive framework for modelling multi-source information diffusion in dynamic, evolving networks, incorporating user behavioral traits and adaptive node thresholds, with planned extensions into Graph Neural Networks and Explainable AI.

Research Assistant, Syracuse University, Syracuse, NY (Summer 2019 - Present)

About Sanup Araballi's Work

Welcome to the personal website of Sanup Araballi, where you can explore his research, academic achievements, interests, and life activities that shape his professional journey.