Theja Tulabandhula Ph.D.

Theja is an Assistant Professor at the University of Illinois Chicago. He obtained his PhD in computer science/machine learning from MIT CSAIL. He has 12+ years of experience in solving novel machine learning problems, and has built and deployed multiple ML solutions at scale. He has also co-authored 50+ research papers at top AI/ML related venues (e.g., JMLR, MLJ, UAI, AAMAS, WWW, Sensors, EJOR, MICCAI, ICASSP etc) and has been awarded 10+ patents.

His professional experiences include working with Apple, Citadel, State Street, Texas Instruments, AI Affinity, Headspace and others. His current research areas of interest include user behavior modeling, reinforcement learning and the use of machine learning techniques in impactful applications. He is motivated towards meaningfully helping improve AI driven service systems and positively contributing to the society.

Google Scholar Profile Linkedin Profile

Research Experiences

Theja conducts research in the areas of machine learning techniques (including supervised, online and reinforcement learning variants) and optimization techniques (both continuous and discrete).

User Behavior

Defining and estimating realistic quantitative user behavior models (both stochastic and strategic) in service systems such as healthcare (including digital health), transportation and digital platforms.

Decision Making

Designing scalable single shot optimization techniques that optimize decisions such as interventions/treatment plans in healthcare, routing and matching in transportation, and pricing and recommendations in digital platforms.

RL & Bandits

Designing well-behaved online and reinforcement learning methods for personalization in service systems by carefully accounting for the underlying domain specific structure.

Decision-aware ML

Enhancing machine learning and estimation processes in the presence of operational information and downstream decision preferences.

Sample Research Talk

Below is a video about user behavior modeling and multi-armed bandits. Similar videos can be viewed on his youtube channel.

Publication Links

In addition to the Google Scholar profile (also linked at the top), most manuscripts can be accessed from the pre-print archives and indices below. There are multiple links due to the interdisciplinary nature of his research.

Specific journal names have been omitted in the sections below for simplicity. Follow the links to know more about these and other papers.

Research Application Area: Transportation

Select research threads are summarized below.

Reinforcement learning for combinatorial optimization

We propose a deep reinforcement learning (DRL)-based approach to the dynamic on-demand crowdshipping problem in which requests constantly arrive in a crowdshipping system for pickup and delivery within limited time windows. We develop novel heuristics-embedded Deep Q-Network (DQN) algorithms that incorporate double and dueling structures as well as hard constraints, to train DRL agents.

Learning ride-sharing preferences

In this work, we focused on learning sharing preferences of participants in an online manner, and deciding how to group them to achieve maximum participant satisfaction (a combinatorial optimization setting). Under certain user behavior models, we show how statistically efficient learning and computationally efficient decision making can be achieved.

Pricing in ride-sharing when users are utility-maximizing

Here, we proposed the notion of sequential fairness and sequential individual rationality that capture mutual inconveniences (such as detours). Further, assuming users follow a particular choice model, we show how a firm (e.g., Uber) can price the exclusive and shared ride options to maximize revenue. We also investigated how offering fairness guarantees can influence market share (a collective user behavior) and improve profits, with extensive simulations using the New York City Taxi dataset.

Research Application Area: Digital

Select research threads are summarized below.

Thompson sampling when users have fatigue

We assume a model where users are shown recommendation lists, and are susceptible to both positional and fatigue effects. We show how a Thompson sampling based scheme can learn the relevant parameters and achieve sublinear regret with a milder dependence on the number of actions (quadratic in the number of recommendations versus exponential).

Bandit methods when users have repetition effects

Here we assume that users are susceptible to priming effects (for instance, repeating the same recommendation increases the propensity to buy) and show how a bandit algorithm can be designed to maximize revenue with near-optimal regret guarantees.

Optimizing recommendations under a multi-purchase choice model

We define a multi-purchase customer choice model within the random utility maximization framework and compare its fit to real world datasets. We further develop multiple optimization approaches to optimize the recommendation sets given such a choice model.

Research Application Area: Healthcare

Select research threads are summarized below.

Unifying functional and structural connectomes via Graph Neural Networks

Most published connectome studies have focused on either structural or functional connectome separately, yet complementary information between them, when available in the same dataset, can be jointly leveraged to improve our understanding of the brain. We propose a function-constrained structural graph variational autoencoder (FCS-GVAE) capable of incorporating information from both functional and structural connectome in an unsupervised fashion, leading to improvements in embeddings and downstream prediction/exploration tasks.

Exploration using payments in mobile health (mHealth)

We assume users follow a linear random utility model and focus on learning via additional payments while minimizing regret in a contextual bandit framework. We develop a variety of schemes, including modifications to existing contextual bandit algorithms to make them amenable to principal-agent settings, and give guarantees on total expected payments.

Routing patients to ER via reinforcement learning

We extend bandit methods to an RL setting where patient arrivals and service levels of ERs are not known a-priori and the goal is to learn a policy for routing that minimizes waiting times. The key idea here is to exploit the structure of the optimal policy of the unknown MDP to restrict the search space.

Current Students

Theja is currently collaborating with the following PhD students.

Teaching Experiences

A playlist containing a sample set of lectures from his deep learning course is embedded below. Most of the course materials are available on Github, and brief descriptions form them are also shown below the video.

Statistical Models and Methods for Business Analytics

The goal of this class is to cover the foundations of modern statistics and machine learning methods complementing the data mining focus of sibling courses. In other words, you will get up to speed with the requisite background, as well as the key theoretical underpinnings of modern analytics. We will do so through the lens of statistical machine learning. Lectures will be complemented with hands-on exercises.

Deep Learning and Modern Applications

Broadly, we will cover topics spanning deep learning and reinforcement learning. In particular, we will study popular deep learning architectures, their design choices and how they are trained. This will be motivated by business applications dealing with image, text and tabular data. Finally, we will look at online and reinforcement learning frameworks and their role in sequential decision making settings such as retail.

MLOps: A Gentle Introduction to Machine Learning Deployment

This practice-oriented course surveys modern best practices around getting machine learning (ML) models into production. We will learn multiple ways of operationalizing machine learning workflows and models in the context of the larger business end-goals. We will gain a better understanding of strategies for model management, monitoring and deployment. Further, we intertwine these topics with online experimentation techniques (A/B testing) and software engineering ideas such as version control, containerization, and continuous integration/continuous deployment.

Other Links

Theja's other interests include reading, hiking, biking, travel and many more. Some links to his social media are given below.