My research projects align with my interest in studying decision-making under uncertainty from a computational perspective and bridging the gap between predictions and actions. My PhD work focused on sequential decision analytics with applications in process science and has been published in top-tier conferences. Specifically, I explored complex problems such as utilizing Memory-Augmented Neural Networks, which have reasoning capabilities, to generate recommendations that aid businesses during their process execution from an operational perspective. Additionally, I developed a Clinical Decision Support System for Sepsis Management using Deep Reinforcement Learning to showcase how we can build agents that provide decision support to knowledge workers during the execution of knowledge-intensive processes.

Process Science

Process science is the interdisciplinary study of socio-technical processes over time. By socio-technical process, we mean a coherent series of changes that involve actions including humans and technologies, occurring at different levels. This broader discipline combines knowledge from information technology and management sciences to improve and run operational processes. I have written a survey paper providing a broad overview of techniques for supporting and improving business processes.

A few of my research projects were aimed at building tools and frameworks that provide decision support to process users at operational and strategic levels during the execution of knowledge-intensive processes. In terms of technologies, AI and ML have the potential to play a key role in operational and strategic business decision-making.

  • At the operational level, advanced ML techniques can provide recommendations to support business process executions. Deep learning architectures like LSTMs, MANNs, and Transformers offer methods for modeling sequential data problems. In DeepProcess: project, we applied Memory Augmented Neural Networks for tackling the challenging problem of Predictive Process Monitoring. see [Paper] [Blog_post]
  • At the Strategic level, AI can help firms come up with robust plans. We explore this in the Alpha-GS project: Alpha-GS - Decision making in adversarial settings using Game tree Search Combined With satisfiability(SAT) [Blog Post]
Decision Support Systems for Human-AI Collaboration:

Designing systems where AI predictions and recommendations can effectively assist human decision-makers while taking into account human expertise and preferences remains a challenge. During my PhD, I was particularly interested in the application of AI to assist clinicians in patient treatment while preserving the privacy of their sensitive data. The digitization of healthcare data, coupled with algorithmic breakthroughs in AI, is poised to significantly impact healthcare delivery in the coming years. Although scientific knowledge can guide interventions, there is a crucial need to swiftly navigate the space of decision-making policies to identify effective strategies that support patients throughout the care process.

Real-world environments are dynamic and can change over time. AI systems need to adapt their actions based on changing contexts and feedback. In MIMIC-RL I investigated the problem of developing a Clinical Decision Support System for Sepsis Management using Deep Reinforcement Learning. In our implementation, we sought to answer the following question: Given a patient's specific characteristics and physiological information at each time step as input, can our proposed framework learn an optimal treatment policy that prescribes the appropriate intervention (e.g., use of a ventilator) at each stage of the treatment process, in order to improve the final outcome (e.g., patient mortality)? see [Blog_post] for more details.

Data Privacy and Federated Learning:

Another significant challenge in this field is data privacy, particularly in sensitive domains, which hinders research and limits access to datasets. In one of my papers, I addressed this issue by applying Federated Learning and Differential Privacy techniques to extract process models from geographically distributed process logs.

Medical data is often geographically dispersed and distributed, and privacy concerns can prevent the construction of a centralized data warehouse. By employing techniques such as federated learning and differential privacy, we can extract process models from distributed healthcare process logs. In this paper, we explore the framework of federated learning in the context of distributed process mining.

Language Understanding and Search:

"80% of all information created today is unstructured (free text with little structural explanation). In 2017 alone the information created was expected to be greater than the previous 5000 years combined. In addition to sheer volume, the rate of information creation is accelerating rapidly, 10x per year by 2025, which is also the year each human is expected to interact with connected devices nearly 5000 times each day."

Recent advances in large language models (LLMs) present an opportunity to reimagine AI systems, using language as a medium for facilitating human-AI interaction. For instance, providing high-quality answers to medical queries necessitates an understanding of the medical context, recall of pertinent medical knowledge, and reasoning with expert information. I have been involved in various academic and industry projects related to information retrieval and knowledge discovery:

Other interests:

My recent focus has largely been on applied machine learning, even though i am been keeping an eye on progress in theoretical realm. Theoretical Machine learning is full of fascinating open problems like generalization, intelligent exploration vs exploitation, counter-factual evaluation, Meta-Learning and Sample Efficient Learning. Another challenge is that Decision-making systems need to be interpretable and explainable so that humans can understand and trust the actions suggested by AI models. Ensuring that the actions recommended by AI systems are fair and unbiased is crucial, especially in sensitive domains like healthcare, finance, and criminal justice. Lastly, Predictions made by many ML models are often based on correlations rather than causal relationships. Designing techniques and algorithms that can suggest actions based on predictions requires understanding how interventions will affect the outcome. This involves causal inference and counterfactual reasoning. For effective decision-making, understanding the causal impact of actions is crucial.

Progress in solving these problems can have a major impact on application domains like industrial automation, healthcare, and education. The decision-making side of machine learning has been relatively neglected but its equally important. It involves individual high-stake decisions, decisions in the context of multiple decision makers, explanations for decisions and dialog about decisions.

Great scientists tolerate ambiguity very well. They believe the theory enough to go ahead; they doubt it enough to notice the errors and faults so they can step forward and create the new replacement theory. If you believe too much you'll never notice the flaws; if you doubt too much you won't get started. It requires a lovely balance. - Richard Hamming