Vahid Ashrafi, Ph.D. Candidate in Information Systems and Data Analytics

Bio

Vahid Ashrafi is a Ph.D. candidate in information systems and data analytics at Stevens Institute of Technology, with extensive experience in behavioral economics, cognitive modeling and the application of AI in decision-making. He has also worked as a research fellow, corporate leader and academic, focusing on cognitive biases, human-computer interaction and decision support systems.

Skillset

Vahid Ashrafi's skillset includes extensive experience in large language models (LLMs), machine learning, data analytics and AI-driven cognitive modeling, alongside expertise in programming languages such as Python and R, business modeling, financial analysis, project management and statistical research design.

Dissertation Summary

Decoding Decisions: An AI-Driven Exploration of Cognitive Errors and Decision-Making Processes

My dissertation explores cognitive errors—systematic deviations from rational judgment, including both cognitive biases and logical fallacies—and their influence on decision-making. Cognitive biases are mental shortcuts that simplify decision-making but can lead to irrational outcomes, while logical fallacies are errors in reasoning that compromise the validity of arguments. Together, these cognitive errors distort decision-making in a range of real-world contexts, from personal finance and health to organizational strategy and leadership. The primary goal of this dissertation is to map, detect, and mitigate these errors using advanced computational methods, including natural language processing (NLP) models and decision-support systems.

The research begins by developing a computational taxonomy of cognitive errors using sentence-based transformer models. This taxonomy provides a structured understanding of how different cognitive biases and logical fallacies relate to each other, illustrating the connections and dependencies among these errors. By analyzing a dataset of cognitive error descriptions, real-world examples, and academic abstracts, the research generates a high-dimensional map of how these errors co-occur across various decision contexts. This framework forms the basis for further exploration into how cognitive errors manifest in real-world decisions.

Building on this theoretical foundation, the research shifts to real-world decision-making scenarios by analyzing a large dataset of advice-seeking scenarios from online sources. These scenarios, which span topics like finance, relationships, and career decisions, are classified using NLP techniques to generate a map of distinct advice categories. This classification offers insights into the common types of decisions individuals face and how they seek guidance for those decisions. The scenarios are analyzed independently of the cognitive error taxonomy, serving as real-world examples of decision-making processes.

Next, the dissertation links cognitive errors to these advice-seeking scenarios by creating a matrix that assigns relevancy scores to each cognitive error for each scenario. This matrix quantifies the likelihood that a particular cognitive error influences a given decision. Key metrics, such as the overall relevance of cognitive errors, their frequency across scenarios, and an irrationality score for each decision, are calculated to provide a detailed analysis of how cognitive errors impact everyday decision-making. This approach allows for a deeper understanding of which biases are most prevalent and which types of decisions are most prone to irrational outcomes.

The focus then shifts to the individual level by developing a personalized system for detecting cognitive errors based on a person’s written and spoken communication. Using NLP models, the system analyzes personal emails, social media posts, and conversations to create personalized cognitive error profiles. These profiles highlight the most common biases that affect an individual’s decision-making, providing actionable insights for self-improvement, professional development, and education. This personalized approach enables individuals to recognize and address their cognitive biases in real time.

The analysis is further extended to group-level decision-making by aggregating personalized cognitive error profiles across communities, organizations, and online groups. By examining the collective decision-making processes within these groups, the research identifies dominant cognitive biases that influence group dynamics and decisions. This group-level analysis is particularly valuable for organizations seeking to improve decision-making by addressing the shared biases that distort collective choices.

The dissertation concludes by evaluating a decision advisor bot designed to reduce cognitive errors in real-world decision-making scenarios. Using advice-seeking scenarios, the bot presents respondents with multiple alternative solutions generated by a large language model (LLM). If respondents select a less rational alternative, the bot provides feedback on the cognitive errors underlying their choice and prompts them to reconsider. The effectiveness of the bot is measured by comparing the irrationality scores of the initial and final decisions, demonstrating how the bot helps reduce cognitive biases and improve decision quality.

Overall, my dissertation offers a comprehensive exploration of cognitive errors in decision-making, combining theoretical analysis with practical tools to detect and mitigate biases. By integrating machine learning, NLP, and decision-support systems, the research provides new insights into how cognitive errors can be understood and addressed in both personal and collective decision-making contexts. The findings have broad implications for fields such as behavioral economics, cognitive psychology, and decision-support systems, offering practical solutions for improving decision-making in everyday life.

Academic Advisor

Jordan Suchow