About Me

Hello! 👋

I'm Alireza Keshavarzian, a PhD candidate in Electrical and Computer Engineering at the University of Toronto, working under the supervision of Professor Shahrokh Valaee. My research sits at the intersection of machine learning, signal processing, and healthcare applications, with a particular focus on developing efficient and interpretable methods for time series classification.

I believe that the most impactful machine learning research combines theoretical rigor with practical applicability. This philosophy drives my work on random representation learning—methods that achieve state-of-the-art performance without expensive training procedures—and my commitment to deploying these techniques in real clinical settings where they can make a genuine difference in patient outcomes.

Research Focus

My doctoral research centers on representation learning for time series data, particularly in healthcare contexts where data is often limited, noisy, and high-stakes. I've developed several novel methods that address fundamental challenges in this space:

RASTER (Random Kitchen Sink with Randomized Threshold Exceedance Rate) introduced the concept of temporal localization to random kernel methods, enabling models to capture where discriminative information occurs in time series—not just what patterns exist. This work demonstrated that simple, randomized approaches can outperform complex deep learning models when properly designed for the temporal structure of the data.

Building on this foundation, I developed HIERVAR and CoMIND, hierarchical feature selection frameworks that can reduce the dimensionality of random representations by over 90% while maintaining or improving classification accuracy. These methods address a critical practical concern: random methods generate thousands of features, many of which are redundant or uninformative.

More recently, my work on CURVE explores how contrastive learning principles can be applied to feature selection without labeled data, opening new possibilities for unsupervised and semi-supervised time series analysis.

Healthcare Applications

Since 2020, I've been working as a Data Scientist at University Health Network (UHN), one of Canada's largest academic health science centers. This role has given me invaluable experience translating research ideas into production systems that clinicians actually use.

My primary project involves building ML pipelines for respiratory oscillometry—a technique that measures lung mechanics through small pressure oscillations. The challenges here are representative of healthcare ML more broadly: limited labeled data, class imbalance, noisy measurements, and the need for interpretable, uncertainty-aware predictions.

I developed OscilloFusion, a multi-channel attention-based architecture that fuses different oscillometry signals to achieve ~80% classification accuracy—a 5 percentage-point improvement over conventional approaches. Beyond raw performance, I've implemented calibrated uncertainty estimates and built interactive tools that let clinicians explore model predictions and explanations, fostering trust and enabling human-AI collaboration.

Beyond Academia

My path to academia wasn't direct. Before starting my PhD, I spent several years in industry, which fundamentally shaped how I approach research problems.

As a Software Engineer Team Lead at Atrovan, I architected multi-tenant IoT platforms, built event-driven microservices with Kafka and Golang, and led cross-functional engineering teams. This experience taught me the importance of scalability, maintainability, and the gap between prototype code and production systems.

At SnappFood (Iran's largest food delivery platform), I worked on recommendation systems and delivery time prediction, learning how to design A/B tests, interpret business metrics, and communicate technical results to non-technical stakeholders.

Most recently, I completed an ML internship at DraftAid, where I developed graph neural networks for CAD automation—a fascinating application of ML to structured, geometric data. This project achieved 94% recall on annotation recommendations and demonstrated how machine learning can augment expert workflows rather than replace them.

Research Philosophy

I'm drawn to research that is both theoretically grounded and practically useful. Too often, machine learning research optimizes for benchmark performance without considering computational costs, interpretability, or deployment constraints. I try to ask: Can this method be understood? Can it be trusted? Can it be deployed?

Random representation methods appeal to me because they challenge the assumption that powerful ML requires powerful (and expensive) learning. By understanding why random features work, we gain insights into the structure of learning problems that transfer across methods and domains.

I'm also committed to reproducible research. I evaluate methods on complete benchmark suites (like the full 112-dataset UCR archive), report statistical significance, and share code. Science advances when others can build on your work.

Personal Interests

Outside of research, I enjoy building side projects that solve real problems. My recent LLM-powered job application tracker combines LangGraph, RAG pipelines, and careful prompt engineering to help job seekers manage applications and tailor resumes—technologies I learned by doing.

I'm based in Toronto, a city I've come to love for its diversity, its tech scene, and its surprisingly good food options for late-night coding sessions.