Connor Lawless

I am currently a postdoc at Stanford Management Science & Engineering advised by Madeleine Udell and Ellen Vitercik. Previously, I did my PhD at Cornell in Operations Research and Information Engineering, where I was fortunate enough to be advised by the incomparable Oktay Gunluk .

My work leverages tools from computational integer programming to build scalable algorithms for interpretable and fair machine learning. Prior to joining Cornell, I completed my undergraduate at the University of Toronto studying industrial engineering and spent a year working at the Royal Bank of Canada building deep reinforcement learning based trade execution algorithms. I was also lucky enough to spend the summer of 2021 working remotely with the Applied AI group at IBM research on interpretable clustering, and the summer of 2023 working at Microsoft Research with the Human Understanding and Empathy Group and the Office of Applied Research on LLMs for Constraint Programming.

In addition to my research, I'm passionate about training the next generation of data scientists through programs such as iXperience where I've taught introductory practical data science tools in Python and R to students around the world. Outside of the office, you're most likely to find me hiking (27 national parks and counting).

Email  /  CV  /  Google Scholar

This paper considers the problem of performing k-means clustering while ensuring groups (e.g. demographic groups) have a minimum level of representation in a specified number of clusters. We show that the popular k-means algorithm (Lloyd's heuristic) can result in unfair outcomes where certain groups lack sufficient representation past the minimum threshold in a proportional number of clusters. We formulate the problem through a mixed-integer optimization framework and present a variant of Lloyd's algorithm, called MiniReL, that directly incorporates the fairness constraints.

Mathematical Programming has been a transformative technology for improving decision making in a number of industrial applications but its impact has been largely realized by large institutions that can afford the optimization expertise. We present a novel framework for interactive decision support for non-expert users that leverages large language models (LLM) to translate natural language requests into operations on an underlying constraint programming model. We investigate this framework through the lens of meeting scheduling, and showcase its potential via a user study with a prototype system.

This paper considers the learning of Boolean rules in disjunctive normal form (DNF, OR-of-ANDs, equivalent to decision rule sets) as an interpretable model for classification. An integer program is formulated to optimally trade classification accuracy for rule simplicity. We also consider the fairness setting and extend the formulation to include explicit constraints on two different measures of classification parity: equality of opportunity and equalized odds. Check out a short presenation on this paper at the Machine Learning NeEDS Mathematical Optimization Seminar.

We extend the idea of using polyhedra to define clusters to the cluster description setting where cluster assignments are fixed and the goal is to explain them as simply as possible. We model the problem as an exponential-sized IP which we use column generation to solve, and introduce a novel grouping scheme to help our approach tackle large datasets.

We propose a novel approach for interpretable clustering that both clusters data points and constructs polyhedra around the discovered clusters to explain them. Our framework allows for additional constraints on the polyhedra including ensuring that the hyperplanes constructing the polyhedra are axis-parallel or sparse with integer coefficients.

As part of the Amazon-MIT last mile routing challenge we presented a novel hierarchical approach to constructing vehicle routes based on historical driver preferences.

In this work we explore learning boolean rule sets with constraints on group fairness and rule set complexity. Our algorithm FairCG is able to achieve comparable performance to other interpretable models with stronger fairness guarantees.

As part of Cornell University's re-opening efforts during the COVID-19 pandemic the ORIE department led an effort to re-optmize Cornell's class schedule and room assignments for social distancing and a hybrid education model. I led the implementation of the primary optimization models for both class scheduling and room allocation.

Work done while at RBC Capital Markets. This patent outlines our team's approach for building a production grade deep reinforcmenet learning algorithm for trade execution (i.e. buying/selling large quantities of shares with low market impact). Check out our feature on Bloomberg !

I've been fortunate enough to spend the last few years teaching undergraduate students across 5 continents the fundamentals of data science in Python. Topics included pandas and numpy, machine learning with scikit-learn, deep learning in keras, and data visualization. As part of the program I also supervised 30+ remote data science projects in industries ranging from fintech to social media marketing.

This course offers a broad overview of computational techniques and mathematical skills useful for data scientists. Topics include: unix shell, regular expressions, version control: (git), data structures and algorithms, working with databases, data analysis using Python and related libraries (Pandas, NumPy/Scipy, scikit-learn), parallel computing (Map-Reduce, Spark, Hadoop), and an overview of standard machine learning and optimization algorithms.