Student
Researcher
Developer
Hello, I'm Jenny. Welcome to my website!
I am an incoming PhD student in the
Technology & Operations Management program at Harvard Business School, supported by an
NSF Graduate Research Fellowship.
Currently, I'm a
Pre-Doctoral Researcher in the
Computational Social Science Lab at Microsoft Research working with
David Rothschild,
Jake Hofman, and
Dan Goldstein.
Before this, I was an undergraduate student at
Wellesley College, where I graduated with majors in Computer Science and Economics.
I am interested in improving existing methods used to answer social and policy-relevant questions. My recent projects focus on
understanding how LLMs change the way humans interact with technology. Specifically, I have explored how GPT-based models
can be prompt-engineered for synthetic data generation and how BERT models fine-tuned on biography text can mitigate bias in race
classification models.
I discovered my interest in research after several years of building software. I am still eager to learn new skills and technologies, and I hope to design tools that can benefit the research community.
I worked on the Media Bias Detector while at Microsoft Research, in collaboration with the Penn Media Accountability Project. We have been working hard to track and classify (using large language models) the top stories published by a collection of prominent publishers spanning the political spectrum in close to real time. Please feel free to visit the site and explore the data! Any feedback is appreciated.
This paper was a follow-up to my undergraduate thesis. I worked with Wellesley professors, Kyung Park and Eni Mustafaraj, to write a succinct analysis of how name and biography-based race classification models perform when tested on several thousand cast and crew members from IMDb.
It is currently under review. Please do not distribute this file!
This thesis explores the use of web scraping and machine learning to enhance the predictive power of current race classification models. With the rise of online data as an important source for research, the availability of race labels typically acquired through traditional survey methods is increasingly limited. Instead, we have access to many features that may relate to race, but not the explicit race label itself. Social scientists have turned to name-based racial classifiers as a predictive source for a person's racial identity. However, relying solely on name-based racial classifiers may be an unreliable approach to inferring racial identity due to potential measurement errors. Therefore, I explore and fine-tune alternative methods for racial classification, focusing on three types of data sources: names, images, and unstructured text.
My new methods are relevant given the growing and influential literature that uses name-based classification to study racial cleavages in America, particularly in the social sciences. While my conceptual framework is motivated by applications in economics, the research methodology and analysis rely on computer science techniques. I evaluate the performance of existing name and image-based race classifiers using a well-balanced dataset consisting of 5,201 individuals. Additionally, I experiment with several machine learning models, such as support vector machines, logistic regression, and transformers, by leveraging biographical information from IMDb, an online repository of movies, TV series, and cast and crew members. Furthermore, I compare the strengths of linear and stacked generalization ensemble methods in combining predictions from name, image, and unstructured text-based racial classifiers. Finally, I construct a linear multi-modal ensemble race classifier that leverages the strengths of several models to enhance predictive accuracy. As an illustration of its benefits, I outline the usage of web scraping and my multi-modal classifier in a case study centered on examining the representation of race on television.
This group project was written for Wellesley College's ECON 320 Economic Development course. For the course final project, we were tasked with proposing a research project that would be theoretically feasible. We used various techniques discussed in class to come up with a research design.
In November 2020, I participated in the Citi x Wellesley Case Competition hosted by Wellesley in Business and Citi Ventures. My team, comprised of four Wellesley students in the Class of 2023, came in first place in the Citi Markets and Securities Services challenge. This was an incredible opportunity to grow and learn, while solving real-world challenges at #Citi.
As a member of a community, whether it’s a company, school, or local restaurant, you have a voice. Devices like TP-Link or Google Nest give you the autonomy to control your home temperature from your phone, but ambient allows entire communities to provide feedback on their temperature preferences.
Ambient gives all employees and customers the right to work and spend quality time in a comfortable atmosphere. With respect to all the requests submitted, ambient will analyze and send the data to the thermostat administrator. The administrator does not have to worry what temperature to set to satisfy everyone; Ambient will do it for them. Join Ambient in helping us all ‘love the air around us’.
Technologies Used: Front-End: Figma, Flask, Bootstrap, HTML/CSS, Back-End: AWS DynamoDB, SQLite, Other: Wolfram API, Google Cloud Compute Engine Main Language: Python BootStrap,html,css, Amazon Web service dynamo db for user database and sqlite for management database, Google Cloud Compute Engine to create VM instances and run the entire Flask app in the cloud.
This project came to life as an answer to the rhetorical question: why is it so cold in here? I worked in a large office building last year, and always came to work with a jacket. It would be 90 degrees in the summer, yet the AC was cranked down to what felt like subzero temperatures. I wasn’t the only one who felt this way. From conversations with friends, I learned that many others felt the same way: their offices/schools/place-of-work was too cold in the summer, and too hot in the winter.
During TechTogether, our group brainstormed many approaches to solving this problem. We chose to implement it as an app that takes in mass data and calculates the optimal temperature change based on these individual preferences (which we quantified). It has three major components: (1) front end mobile app, designed in Figma, that shows what users see, and a Flask web app for greater control/features. (2) database that collects data from each user tap in the app, which would mean connecting the front end to the AmazonDB database. (3) backend that uses the data from the database to calculate the suggested ambient temperature increase/decrease using the Wolfram API in Python. Finally, we learned how to use the Google Cloud Compute Engine to create VM instances and run our entire Flask app in the cloud. We were able to complete part of each component, but we didn’t have enough time to learn how to connect each piece fully together. However, we learned a lot along the way ☺.
