This Spring 2025 semester, students came up with a number of exciting projects in the area of human and data-centred AI. This newsletter provides some project highlights! (Keeping with the themes of the course -- in which we heavily discussed data governance, markets, sharing preferences, and provenance -- these are from students who opted in).
I sharing this, I'm hoping to celebrate student work, but also highlight some skilled students in case other SFU colleagues are looking for RAs or similar opportunities.
| "Linguistic Bias in Multilingual AI: A Study of ChatGPT's Language Variations" |
| - A combined benchmarking and system-building effort! - Team: Justine Gerrard |
In this project, the team built a browser extension for quickly asking LLMs question across six different language (first, the question is translated using a dedicated translation API, and then is sent as a query to an LLM, and that answer is translated back to English for comparison). The team also built a benchmark dataset to compare how answers vary when asked across these languages   |
"Fourth Down: Human-Centered Analytics in the NFL"
- Applying domain knowledge to the context of ML for sports!
- Team: Paul Atwal
In this project, the team set out train a model to predict fourth down outcomes. The project report includes a very nice intro to the concept of fourth downs for a sports-averse reader. The final approach ended up involving an ensemble of four sub-models (win probability, punt outcomes, field goal outcomes, and "should you go for it") that provide some nice natural explainability benefits, and some Shapley value-based feature analysis. This project also involved substantial domain knowledge to process and interpret the data. 
"Effects of Privacy-Driven Data Leverage on Recommender System Performance"
- What happens to recommender systems performance if users take privacy preserving actions?
- Team: Nida Anwar, Efe Erhan, Katya Kubyshkin, Zoe Stanley
Here, the team replicated results from a research paper in recommender systems that combined demographic-based user clustering with traditional rating-based collaborative filtering (using MovieLens data) and then conducted experiments to simulate how privacy interventions (removing ratings, removing demographic data) might affect model performance. 
A Data Scaling Law of a Manifold’s Resolution
- Working through the concept of "data scaling" mathematically!
- Team: Abiel Kim
The team produced a mathematical analysis of data scaling through the lens of the manifold hypothesis, and is working to extend this analysis with additional proofs and exploration. 
Patent Compass Navigate the wold of patents with AI-powered precision
- Team: Daniel Surina
In this project, the team built a semantic search tool that can identify patents similar to a user's query. This involved data collection, pre-processing, an embedding model, and a search API. 
"Learning to Phish: Evaluating Machine Learning Models Across Diverse Email Datasets" Classic phishing detection across a variety of datasets!
- Team: Nathan Chan
Here, the team took a classic problem in ML -- phishing detection -- and assess the impact of different data distributions on performance through some clever approaches to (1) generating synthetic data and (2) acquiring new human generated data. The analysis involved a comparison across domains and Shapley value-based investigation of token-level influence.
