Using Stochastic Modeling to Determine Optimal Resource Allocation in Classroom Teaching
Master's Thesis. Publication pending.
- Decreasing class sizes and increasing formative assessments improve student learning, but with diminishing returns.
- Neither hiring more teachers nor increasing formative assessments is always the optimal strategy for improving student learning on a per-dollar level.
- Stochastic modelling can effectively replicate classroom dynamics, enabling educational psychology research without incurring significant monetary or temporal costs and without violating ethics considerations with human subjects.
Mentored by Michael Frank
- Paper: Publication pending
What Will You Code Next? Deep knowledge tracing on non-binary data
Publication accepted to the Journal of Educational Data Mining, publication pending. Presenter at Women in Machine Learning Workshop 2016, Learning at Scale 2017, and Educational Data Mining 2017 Conference.
- Recurrent neural nets (RNNs) with long short-term memory (LSTMs) can predict student performance on coding exercises better than various naive classifiers.
- RNNs can use incremental code submissions to predict future student performance and anticipate student struggles better than logistic regression or Bayesian knowledge tracing methods.
- Complex inputs and outputs can be represented as vector embeddings generated by recursive neural nets without substantial accuracy dropoff, which enables future work on automating feedback for student learning on complex tasks.
Mentored by Chris Piech
Mirror, mirror on the wall: How perception of personal attractiveness affects negotiation performance
Organizational Behavior and Human Decision Processes, Vol. 124, Issue 2, Jul. 2014, p.133-149
- Higher self-perceived attractiveness (SPA) increased support for inequality.
- By contrast, lower SPA decreased support for inequality and hierarchies.
- These effects occurred because SPA influenced perceived social class.
- Power, status, and self-esteem did not account for the effects.
- Correlational and experimental evidence are presented.
Mentored by Peter Belmi and Margaret Neale
- Journal link: http://www.sciencedirect.com/science/article/pii/S0749597814000223
- Stanford GSB article: Researchers: A Few Bad Hair Days Can Change Your Life
- Poster: pdf, jpg
Rethinking natural altruism: Simple reciprocal interactions trigger children's benevolence
Proceedings of the National Academy of Sciences, Vol. 111, Issue 48, Dec. 2014, p.17071-17074
- Reciprocal interactions guide children’s acquisition of cultural knowledge.
- Reciprocal interactions guide children's efforts to transfer information that they have learned.
- Reciprocity may be an implicit pedagogical cue (Gergely & Csibra, 2009) that allows children to become culturally competent.
- Non-reciprocal activity does not trigger altruism and socialization.
Mentored by Rodolfo Cortes and Carol Dweck
Find Your Food Soulmate: Community detection on food preference models using Yelp dataset
- On average, 9.2 people have a particular distinct set of food preferences, each represented by a community/cluster.
- On average, only 2.43 of your friends share the same food preferences. This implies there are ~6 potential individuals we could recommend who would share your food tastes that you don't already know.
- We have found a way to predict with 47% accuracy (far better than random chance, given there are 533 communities) whether or not two people have food tastes similar enough that they rate a given restaurant within 1 star of each other.
This research project was written in Python, and leveraged the Stanford Network Analysis Project's Python package, Snap.py, for network functions. This project was made with Angela Sy and Blanca Villanueva for CS 224w: Social and Information Network Analysis.