Latent-variable models for data-driven social good

Friday, June 14, 2019, 3:00 pm - 4:00 pm PDTiCal
This event is open to the public.
AI Seminar
Sabina Tomkins

In social good domains it is often important to understand why predictions are made, as well as ensuring the quality of those predictions. Latent-variable models offer one approach to improving transparency, while also improving predictive performance. Additionally, when there are dependencies between unknown variables, relational models which jointly infer their values can improve over off-the-shelf techniques which assume independence. I will present latent-variable models in three diverse problem settings: human trafficking, sustainable recommender systems and online education. In the area of human trafficking I will present results on the relationship between environmental events and human trafficking, as well as present a spatio-temporal model for route prediction. Next, I will show how latent representations of sustainability can allow us to discover potentially sustainable products. Finally, I will present a collective model of student learning. In each of these settings I will show how incorporating knowledge of structure between variables can improve performance, while latent-variable values can improve understanding. 


Sabina Tomkins is currently a postdoctoral scholar at Harvard University where she works on reinforcement learning for mobile health interventions. Previously, she obtained her PhD from the University of California Santa Cruz where she developed probabilistic models for data-driven social good. 

« Return to Upcoming Events