Seattle Women in Tech Regatta 2020

This technical session brings together three perspectives and the application of machine learning techniques in very different situations. Attendees will have an opportunity to learn about reinforcement learning for personalization, Bayesian techniques applied the the search for missing children, and network graph modeling in pursuit of active users.

Creating Hooking Experiences Using Reinforcement Learning and Responsible AI
Samia Khalid
Netflix, Spotify, Instagram, Xbox – you name it – personalized experiences are what keep users hooked! How do the pros tackle this complex, dynamic problem? In this session, we will first deconstruct personalization and then we will learn how you can integrate it in your scenarios as well. Since with great power comes great responsibility, we will also talk about Responsible AI.

Bayesian Optimal Search Plan Construction to Assist the Search for Missing Child
Yuchao Ma 
More than 420,000 entries for missing children were reported to National Crime Information Center each year since 2014. Considering the vast uncertainties in the circumstances surrounding the missing child, scientific methods can be used to systematically evaluate potential scenarios, suggest searching directions and allocate time resources. In this pilot study, we introduce a Bayesian framework to construct an optimal search plan that maximizes the overall probability of detection, to assist the search for missing child. It first identifies a potential search area after information gathering, and then divides it into equal-sized units for prioritization based on a programmable granularity. A prior probability of detection is computed over all search units, and a detection model is developed in line with the optimal search theory. Along with the progress of search action, the posterior probability of finding the missing child is estimated using Bayesian method, and a maximum probability search plan is constructed, which assigns search effort to the unit with the highest posterior probability, and has been proved to be optimal for a given cumulative time investment. For the purpose of validation and visualization, we have developed a web-based application to render the proposed framework on an interactive GIS map. The implementation for algorithms and web-app will be shared for open access.