Matthew Howard

Prototyped research models for resolving unique users across devices from anonymous and semi-anonymous web logs using the Java-based probabilistic soft logic (PSL) framework for probabilistic relational learning.

Aggregated web log data (e.g., geographic clustering, device similarities) to engineer features to capture transitive relationships of users across multiple modes of access

Trained and tested large-scale probabilistic models at scale with AWS.

Developed novel probabilistic model to predict future mobile device usage activity based upon user context features (e.g., phone state, recent device actions) and historical application usage trends with Java-based probabilistic soft logic (PSL) framework.

Model successfully predicted future action in 75% of cases within 1 time step vs. 68% for baseline SVM-HMM model for 103 user sample

Wrote SQL/Python pipeline to digest terabyte-scale DeviceAnaylzer mobile device dataset into model-ready feature databases.

Automated device usage visualizations across timescales with Matplotlib to identify contextual dependence of features and user actions.

Research conducted as Ph.D. student in the area of statistical relational learning and AI.

Developed probabilistic approaches to ranking/recommendation problems within the setting of relational networks (e.g., social nets).

Designed method to cast non-convex optimization of the ranking metric NDCG as a convex optimization for utilization in learning of rankings via graphical models known as Hinge-loss Markov random fields (HL-MRFs).

Created early-stage probabilistic implementation of the PC-algorithm for causal discovery within probabilistic soft logic (PSL) to identify causal structure within networked data.