I am a data science team lead for an innovation research team at Comscore. My passion is researching, building, and packaging up new custom probabilistic and machine learning models to improve predictions, decision making, and inferences.
One recent example is using a custom low-dimensional Gaussian process kernel to vary the certainty of one set of predictions from one dataset based on how the two datasets align on a separate predictive metric; this allows for probabilistic data fusion and variable information sharing from two distinct data sources.
Another recent example is creating a torch module for hierarchical embeddings; this in turn greatly improves embedding performance for ‘hierarchically ordered’ entities, such as embedding websites within domains within a network within a company. This also permits embeddings to exist at multiple levels of specificity, and to allow efficient information sharing ala Bayesian multilevel models.