`2018-01-30 17:00:00``2018-01-30 18:00:00``Topology, Geometry and Data Seminar - Rob Littleton``Title: Optimization Considerations for Gradient Boosted ClassificationSpeaker: Rob Littleton (CoverMyMeds)Abstract: The mission of CoverMyMeds (CMM) is to help patients get the medicines they need to be healthy, which is achieved mainly by making the prior-authorization (PA) process more efficient. This discussion will focus on the optimization of a product called “Indicators.” The purpose of Indicators is to predict, at the point of prescription, whether a PA will be required for an insurance company to pay for a drug. In order to make this prediction, CMM leverages visibility into data streams between pharmacies and insurance companies, which provides insight into which claims are rejected by the insurance company, and which are paid.The data set is large and contains mostly categorical variables. As the decision criteria is binary, a tree-based classifier is an obvious choice for building the predictive model. Therefore, gradient boosted machine (GBM) is ideal due to its ability to handle large amounts of categorical data and its predictive capabilities. While GBM is itself a powerful classifier, optimization of the model requires tuning a combination of interconnected hyperparameters.Running the entire GBM algorithm is a computationally intensive process. Running multiple iterations to determine optimal hyperparameters therefore must utilize a simple, fast, multi-objective tuning methodology. We are presently testing multiple optimization modalities including Particle Swarm Optimization (PSO) variants and Bayesian optimization.This presentation will cover the overall problem, the optimization algorithms we’re using, some of the results achieved, and future directions.Seminar URL: https://research.math.osu.edu/tgda/tgda-seminar.html``Cockins Hall 240``OSU ASC Drupal 8``ascwebservices@osu.edu``America/New_York``public`

`2018-01-30 16:00:00``2018-01-30 17:00:00``Topology, Geometry and Data Seminar - Rob Littleton``Title: Optimization Considerations for Gradient Boosted ClassificationSpeaker: Rob Littleton (CoverMyMeds)Abstract: The mission of CoverMyMeds (CMM) is to help patients get the medicines they need to be healthy, which is achieved mainly by making the prior-authorization (PA) process more efficient. This discussion will focus on the optimization of a product called “Indicators.” The purpose of Indicators is to predict, at the point of prescription, whether a PA will be required for an insurance company to pay for a drug. In order to make this prediction, CMM leverages visibility into data streams between pharmacies and insurance companies, which provides insight into which claims are rejected by the insurance company, and which are paid.The data set is large and contains mostly categorical variables. As the decision criteria is binary, a tree-based classifier is an obvious choice for building the predictive model. Therefore, gradient boosted machine (GBM) is ideal due to its ability to handle large amounts of categorical data and its predictive capabilities. While GBM is itself a powerful classifier, optimization of the model requires tuning a combination of interconnected hyperparameters.Running the entire GBM algorithm is a computationally intensive process. Running multiple iterations to determine optimal hyperparameters therefore must utilize a simple, fast, multi-objective tuning methodology. We are presently testing multiple optimization modalities including Particle Swarm Optimization (PSO) variants and Bayesian optimization.This presentation will cover the overall problem, the optimization algorithms we’re using, some of the results achieved, and future directions.Seminar URL: https://research.math.osu.edu/tgda/tgda-seminar.html``Cockins Hall 240``Department of Mathematics``math@osu.edu``America/New_York``public`**Title**: Optimization Considerations for Gradient Boosted Classification

**Speaker**: Rob Littleton (CoverMyMeds)

**Abstract**: The mission of CoverMyMeds (CMM) is to help patients get the medicines they need to be healthy, which is achieved mainly by making the prior-authorization (PA) process more efficient. This discussion will focus on the optimization of a product called “Indicators.” The purpose of Indicators is to predict, at the point of prescription, whether a PA will be required for an insurance company to pay for a drug. In order to make this prediction, CMM leverages visibility into data streams between pharmacies and insurance companies, which provides insight into which claims are rejected by the insurance company, and which are paid.

The data set is large and contains mostly categorical variables. As the decision criteria is binary, a tree-based classifier is an obvious choice for building the predictive model. Therefore, gradient boosted machine (GBM) is ideal due to its ability to handle large amounts of categorical data and its predictive capabilities. While GBM is itself a powerful classifier, optimization of the model requires tuning a combination of interconnected hyperparameters.

Running the entire GBM algorithm is a computationally intensive process. Running multiple iterations to determine optimal hyperparameters therefore must utilize a simple, fast, multi-objective tuning methodology. We are presently testing multiple optimization modalities including Particle Swarm Optimization (PSO) variants and Bayesian optimization.

This presentation will cover the overall problem, the optimization algorithms we’re using, some of the results achieved, and future directions.

**Seminar URL**: https://research.math.osu.edu/tgda/tgda-seminar.html