CAREER: Data to Operational Decisions: A Predictive Analytics Approach

New York University, New York NY

Investigators

Abstract

This Faculty Early Career Development (CAREER) Program grant will develop data-driven modeling and learning techniques to improve the accuracy of decision making in operations. The focus will be on easy to implement techniques that work "out-of-the-box" for a wide-range of decisions faced by businesses and society: the right products to design, the right products and prices to offer to customers, the right quantity of each product to carry, and even the right policy mix (taxes, subsidies, etc.) to offer to individual industries or demographic groups. The existing approach to this problem is to select an appropriate model, learn model parameters from data, and then solve the decision problem under the model. The focus is mainly on either learning the model (typically, independent of the specific decision context) or efficiently solving the decision problem given the model, leaving model selection to an expert. In contrast, this research will take an end-to-end approach: starting with a type of data (purchase transactions, click-streams, marketing studies, choice of insurance policies, etc.) and ending with an operational decision. The integrated approach will work "out-of-the-box" by automatically selecting a model customized to the data and the decision. By limiting the need for expert input, the research will significantly increase access to data-driven decision methodologies to a wide-range of industries - greatly benefiting the U.S. economy and society. For the end-to-end approach, the research will develop modeling techniques that not only have high predictive power, but are amenable to efficient optimization (solving of the decision problem). The challenge however is that complex predictive models lack structure for efficient optimization, whereas parsimonious models that can be efficiently optimized lack predictive power. In order to balance this tension, the research will focus on the model class of distributions over preference (ranked) lists. This model class is very general (is rooted in the classical economic utility theory) and captures a wide range of responses of individuals to product or policy offerings. Despite its generality, the model class also possesses rich transitivity structure. The research will develop algorithmic solutions that use data to identify an instance of the general model class and then exploit the transitivity structure to efficiently optimize. Further, the research will develop novel analysis techniques to quantify the performance of the algorithms. Performance analysis is challenging because it requires quantifying "model complexity" and "data complexity". The research will develop novel "complexity" metrics using traditional engineering techniques such as compressive sensing and Boolean function analysis.