BRC FY17 Topic 6:Predictive and causal modeling with causal graphical models
Abstract
Significant progress has been made in the machine learning (ML) field in developing predictive models. For example, deep learning (DL) is an emerging area of ML in the last five to ten years shown to produce state-of-the-art results in large-scale industrial applications in computer vision, automatic speech recognition, and natural language processing. Unfortunately, most of the ML models including DL are black-box approaches that have little explanatory power. The prediction power of ML does not translate into insights about the underlying data-generating mechanisms. In reality many pressing questions we face are often causal in nature. Understanding the causal mechanisms underlying an observed phenomenon is one of the primary goals of science and engineering. There exists a mismatch, therefore, between the aspirations of the scientists and the current ML tools for data analysis.On the other hand, recent progress in the theory of causal graphical models (CGMs) has provided a generalmethodology for representing and reasoning about causal relationships. However, CGMs have not achieved great impact as ML models in real-world applications. The main objective of this project is to develop a principled theory that is both predictive and explanatory, based on the framework of CGMs.We will develop the theoretical foundations and algorithms necessary to equip scientists with effective tools to pursue sound predictive and causal modeling from big data. We aim to bridge the methodology gap between ML and causal modeling and to investigate in what ways the advantages of one can be made to mitigate the shortcomings of the other. The central research challenge we will address is to develop efficient methods for learning CGMs from highdimensional data. We propose to attack the problem using some of the successful ML techniques developed over the years, including curriculum learning, sparse learning, feature selection, and parallel computations. We aim to develop strategies to exploit unique opportunities that appear only in a big data environment to develop algorithms to learn CGMs from big heterogeneous data from interventions, natural experiments, or different resources. We propose tocollect or generate benchmark datasets to conduct extensive experimental studies to evaluate the predictionperformance of CGMs against traditional ML algorithms and their ability to accurately predict causal effectsrelationships.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712140
Entities
People
- Jin Tian
Organizations
- Iowa State University
- Office of Naval Research
- United States Navy