Research

My academic work, done for my Ph.D. in computer science at the University of Texas at Austin (2008, advised by Inderjit Dhillon and Emmett Witchel), was about learning from correlations: inferring how features relate to one another, measuring that structure, and scaling it with high performance numerical methods.

Metric learning

Algorithms like k-means and nearest-neighbor search are only as good as their notion of distance. Information-Theoretic Metric Learning (ICML 2007) learns a Mahalanobis distance from pairwise constraints by minimizing the LogDet divergence. It needs no eigenvalue computations or semidefinite programming, it can incorporate a prior, and it has an online variant with regret bounds. Follow-on work scaled the approach linearly with dimensionality (KDD 2008) and kernelized it to learn metrics over arbitrarily high-dimensional spaces (JMLR).

Clustering distributions

Sometimes each object is best described by a whole distribution rather than a point. Examples are a movie rated by many users, or a sensor emitting a time series. Differential Entropic Clustering (NIPS 2006) clusters multivariate Gaussians directly. It connects the problem to Bregman divergences and to the same LogDet geometry that underlies the metric learning work.

Statistical software analysis

The methods were validated on real software-engineering problems. Clarify (PLDI 2007) improves cryptic error reports by classifying a program's behavior profile. Cost-sensitive decision trees (ECML 2006) account for the overhead of instrumenting a program. The dissertation adds metric learning for nearest-neighbor software support, and clustering-based ordering to search large unit-test suites efficiently.

Web search

Localized search engines index one community of the web but lack the global PageRank needed to judge a page's importance. Estimating the Global PageRank of Web Communities (KDD 2006) estimates it using only O(n) computation, bandwidth, and storage for a local domain of n pages.

Selected publications

Metric and Kernel Learning Using a Linear Transformation
P. Jain, B. Kulis, J. V. Davis, I. S. Dhillon

Journal of Machine Learning Research (JMLR), 2012

arXiv JMLR
Mining Statistical Correlations with Applications to Software Analysis
J. V. Davis

Ph.D. Dissertation, The University of Texas at Austin, 2008

PDF UT Austin
Structured Metric Learning for High-Dimensional Problems
J. V. Davis, I. S. Dhillon

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2008

PDF ACM
Information-Theoretic Metric Learning
J. V. Davis, B. Kulis, P. Jain, S. Sra, I. S. Dhillon

International Conference on Machine Learning (ICML), 2007 2,000+ citations

PDF ACM
Improved Error Reporting for Software that Uses Black-Box Components
J. Ha, C. J. Rossbach, J. V. Davis, I. Roy, H. E. Ramadan, D. E. Porter, D. L. Chen, E. Witchel

ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), 2007

ACM
Differential Entropic Clustering of Multivariate Gaussians
J. V. Davis, I. S. Dhillon

Advances in Neural Information Processing Systems (NIPS), 2006

PDF NeurIPS
Cost-Sensitive Decision Tree Learning for Forensic Classification
J. V. Davis, J. Ha, C. J. Rossbach, H. E. Ramadan, E. Witchel

European Conference on Machine Learning (ECML), 2006

Springer
Estimating the Global PageRank of Web Communities
J. V. Davis, I. S. Dhillon

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2006

PDF ACM