Jason V. Davis

Information-theoretic metric learning

In this paper, we present an information-theoretic approach to learning a Mahalanobis distance function. We formulate the problem as that of minimizing the differential relative entropy between two multivariate Gaussians under constraints on the distance function. We express this problem as a particular Bregman optimization problem---that of minimizing the LogDet divergence subject to linear constraints. Our resulting algorithm has several advantages over existing methods. First, our method can handle a wide variety of constraints and can optionally incorporate a prior on the distance function. Second, it is fast and scalable. Unlike most existing methods, no eigenvalue computations or semi-definite programming are required. We also present an online version and derive regret bounds for the resulting algorithm. Finally, we evaluate our method on a recent error reporting system for software called Clarify, in the context of metric learning for nearest neighbor classification, as well as on standard data sets.

Structured metric learning for high dimensional problems

The success of popular algorithms such as k-means clustering or nearest neighbor searches depend on the assumption that the underlying distance functions reflect domain-specific notions of similarity for the problem at hand. The distance metric learning problem seeks to optimize a distance function subject to constraints that arise from fully-supervised or semisupervised information. Several recent algorithms have been proposed to learn such distance functions in low dimensional settings. One major shortcoming of these methods is their failure to scale to high dimensional problems that are becoming increasingly ubiquitous in modern data mining applications. In this paper, we present metric learning algorithms that scale linearly with dimensionality, permitting efficient optimization, storage, and evaluation of the learned metric. This is achieved through our main technical contribution which provides a framework based on the log-determinant matrix divergence which enables efficient optimization of structured, low-parameter Mahalanobis distances. Experimentally, we evaluate our methods across a variety of high dimensional domains, including text, statistical software analysis, and collaborative filtering, showing that our methods scale to data sets with tens of thousands or more features. We show that our learned metric can achieve excellent quality with respect to various criteria. For example, in the context of metric learning for nearest neighbor classification, we show that our methods achieve 24% higher accuracy over the baseline distance. Additionally, our methods yield very good precision while providing recall measures up to 20% higher than other baseline methods such as latent semantic analysis.

Improved error reporting for software that uses black-box components

An error occurs when software cannot complete a requested action as a result of some problem with its input, configuration, or environment. A high-quality error report allows a user to understand and correct the problem. Unfortunately, the quality of error reports has been decreasing as software becomes more complex and layered. End-users take the cryptic error messages given to them by programsand struggle to fix their problems using search engines and support websites. Developers cannot improve their error messages when they receive an ambiguous or otherwise insufficient error indicator from a black-box software component. We introduce Clarify, a system that improves error reporting by classifying application behavior. Clarify uses minimally invasive monitoring to generate a behavior profile, which is a summary of the programs execution history. A machine learning classifier uses the behavior profile to classify the applications behavior, thereby enabling a more precise error report than the output of the application itself. We evaluate a prototype Clarify system on ambiguous error messages generated by large, modern applications like gcc, La-TeX, and the Linux kernel. For a performance cost of less than 1% on user applications and 4.7% on the Linux kernel, the proto type correctly disambiguates at least 85% of application behaviors that result in ambiguous error reports. This accuracy does not degrade significantly with more behaviors: a Clarify classifier for 81 La-TeX error messages is at most 2.5% less accurate than a classifier for 27 LaTeX error messages. Finally, we show that without any human effort to build a classifier, Clarify can provide nearest-neighbor software support, where users who experience a problem are told about 5 other users who might have had the same problem. On average 2.3 of the 5 users that Clarify identifies have experienced the same problem.

Estimating the global pagerank of web communities

Localized search engines are small-scale systems that index a particular community on the web. They offer several benefits over their large-scale counterparts in that they are relatively inexpensive to build, and can provide more precise and complete search capability over their relevant domains. One disadvantage such systems have over large-scale search engines is the lack of global PageRank values. Such information is needed to assess the value of pages in the localized search domain within the context of the web as a whole. In this paper, we present well-motivated algorithms to estimate the global PageRank values of a local domain. The algorithms are all highly scalable in that, given a local domain of size n, they use O(n) resources that include computation time, bandwidth, and storage. We test our methods across a variety of localized domains, including site-specific domains and topic-specific domains. We demonstrate that by crawling as few as n or 2n additional pages, our methods can give excellent global PageRank estimates.

Cost-Sensitive decision tree learning for forensic classification

In some learning settings, the cost of acquiring features for classification must be paid up front, before the classifier is evaluated. In this paper, we introduce the forensic classification problem and present a new algorithm for building decision trees that maximizes classification accuracy while minimizing total feature costs. By expressing the ID3 decision tree algorithm in an information theoretic context, we derive our algorithm from a well-formulated problem objective. We evaluate our algorithm across several datasets and show that, for a given level of accuracy, our algorithm builds cheaper trees than existing methods. Finally, we apply our algorithm to a real-world system, Clarify. Clarify classifies unknown or unexpected program errors by collecting statistics during program runtime which are then used for decision tree classification after an error has occurred. We demonstrate that if the classifier used by the Clarify system is trained with our algorithm, the computational overhead (equivalently, total feature costs) can decrease by many orders of magnitude with only a slight (<1%) reduction in classification accuracy.

Prediction of Protein–Protein Interactions: A Study of the Co-evolution Model

The concept of molecular co-evolution drew attention in recent years as the basis for several algorithms for the prediction of protein-protein interactions. While being successful on specific data, the concept has never been tested on a large set of proteins. In this chapter we analyze the feasibility of the co-evolution principle for protein-protein interaction prediction through one of its derivatives, the correlated divergence model. Given two proteins, the model compares the patterns of divergence of their families and assigns a score based on the correlation between the two. The working hypothesis of the model postulates that the stronger the correlation the more likely is that the two proteins interact. Several novel variants of this model are considered, including algorithms that attempt to identify the subset of the database proteins (the homologs of the query proteins) that are more likely to interact. We test the models over a large set of protein interactions extracted from several sources, including BIND, DIP, and HPRD.

CO 2 Capture by Absorption with Potassium Carbonate First Quarterly Report 2007

The objective of this work is to improve the process for CO{sub 2} capture by alkanolamine absorption/stripping by developing an alternative solvent, aqueous K{sub 2}CO{sub 3} promoted by piperazine. Modeling of stripper performance suggests that vacuum stripping may be an attractive configuration for all solvents. Flexipac 1Y structured packing performs in the absorber as expected. It provides twice as much mass transfer area as IMTP No.40 dumped packing. Independent measurements of CO{sub 2} solubility give a CO{sub 2} loading that is 20% lower than that Cullinanes values with 3.6 m PZ at 100-120 C. The effective mass transfer coefficient (K{sub G}) in the absorber with 5 m K/2.5 m PZ appears to be 0 to 30% greater than that of 30 wt% MEA.