Evan R. Kirshenbaum

Extremely fast text feature extraction for classification and indexing

Most research in speeding up text mining involves algorithmic improvements to induction algorithms, and yet for many large scale applications, such as classifying or indexing large document repositories, the time spent extracting word features from texts can itself greatly exceed the initial training time. This paper describes a fast method for text feature extraction that folds together Unicode conversion, forced lowercasing, word boundary detection, and string hash computation. We show empirically that our integer hash features result in classifiers with equivalent statistical performance to those built using string word features, but require far less computation and less memory.

Pragmatic text mining: minimizing human effort to quantify many issues in call logs

We discuss our experiences in analyzing customer-support issues from the unstructured free-text fields of technical-support call logs. The identification of frequent issues and their accurate quantification is essential in order to track aggregate costs broken down by issue type, to appropriately target engineering resources, and to provide the best diagnosis, support and documentation for most common issues. We present a new set of techniques for doing this efficiently on an industrial scale, without requiring manual coding of calls in the call center. Our approach involves (1) a new text clustering method to identify common and emerging issues; (2) a method to rapidly train large numbers of categorizers in a practical, interactive manner; and (3) a method to accurately quantify categories, even in the face of inaccurate classifications and training sets that necessarily cannot match the class distribution of each new months data. We present our methodology and a tool we developed and deployed that uses these methods for tracking ongoing support issues and discovering emerging issues at HP.

A live comparison of methods for personalized article recommendation at forbes.com

We present the results of a multi-phase study to optimize strategies for generating personalized article recommendations at the Forbes.com web site. In the first phase we compared the performance of a variety of recommendation methods on historical data. In the second phase we deployed a live system at Forbes.com for five months on a sample of 82,000 users, each randomly assigned to one of 20 methods. We analyze the live results both in terms of click-through rate (CTR) and user session lengths. The method with the best CTR was a hybrid of collaborative-filtering and a content-based method that leverages Wikipedia-based concept features, post-processed by a novel Bayesian remapping technique that we introduce. It both statistically significantly beat decayed popularity and increased CTR by 37%.

Using Genetic Programming to Obtain a Closed-Form Approximation to a Recursive Function

We demonstrate a fully automated method for obtaining a closedform approximation of a recursive function. This method resulted from a realworld problem in which we had a detector that monitors a time series and where we needed an indication of the total number of false positives expected over a fixed amount of time. The problem, because of the constraints on the available measurements on the detector, was formulated as a recursion, and conventional methods for solving the recursion failed to yield a closed form or a closed-form approximation. We demonstrate the use of genetic programming to rapidly obtain a high-accuracy approximation with minimal assumptions about the expected solution and without a need to specify problem-specific parameterizations. We analyze both the solution and the evolutionary process. This novel application shows a promising way of using genetic programming to solve recurrences in practical settings.