Fateh A. Tipu

Segmentation-based modeling for advanced targeted marketing

Fingerhut Business Intelligence (BI) has a long and successful history of building statistical models to predict consumer behavior. The models constructed are typically segmentation-based models in which the target audience is split into subpopulations (i.e., customer segments) and individually tailored statistical models are then developed for each segment. Such models are commonly employed in the direct-mail industry; however, segmentation is often performed on an ad-hoc basis without directly considering how segmentation affects the accuracy of the resulting segment models. Fingerhut BI approached IBM Research with the problem of how to build segmentation-based models more effectively so as to maximize predictive accuracy. The IBM Advanced Targeted Marketing-Single EventsTM (IBM ATM-SETM) solution is the result of IBM Research and Fingerhut BI directing their efforts jointly towards solving this problem. This paper presents an evaluation of ATM-SEs modeling capabilities using data from Fingerhuts catalog mailings.

Data-intensive analytics for predictive modeling

The Data Abstraction Research Group was formed in the early 1990s, to bring focus to the work of the Mathematical Sciences Department in the emerging area of knowledge discovery and data mining (KD & DM). Most activities in this group have been performed in the technical area of predictive modeling, roughly at the intersection of machine learning, statistical modeling, and database technology. There has been a major emphasis on using business and industrial problems to motivate the research agenda. Major accomplishments include advances in methods for feature analysis, rule-based pattern discovery, and probabilistic modeling, and novel solutions for insurance risk management, targeted marketing, and text mining. This paper presents an overview of the groups major technical accomplishments.

Probabilistic estimation-based data mining for discovering insurance risks

IBMs underwriting profitability analysis application mines property and casualty insurance policy and claims data to construct predictive models for insurance risks. UPA uses the ProbE data-mining kernel to discover risk-characterization rules by analyzing large, noisy data sets.

A probabilistic estimation framework for predictive modeling analytics

IBM ProbE (for probabilistic estimation) is an extensible, embeddable, and scalable modeling engine, particularly well-suited for implementing segmentation-based modeling techniques, wherein data records are partitioned into segments and separate predictive models are developed for each segment. We describe the ProbE framework and discuss two key business solutions that have been built using ProbE: the IBM Underwriting Profitability Analysis for insurance risk management, and the IBM Advanced Targeted Marketing for Single Events for direct mail database marketing.

Embedded predictive modeling in a parallel relational database

A methodology for embedding predictive modeling algorithms in a commercial parallel database is described; specifically, the parallel editions of IBM DB2 Universal Database, although many aspects of the overall approach can be used with other commercial parallel databases. This parallelization approach was implemented in the Version 8.2 release of DB2 Intelligent Miner Modeling to support a new predictive modeling algorithm called Transform Regression. This database-embedded mining algorithm provides all the usual benefits, including easier integration into large enterprise applications, the ability to perform entire data mining workflows directly from an SQL-based programming interface, reduced data transfer costs between the database and the data mining application, and faster, parallel data access during query processing. However, in addition to the these benefits, a significant part of the data mining computations are also parallelized without the use of any sophisticated parallel programming constructs, or any specialized message passing and parallel synchronization libraries.

On optimizing the selection of business transformation projects

To compete and thrive in a changing business environment, a business can adapt by initiating and successfully carrying out business transformation projects. In this paper we propose a methodology for the optimal selection of such transformational projects. We propose a two-stage methodology based on (1) correlation analytics for identifying key drivers of business performance and (2) advanced portfolio-optimization techniques for selecting optimal business-transformation portfolios in the face of resource constraints, budget constraints, and a rich variety of business rules. We illustrate our methodology through a case study from the electronics industry.

A Unified Framework for Outlier Detection in Trace Data Analysis

Process trace data (PTD) is an important data type in semiconductor manufacturing and has a very large aggregate volume. While data mining and statistical analysis play a key role in the quality control of wafers, the existence of outliers adversely affects the applications benefiting from PTD analysis. Due to the complexities of PTD and the resultant outlier patterns, this paper proposes a unified outlier detection framework which takes advantages of data complexity reduction using entropy and abrupt change detection using cumulative sum (CUSUM) method. To meet the practical needs of PTD analysis, a two-step algorithm taking into account of the related domain knowledge is developed, and its effectiveness is validated by using real PTD sets and a production example. The experimental results show that the proposed method outperforms the Fast Greedy Algorithm (FGA) and the Grubbs test, two commonly used outlier detection techniques for univariate data.