Varghese S. Jacob

A Study of the Classification Capabilities of Neural Networks Using Unsupervised Learning: A Comparison with K-Means Clustering

Several neural networks have been proposed in the general literature for pattern recognition and clustering, but little empirical comparison with traditional methods has been done. The results reported here compare neural networks using Kohonen learning with a traditional clustering method (K-means) in an experimental design using simulated data with known cluster solutions. Two types of neural networks were examined, both of which used unsupervised learning to perform the clustering. One used Kohonen learning with a conscience and the other used Kohonen learning without a conscience mechanism. The performance of these nets was examined with respect to changes in the number of attributes, the number of clusters, and the amount of error in the data. Generally, theK-means procedure had fewer points misclassified while the classification accuracy of neural networks worsened as the number of clusters in the data increased from two to five.

Comparative performance of the FSCL neural net and K-means algorithm for market segmentation

Abstract Given the success of neural networks in a variety of applications in engineering, such as speech and image quantization, it is natural to consider its application to similar problems in other domains. A related problem that arises in business is market segmentation for which clustering techniques are used. In this paper, we explore the ability of a specific neural network, namely the Frequency-Sensitive Competitive Learning Algorithm (FSCL), to cluster data for developing strategic marketing decisions. To this end, we investigate the comparative performance of FSCL vis-a-vis the K-means clustering technique. A cluster analysis conducted on brand choice data for the coffee category revealed that the two methodologies resulted in widely differing cluster solutions. In an effort to address the dispute over the appropriate methodology, a comparative performance investigation was undertaken using simulated data with known cluster solutions in a fairly large experimental design to mimic varying data quality to reflect data collection and measurement error. Based on the results of these studies, it is observed that a combination of the two methodologies, wherein the results of the FSCL network are input as seeds to the K-means, seems to provide more managerially insightful segmentation schemes.

Assessing Data Quality for Information Products: Impact of Selection, Projection, and Cartesian Product

The cost associated with making decisions based on poor-quality data is quite high. Consequently, the management of data quality and the quality of associated data management processes has become critical for organizations. An important first step in managing data quality is the ability to measure the quality of information products (derived data) based on the quality of the source data and associated processes used to produce the information outputs. We present a methodology to determine two data quality characteristics--accuracy and completeness--that are of critical importance to decision makers. We examine how the quality metrics of source data affect the quality for information outputs produced using the relational algebra operations selection, projection, and Cartesian product. Our methodology is general, and can be used to determine how quality characteristics associated with diverse data sources affect the quality of the derived data.

Scheduling advertisements on a web page to maximize revenue

Many web sites (e.g. Hotmail, Yahoo) provide free services to the users while generating revenues from advertising. Advertising revenue is, therefore, critical for these sites. This in turn raises the question, how should advertisements at a web site be scheduled in a planning horizon to maximize revenue. Advertisements on the web are specified by geometry and display frequency and both of these factors need to be considered in developing a solution to the advertisement scheduling problem. Since this problem belongs to the class of NP-hard problems, we first develop a heuristic called LSMF to solve the problem. This heuristic is then combined with a genetic algorithm (GA) to develop a hybrid GA. The hybrid GA solution is first compared with the upper bound obtained by running CPLEX for the integer programming formulation of the problem. It is then also compared with an existing algorithm proposed in the literature. Our computational results show that the hybrid GA performs exceptionally well in the sense that it provides optimal or near optimal solution for a wide range of problem instances of realistic sizes and the improvements over existing algorithm are substantial. Finally we present a case study to illustrate how revenue could be significantly increased with a small improvement in the advertisement schedule. It is the first such study in this setup.

Organizational decision support systems

Abstract Decision support systems have traditionally been discussed within the context of individual or group decision making. In this paper we study decision support systems from an organizational perspective. We propose a framework for designing an organizational decision support system that is based on a network of knowledgebased systems. Nodes of this network interact with each other, as well as various other organizational systems, to provide comprehensive decision support. This network is also utilized to provide effective support for formal multi-participant decision making.

Development of hybrid genetic algorithms for product line designs

In this paper, we investigate the efficacy of artificial intelligence (AI) based meta-heuristic techniques namely genetic algorithms (GAs), for the product line design problem. This work extends previously developed methods for the single product design problem. We conduct a large scale simulation study to determine the effectiveness of such an AI based technique for providing good solutions and bench mark the performance of this against the current dominant approach of beam search (BS). We investigate the potential advantages of pursuing the avenue of developing hybrid models and then implement and study such hybrid models using two very distinct approaches: namely, seeding the initial GA population with the BS solution, and employing the BS solution as part of the GA operators process. We go on to examine the impact of two alternate string representation formats on the quality of the solutions obtained by the above proposed techniques. We also explicitly investigate a critical managerial factor of attribute importance in terms of its impact on the solutions obtained by the alternate modeling procedures. The alternate techniques are then evaluated, using statistical analysis of variance, on a fairly large number of data sets, as to the quality of the solutions obtained with respect to the state-of-the-art benchmark and in terms of their ability to provide multiple, unique product line options.

An Experimental Analysis of Face to Face versus Computer Mediated Communication Channels

Experimental research on Group Decision Support Systems (GDSS) has generally focused on democratic groups whose members typically share the same objectives. In organizations, however, there are many situations where groups have a leader who has the power to override the groups recommendation, the objective of the leader may not be the same as the objective of each member, and not everyone may have the same information. This paper reports the results of an experiment in which the groups, having a designated leader, worked on a mixed-motive task. Within this context, we analyze group decision outcomes and processes for groups that use a face-to-face channel of cormnunication and those that utilize computer mediated communication. We compare performance of the leader and members with respect to an objective measure of performance, the efficient frontier. The results indicate that for this task groups using face-to-face channel outperform groups using computer mediated communication.

Breast cancer prediction using the isotonic separation technique

A recently developed data separation/classification method, called isotonic separation, is applied to breast cancer prediction. Two breast cancer data sets, one with clean and sufficient data and the other with insufficient data, are used for the study and the results are compared against those of decision tree induction methods, linear programming discrimination methods, learning vector quantization, support vector machines, adaptive boosting, and other methods. The experiment results show that isotonic separation is a viable and useful tool for data classification in the medical domain. 2006 Elsevier B.V. All rights reserved.

Internet content filtering using isotonic separation on content category ratings

The World Wide Web has enabled anybody with a low cost Internet connection to access vast information repositories. Some of these repositories contain information (e.g., hate speech and pornography) that is considered objectionable, especially for children to view. Several efforts---legal and technical---are underway to protect children and the generic public from accessing this type of content. We propose a technical approach utilizing a recently proposed technique called isotonic separation for filtering with content metadata if they satisfy monotone conditions. We illustrate this approach using a category rating method of PICS. In essence, we formulate the Internet content filtering problem as a classification problem on content metadata and report on experiments we conducted with the isotonic separation technique.

Isotonic Separation

Data classification and prediction problems are prevalent in many domains. The need to predict to which class a particular data point belongs has been seen in areas such as medical diagnosis, credit rating, Web filtering, prediction, and stock rating. This has led to strong interest in developing systems that can accurately classify data and predict outcome. The classification is typically based on the feature values of objects being classified. Often, a form of ordering relation, defined by feature values, on the objects to be classified is known. For instance, the objects belonging to one class have larger (or smaller) feature values than do those in the other class. Exploiting this characteristic of isotonicity, we propose a data-classification method called isotonic separation based on linear programming, especially network programming. The paper also addresses an extension of the isotonic-separation method for continuous outcome prediction. Applications of the isotonic separation for discrete outcome prediction and its extension for continuous outcome prediction are shown to illustrate its applicability.