Michael Egmont-Petersen

IMAGE PROCESSING WITH NEURAL NETWORKS–A REVIEW

Abstract We review more than 200 applications of neural networks in image processing and discuss the present and possible future role of neural networks, especially feed-forward neural networks, Kohonen feature maps and Hopfield neural networks. The various applications are categorised into a novel two-dimensional taxonomy for image processing algorithms. One dimension specifies the type of task performed by the algorithm: preprocessing, data reduction/feature extraction, segmentation, object recognition, image understanding and optimisation. The other dimension captures the abstraction level of the input data processed by the algorithm: pixel-level, local feature-level, structure-level, object-level, object-set-level and scene characterisation. Each of the six types of tasks poses specific constraints to a neural-based approach. These specific conditions are discussed in detail. A synthesis is made of unresolved problems related to the application of pattern recognition techniques in image processing and specifically to the application of neural networks. Finally, we present an outlook into the future application of neural networks and relate them to novel developments.

Bayesian network classifiers for identifying the slope of the customer lifecycle of long-life customers

Undoubtedly, Customer Relationship Management (CRM) has gained its importance through the statement that acquiring a new customer is several times more costly than retaining and selling additional products to existing customers. Consequently, marketing practitioners are currently often focusing on retaining customers for as long as possible. However, recent findings in relationship marketing literature have shown that large differences exist within the group of long-life customers in terms of spending and spending evolution. Therefore, this paper focuses on introducing a measure of a customers future spending evolution that might improve relationship marketing decision making. In this study, from a marketing point of view, we focus on predicting whether a newly acquired customer will increase or decrease his/her future spending from initial purchase information. This is essentially a classification task. The main contribution of this study lies in comparing and evaluating several Bayesian network classifiers with statistical and other artificial intelligence techniques for the purpose of classifying customers in the binary classification problem at hand. Certain Bayesian network classifiers have been recently proposed in the artificial intelligence literature as probstudy. We discuss and evaluate several types of Bayesian network classifiers and their corresponding structure learning algorithms. We contribute to the literature by providing experimental evidence that: (1) Bayesian network classifiers offer an interesting and viable alternative for our customer lifecycle slope estimation problem; (2) the Markov Blanket concept allows for a natural form of attribute selection that was very effective for the application at hand; (3) the sign of the slope can be predicted with a powerful and parsimonious general, unrestricted Bayesian network classifier; (4) a set of three variables measuring the volume of initial purchases and the degree to which customers originally buy in different categories, are powerful predictors for estimating the sign of the slope, and might therefore provide desirable additional information for relationship marketing decision making.

Learning Bayesian network classifiers for credit scoring using Markov chain Monte Carlo search

In this paper, we evaluate the power and usefulness of Bayesian network classifiers (probabilistic networks) for credit scoring. Various types of Bayesian network classifiers are evaluated and contrasted including unrestricted Bayesian network classifiers learning using Markov Chain Monte Carlo (MCMC) search. Experiments were carried out on three real life credit scoring data sets. It is shown that MCMC Bayesian network classifiers have a very good performance and by using the Markov blanket concept, a natural form of feature selection is obtained, which results in parsimonious and powerful models for financial credit scoring.

The Effects of Disregarding Test Characteristics in Probabilistic Networks

In most medical disciplines, the results from diagnostic tests are not unequivocal. To capture the uncertainties in test results, the notions of sensitivity and specificity of diagnostic tests have been introduced. Although the importance of taking these test characteristics into account in medical reasoning is stressed throughout the literature, they are often not modelled explicitly in real-life probabilistic networks. In this paper, we study the effects that disregarding the characteristics of diagnostic tests can have on the performance of a probabilistic network. We feel that the effects that we observed in a real-life network for the staging of oesophageal cancer, are likely to be found in networks for other applications in medicine as well.

Confidence intervals for probabilistic network classifiers

Probabilistic networks (Bayesian networks) are suited as statistical pattern classifiers when the feature variables are discrete. It is argued that their white-box character makes them transparent, a requirement in various applications such as, e.g., credit scoring. In addition, the exact error rate of a probabilistic network classifier can be computed without a dataset. First, the exact error rate for probabilistic network classifiers is specified. Secondly, the exact sampling distribution for the conditional probability estimates in a probabilistic network classifier is derived. Each conditional probability is distributed according to the bivariate binomial distribution. Subsequently, an approach for computing the sampling distribution and hence confidence intervals for the posterior probability in a probabilistic network classifier is derived. Our approach results in parametric bootstrap confidence intervals. Experiments with general probabilistic network classifiers, the Naive Bayes classifier and tree augmented Naive Bayes classifiers (TANs) show that our approximation performs well. Also simulations performed with the Alarm network show good results for large training sets. The amount of computation required is exponential in the number of feature variables. For medium and large-scale classification problems, our approach is well suited for quick simulations. A running example from the domain of credit scoring illustrates how to actually compute the sampling distribution of the posterior probability.

Discovery of regulatory connections in microarray data

In this paper, we introduce a new approach for mining regulatory interactions between genes in microarray time series studies. A number of preprocessing steps transform the original continuous measurements into a discrete representation that captures salient regulatory events in the time series. The discrete representation is used to discover interactions between the genes. In particular, we introduce a new across-model sampling scheme for performing Markov Chain Monte Carlo sampling of probabilistic network classifiers. The results obtained from the microarray data are promising. Our approach can detect interactions caused both by co-regulation and by control-regulation.

Feature Selection by Markov Chain Monte Carlo Sampling – A Bayesian Approach

We redefine the problem of feature selection as one of model selection and propose to use a Markov Chain Monte Carlo method to sample models. The applicability of our method is related to Bayesian network classifiers. Simulation experiments indicate that our novel proposal distribution results in an ignorant proposal prior. Finally, it is shown how the sampling can be controlled by a regularization prior.