Search

Abstract: The paper deals with the problem of predicting the time to default in credit behavioural scoring. This area opens a possibility of including a dynamic component in behavioural scoring modelling which enables making decisions related to limit, collection and recovery strategies, retention and attrition, as well as providing an insight into the profitability, pricing or term structure of the loan. In this paper, we compare survival analysis and neural networks in terms of modelling and results. The neural network architecture is designed such that its output is comparable to the survival analysis output. Six neural network models were created, one for each period of default. A radial basis neural network algorithm was used to test all six models. The survival model used a Cox modelling procedure. Further, different performance measures of all models were discussed since even in highly accurate scoring models, misclassification patterns appear. A systematic comparison ‘3+2+2’ procedure is suggested to find the most effective model for a bank. Additionally, the survival analysis model is compared to neural network models according to the relative importance of different variables in predicting the time to default. Although different models can have very similar performance measures they may consist of different variables. The dataset used for the research was collected from a Croatian bank and credit customers were observed during a 12-month period. The paper emphasizes the importance of conducting a detailed comparison procedure while selecting the best model that satisfies the users’ interest. [Copyright &y& Elsevier]

Abstract: Kernel-based nonlinear characteristic extraction and classification algorithms are popular new research directions in machine learning. In this paper, we propose an improved photometric stereo scheme based on improved kernel-independent component analysis method to reconstruct 3D human faces. Next, we fetch the information of 3D faces for facial face recognition. For reconstruction, we obtain the correct normal vector’s sequence to form the surface, and use a method for enforcing integrability to reconstruct 3D objects. We test our algorithm on a number of real images captured from the Yale Face Database B, and use three kinds of methods to fetch characteristic values. Those methods are called contour-based, circle-based, and feature-based methods. Then, a three-layer, feed-forward neural network trained by a back-propagation algorithm is used to realize a classifier. All the experimental results were compared to those of the existing human face reconstruction and recognition approaches tested on the same images. The experimental results demonstrate that the proposed improved kernel independent component analysis (IKICA) method is efficient in reconstruction and face recognition applications. [ABSTRACT FROM AUTHOR]

Abstract: Software defect predictors which classify the software modules into defect-prone and not-defect-prone classes are effective tools to maintain the high quality of software products. The early prediction of defect-proneness of the modules can allow software developers to allocate the limited resources on those defect-prone modules such that high quality software can be produced on time and within budget. In the process of software defect prediction, the misclassification of defect-prone modules generally incurs much higher cost than the misclassification of not-defect-prone ones. Most of the previously developed predication models do not consider this cost issue. In this paper, three cost-sensitive boosting algorithms are studied to boost neural networks for software defect prediction. The first algorithm based on threshold-moving tries to move the classification threshold towards the not-fault-prone modules such that more fault-prone modules can be classified correctly. The other two weight-updating based algorithms incorporate the misclassification costs into the weight-update rule of boosting procedure such that the algorithms boost more weights on the samples associated with misclassified defect-prone modules. The performances of the three algorithms are evaluated by using four datasets from NASA projects in terms of a singular measure, the Normalized Expected Cost of Misclassification (NECM). The experimental results suggest that threshold-moving is the best choice to build cost-sensitive software defect prediction models with boosted neural networks among the three algorithms studied, especially for the datasets from projects developed by object-oriented language. [Copyright &y& Elsevier]

This paper presents a method for computing a thesaurus from a text corpus, and combined with a revised back-propagation neural network (BPNN) learning algorithm for document categorization. Automatically constructed thesaurus is a data structure that accomplished by extracting the relatedness between words. Neural network is one of the efficient approaches for document categorization. However the conventional BPNN has the problems of slow learning and easy to involve into the local minimum. We use a revised algorithm to improve the conventional BPNN that can overcome these problems. A well constructed thesaurus has been recognized as valuable tool in the effective operation of document categorization, it overcome some problem for the document categorization based on bag of words which ignored the relationship between words. To investigate the effectiveness of our method, we conducted the experiments on the standard Reuter-21578. The experimental results show that the proposed model was able to achieve higher categorization effectiveness as measured by the precision, recall and F-measure. [Copyright &y& Elsevier]

In this paper, we focus on the experimental analysis on the performance in artificial neural networks with the use of statistical tests on the classification task. Particularly, we have studied whether the sample of results from multiple trials obtained by conventional artificial neural networks and support vector machines checks the necessary conditions for being analyzed through parametrical tests. The study is conducted by considering three possibilities on classification experiments: random variation in the selection of test data, the selection of training data and internal randomness in the learning algorithm. The results obtained state that the fulfillment of these conditions are problem-dependent and indefinite, which justifies the need of using non-parametric statistics in the experimental analysis. [Copyright &y& Elsevier]

Abstract: This paper presents an overview of a data analysis method based on self-organizing maps (SOM), a well-known unsupervised neural network learning algorithm, which was applied to a lead-free wave soldering process. The aim of the study was to determine whether the neural network modeling method could be a useful and time-saving way to analyze data from a discrete manufacturing process, such as wave soldering, which is a widely used technique in the electronics industry to solder components on printed circuit boards. The data variables were mostly various process parameters, but also some solder defect numbers were present in the data as a measure of the product quality. The data analysis procedure went as follows. At first, the process data were modeled using the SOM-algorithm. Next, the neuron reference vectors of the formed map were clustered to reveal the desired dominating elements of each territory of the map. At the final stage, the clusters were utilized as sub-models to indicate variable dependencies in these sub-models. The results show that the method presented here can be a good way to analyze this type of process data because interesting interactions between certain process parameters and solder defects were found by means of this data-driven modeling method. [Copyright &y& Elsevier]

In this paper, a neural network model is presented to characterize the thickness and the uniformity of the cellgap process for flexible liquid crystal display (LCD). Input factors are explored via a D-optimal design with 15 runs and used as training data in the neural network. In order to verify the fitness of the model, three more runs are added as test data. Latin hypercube sampling and error back-propagation algorithm are used to build the model. Latin hypercube sampling is used to generate initial weights and biases of the network. The thickness of cellgap is measured at five points: one at the center and four at the edges. The average thickness is used as cellgap thickness, and the uniformity is obtained by comparing the thickness at the center and edge points. [Copyright &y& Elsevier]

Abstract: Nowadays license plate recognition became a key technique to many automated systems such as road traffic monitoring, automated payment of tolls on high ways or bridges, security access, and parking lots access control. Most of the previous license plate locating (LPL) approaches are not robust in case of low-quality images. Some difficulties result from illumination variance, noise, complex and dirty background. This paper presents a real-time and robust method for license plate location and recognition. Edge features of the car image are very important, and edge density and background color can be used to successfully detect a number plate location according to the characteristics of the number plate. The proposed algorithm can efficiently determine and adjust the plate rotation in skewed images. LP quantization and equalization has been applied as an important step for successful decryption of the LP. It finds the optimal adaptive threshold corresponding to the intensity image obtained after adjusting the image intensity values. An efficient character segmentation algorithm is used in order to segment the characters in the binary license plate image. An optical character recognition (OCR) engine has then been proposed. The OCR engine includes digit dilation, contours adjustment and resizing. Each digit is resized to standard dimensions according to a neural network dataset. The back-propagation neural network (BPNN) is selected as a powerful tool to perform the recognition process. Experiments have been conducted to corroborate the efficiency of the proposed method. Experimental results showed that the proposed method has excellent performance even in case of low-quality images or images exhibiting illumination effects and noise. Experimental results illustrate the great robustness and efficiency of our method. [Copyright &y& Elsevier]

Abstract: One of the significant research problems in multivariate analysis is the selection of a subset of input variables that can predict the desired output with an acceptable level of accuracy. This goal is attained through the elimination of the variables that produce noise or, are strictly correlated with other already selected variables. Feature subset selection (selection of the input variables) is important in correlation analysis and in the field of classification and modeling. This paper presents a hybrid method based on ant colony optimization and artificial neural networks (ANNs) to address feature selection. The proposed hybrid model is demonstrated using data sets from the domain of medical diagnosis, yielding promising results. [Copyright &y& Elsevier]