He has published 450 journal and conference papers, authored or co-authored three books, including the pioneering neural networks text "Introduction to Artificial Neural Systems" (1992), and co-edited a number of volumes in Springer Lecture Notes in Computer Science. His books and articles were cited over 17,000 times (Google Scholar, 2022).
This study investigates the effect of class imbalance in training data when developing neural network classifiers for computer-aided medical diagnosis. The investigation is performed in the presence of other characteristics that are typical among medical data, namely small training sample size, large number of features, and correlations between features. Two methods of neural network training are explored: classical backpropagation (BP) and particle swarm optimization (PSO) with clinically relevant training criteria. An experimental study is performed using simulated data and the conclusions are further validated on real clinical data for breast cancer diagnosis. The results show that classifier performance deteriorates with even modest class imbalance in the training data. Further, it is shown that BP is generally preferable over PSO for imbalanced training data especially with small data sample and large number of features. Finally, it is shown that there is no clear preference between oversampling and no compensation approach and some guidance is provided regarding a proper selection.
Lack of precision is common in property value assessment. Recently non-conventional methods, such as neural networks based methods, have been introduced in property value assessment as an attempt to better address this lack of precision and uncertainty. Although fuzzy logic has been suggested as another possible solution, no other artificial intelligence methods have been applied to real estate value assessment other than neural network based methods. This paper presents the results of using two new non-conventional methods, fuzzy logic and memory-based reasoning, in evaluating residential property values for a real data set. The paper compares the results with those obtained using neural networks and multiple regression. Methods of feature reduction, such as principal component analysis and variable selection, have also been used for possible improvement of the final results. The results indicate that no single one of the new methods is consistently superior for the given data set.
The output of the algorithm is the best global position found during all iterations. Even though PSO convergence to a global optimum has not been proven for the general case (some results on the convergence can be found in (Clerc and Kennedy, 2002)), the algorithm has been shown efficient for many optimization problems including training neural networks (Kennedy and Eberhart, 1995).
Neurocomputing methods are loosely based on a model of the brain as a network of simple interconnected processing elements corresponding to neurons. These methods derive their power from the collective processing of artificial neurons, the chief advantage being that such systems can learn and adapt to a changing environment. In knowledge-based neurocomputing, the emphasis is on the use and representation of knowledge about an application. Explicit modeling of the knowledge represented by such a system remains a major research topic. The reason is that humans find it difficult to interpret the numeric representation of a neural network.
For many years lenders have been using traditional statistical techniques such as logistic regression and discriminant analysis to more precisely distinguish between creditworthy customers who are granted loans and non-creditworthy customers who are denied loans. More recently new machine learning techniques such as neural networks, decision trees, and support vector machines have been successfully employed to classify loan applicants into those who are likely to pay a loan off or default upon a loan. Accurate classification is beneficial to lenders in terms of increased financial profits or reduced losses and to loan applicants who can avoid overcommitment. This paper examines a historical data set from consumer loans issued by a German bank to individuals whom the bank considered to be qualified customers. The data set consists of the financial attributes of each customer and includes a mixture of loans that the customers paid off or defaulted upon. The paper examines and compares the classification accuracy rates of three decision tree techniques as well as analyzes their ability to generate easy to understand rules.
This paper describes a first effort to design and implement an adaptive neuro-fuzzy inference system-based approach to estimate prices for residential properties. The data set consists of historic sales of houses in a market in the Midwest region of the United States and it contains parameters describing typical residential property features and the actual sale price. The study explores the use of fuzzy inference systems to assess real estate property values and the use of neural networks in creating and fine-tuning the fuzzy rules used in the fuzzy inference system. The results are compared with those obtained using a traditional multiple regression model. The paper also describes possible future research in this area.
A neural network was developed that utilizes both clinical and imaging (CT and MRI) data to predict posterior fossa tumor (PFT) type. Data from 33 children with PFTs were used to develop and test the system. When all desired information was available, the network was able to correctly classify 85.7% of the tumors. In cases with incomplete data, it was able to correctly classify 72.7% of the tumors. In both instances, the diagnoses made by the network were more likely to be correct than those made by the neuroradiologists.
de7ca951e2