Two-Group Classification Using Neural Networks*

Artificial neural networks are new methods for classification. We investigate two important issues in building neural network models; network architecture and size of training samples. Experiments were designed and carried out on two-group classification problems to find answers to these model building questions. The first experiment deals with selection of architecture and sample size for different classification problems. Results show that choice of architecture and choice of sample size depend on the objective: to maximize the classification rate of training samples, or to maximize the generalizability of neural networks. The second experiment compares neural network models with classical models such as linear discriminant analysis and quadratic discriminant analysis, and nonparametric methods such as k-nearest-neighbor and linear programming. Results show that neural networks are comparable to, if not better than, these other methods in terms of classification rates in the training samples but not in the test samples.     

The Effect of Misclassification Costs on Neural Network Classifiers

The potential of neural networks for classification problems has been established by numerous successful applications reported in the literature. One of the major assumptions used in almost all studies is the equal cost consequence of misclassification. With this assumption, minimizing the total number of misclassification errors is the sole objective in developing a neural network classifier. Often this is done simply to ease model development and the selection of classification decision points. However, it is not appropriate for many real situations such as quality assurance, direct marketing, bankruptcy prediction, and medical diagnosis where misclassification costs have unequal consequences for different categories. In this paper, we investigate the issue of unequal misclassification costs in neural network classifiers. Through an application in thyroid disease diagnosis, we find that different cost considerations have significant effects on the classification performance and that appropriate use of cost information can aid in optimal decision making. A cross-validation technique is employed to alleviate the problem of bias in the training set and to examine the robustness of neural network classifiers with regard to sampling variations and cost differences.     

Optimal Cost Sampling for Decision Making with Multiple Regression Models

This paper develops an explicit relationship between sample size, sampling error, and related costs for the application of multiple regression models in observational studies. Graphs and formulas for determining optimal sample sizes and related factors are provided to facilitate the application of the derived models. These graphs reveal that, in most cases, the imprecision of estimates and minimum total cost are relatively insensitive to increases in sample size beyond n=20. Because of the intrinsic variation of the regression model, even if larger samples are optimal, the relative change in the total cost function is small when the cost of imprecision is a quadratic function. A model-utility approach, however, may impose a lower bound on sample size that requires the sample size be larger than indicated by the estimation or cost-minimization approaches. Graphs are provided to illustrate lower-bound conditions on sample size. Optimal sample size in view of all considerations is obtained by the maximin criterion, the maximum of the minimum sample size for all approaches.     

Private Ownership versus Institutional Ownership: Exclusion Pricing of Initial Public Offerings

The authors examine the literature with respect to the pricing of initial public offerings and focus upon the relationship of pricing to the structure and conduct of the investment banking industry. Using a data base of all share offerings undertaken in the United States over a two and a half year period, the authors find that there is considerable evidence for the proposition that large, prestigious, and well capitalised investment banks tend to price their share offerings at a higher absolute level than those not meeting such characteristics. Using classical statistical methods, the authors find that the pricing strategy of investment banks is connected to their affiliation with investment funds and unit trusts. The motives for such pricing strategies, the authors argue, lie with the affiliation of investment banks with investment funds, suggesting that the pricing of new share offerings may be a means of excluding retail investors from participating in the strong returns such issues exhibit. The authors raise legal and regulatory implications of their findings in the context of the general consolidation observed within the investment banking industry. This revised version was published online in July 2006 with corrections to the Cover Date.     

Leveraging the Strengths of Choice Models and Neural Networks: A Multiproduct Comparative Analysis*

Choice models and neural networks are two approaches used in modeling selection decisions. Defining model performance as the out‐of‐sample prediction power of a model, we test two hypotheses: (i) choice models and neural network models are equal in performance, and (ii) hybrid models consisting of a combination of choice and neural network models perform better than each stand‐alone model. We perform statistical tests for two classes of linear and nonlinear hybrid models and compute the empirical integrated rank (EIR) indices to compare the overall performances of the models. We test the above hypotheses by using data for various brand and store choices for three consumer products. Extensive jackknifing and out‐of‐sample tests for four different model specifications are applied for increasing the external validity of the results. Our results show that using neural networks has a higher probability of resulting in a better performance. Our findings also indicate that hybrid models outperform stand‐alone models, in that using hybrid models guarantee overall results equal or better than the two stand‐alone models. The improvement is particularly significant in cases where neither of the two stand‐alone models is very accurate in prediction, indicating that the proposed hybrid models may capture aspects of predictive accuracy that neither stand‐alone model is capable of on their own. Our results are particularly important in brand management and customer relationship management, indicating that multiple technologies and mixture of technologies may yield more accurate and reliable outcomes than individual ones.     

Network Models for Fleet Management*

This paper presents network models to determine the optimal replacement policy for a fleet of vehicles over a finite planning horizon. First, a minimum cost-flow model is developed to determine the optimal replacement schedule for a fleet of fixed size consisting of a single type of vehicle of various ages. The model is then extended to allow for restrictions on capital expenditures that limit the purchase of new vehicles in any time period and to allow for fluctuations in the fleet size due to planned expansion or retrenchment. Finally, a multi-commodity network model is developed for a fleet consisting of multiple vehicle types and ages.     

A Cross-Validation Analysis of Neural Network Out-of-Sample Performance in Exchange Rate Forecasting

Econometric methods used in foreign exchange rate forecasting have produced inferior out-of-sample results compared to a random walk model. Applications of neural networks have shown mixed findings. In this paper, we investigate the potentials of neural network models by employing two cross-validation schemes. The effects of different in-sample time periods and sample sizes are examined. Out-of-sample performance evaluated with four criteria across three forecasting horizons shows that neural networks are a more robust forecasting method than the random walk model. Moreover, neural network predictions are quite accurate even when the sample size is relatively small.     

Designing Referral Network Structures and Decision Rules to Streamline Provision of Urgent Health and Human Services*

This study uses fully factorial computer simulation to identify referral network attributes and referral decision rules that streamline the routing of people to urgent, limited services. As an example of a scenario, the model represents vaccine delivery in a city of 100,000 people during the first 30 days of a pandemic. By modeling patterns of communication among health care providers and daily routing of overflow clients to affiliated organizations, the simulations determine cumulative effects of referral network designs and decision rules on citywide delivery of available vaccines. Referral networks generally improve delivery rates when compared with random local search by clients. Increasing the health care organizations’ tendencies to form referral partnerships from zero to about four partners per organization sharply increases vaccine delivery under most conditions, but further increases in partnering yield little or no gain in system performance. When making referrals, probabilistic selection among partner organizations that have any capacity to deliver vaccines is more effective than selection of the highest‐capacity partner, except when tendencies to form partnerships are very low. Implications for designing health and human service referral networks and helping practitioners optimize their use of the networks are discussed. Suggestions for using simulations to model comparable systems are provided.     

Theoretical Distributions of Optima for Univariate Discrimination of Random Data*

Optimal linear discriminant models maximize percentage accuracy for dichotomous classifications, but are rarely used because a theoretical framework that allows one to make valid statements about the statistical significance of the outcomes of such analyses does not exist. This paper describes an analytic solution for the theoretical distribution of optimal values for univariate optimal linear discriminant analysis, under the assumption that the data are random and continuous. We also present the theoretical distribution for sample sizes up to N= 30. The discovery of a statistical framework for evaluating the performance of optimal discriminant models should greatly increase their use by scientists in all disciplines.     

基于神经网络的洪水灾害预测方法

在工程应用的实际预测中,主要有两类不同性质的问题,一类是强结构化问题,另一类是弱结构化问题。而洪水灾害系统的预测问题,基本上属于弱结构化问题。人工神经网络是处理弱结构化问题的强有力的工具,人工神经网络所具有的一些特征表明了它适用于洪水灾害的预测。本文建立了用于洪水灾害预测的人工神经网络模型,阐述了其基本原理,并结合实例说明了其应用。     

Neural Network Earnings per Share Forecasting Models: A Comparative Analysis of Alternative Methods

In this paper, we present a comparative analysis of the forecasting accuracy of univariate and multivariate linear models that incorporate fundamental accounting variables (i.e., inventory, accounts receivable, and so on) with the forecast accuracy of neural network models. Unique to this study is the focus of our comparison on the multivariate models to examine whether the neural network models incorporating the fundamental accounting variables can generate more accurate forecasts of future earnings than the models assuming a linear combination of these same variables. We investigate four types of models: univariate‐linear, multivariate‐linear, univariate‐neural network, and multivariate‐neural network using a sample of 283 firms spanning 41 industries. This study shows that the application of the neural network approach incorporating fundamental accounting variables results in forecasts that are more accurate than linear forecasting models. The results also reveal limitations of the forecasting capacity of investors in the security market when compared to neural network models.     

Inductive,Evolutionary, and Neural Computing Techniques for Discrimination: A Comparative Study*

This paper provides a comparative study of machine learning techniques for two-group discrimination. Simulated data is used to examine how the different learning techniques perform with respect to certain data distribution characteristics. Both linear and nonlinear discrimination methods are considered. The data has been previously used in the comparative evaluation of a number of techniques and helps relate our findings across a range of discrimination techniques.     

Neural Networks: A New Tool for Predicting Thrift Failures*

A neural network model that processes input data consisting of financial ratios is developed to predict the financial health of thrift institutions. The network's ability to discriminate between healthy and failed institutions is compared to a traditional statistical model. The differences and similarities in the two modelling approaches are discussed. The neural network, which uses the same financial data, requires fewer assumptions, achieves a higher degree of prediction accuracy, and is more robust.     

The Efficacy of Neural Networks in Predicting Returns on Stock and Bond Indices*

This paper uses two recently developed tests to identify neglected nonlinearity in the relationship between excess returns on four asset classes and several economic and financial variables. Having found some evidence of possible nonlinearity, it was then investigated whether the predictive power of these variables could be enhanced by using neural network models instead of linear regression or GARCH models. Some evidence of nonlinearity in the relationships between the explanatory variables and large stocks and corporate bonds was found. It was also found that the GARCH models are conditionally efficient with respect to neural network models, but the neural network models outperform GARCH models if financial performance measures are used. In resonance with the results reported for the tests for neglected nonlinearity, it was found that the neural network forecasts are conditionally efficient with respect to linear regression models for large stocks and corporate bonds, whereas the evidence is not statistically significant for small stocks and intermediate-term government bonds. This difference persists even when financial performance measures for individual asset classes are used for comparison.     

基于人工神经网络的三江平原土壤质量综合评价与预测模型

根据土壤质量定量评价指标分级体系生成足够多代表性好的神以网络训练和检验用的样本。建立神经网络模型时,利用删减或扩张准则确定神经网络最佳拓扑结构,避免“过拟合”现象,利用检验样本监控在训练过程中不发生“过学习”现象,使建立的土壤质量的综合评价与预测模型具有较好的泛化能力和预测能力。对三江平原地区主要耕作土壤质量的综合评价与预测结果表明,神经网络方法能较好地应用于土壤质量综合评价与预测,比加权综合指数法能更精细地评价与预测土壤的变化趋。     

Data Mining for Network Intrusion Detection: A Comparison of Alternative Methods*

Intrusion detection systems help network administrators prepare for and deal with network security attacks. These systems collect information from a variety of systems and network sources, and analyze them for signs of intrusion and misuse. A variety of techniques have been employed for analysis ranging from traditional statistical methods to new data mining approaches. In this study the performance of three data mining methods in detecting network intrusion is examined. An experimental design (3times2x2) is created to evaluate the impact of three data mining methods, two data representation formats, and two data proportion schemes on the classification accuracy of intrusion detection systems. The results indicate that data mining methods and data proportion have a significant impact on classification accuracy. Within data mining methods, rough sets provide better accuracy, followed by neural networks and inductive learning. Balanced data proportion performs better than unbalanced data proportion. There are no major differences in performance between binary and integer data representation.     

Decision Models for Robot Selection: A Comparison of Ordinary Least Squares and Linear Goal Programming Methods

The profusion of robot designs, the cost of testing, and the fact that robot operational parameter maximums are often mutually exclusive are factors that create a complex selection decision for the potential user. While formal robot testing standards are now in place, formal techniques to select robots for the testing process have not been addressed. A linear goal programming model is an effective tool for the decision maker for optimizing the robot selection process in terms of requirement priorities. It is also shown that this model provides a more stable result than the ordinary least squares estimator in the presence of statistical outliers of robot parameters. The methodology is illustrated through the use of current robot specifications.     

Optimal Design of a Two-Level Transportation Network with a Different Unit Cost for Each Secondary Link*

The objective of this paper is to further develop Singhal's (1990) framework for designing a two-level hierarchical transportation network consisting of a trunk or primary link and several feeder or secondary links. Secondary links are perpendicular lines from each of the given points and the primary link is a straight line or curve connecting the feet of two extreme perpendicular lines. The problem and the associated strategic and operational considerations such as cost, time, feasibility, and preferred regions for the primary link in the context of rural highway planning are discussed. Two-level networks are also common in electricity transmission, pipelines, and telecommunication design. The core of the framework is a model for finding the path of a primary link such that a weighted sum of the lengths of the perpendicular lines from each point to a linear primary link and the distance between the feet of the two extreme perpendicular lines is minimized. The analysis shows that for almost every problem there exists a wide range of solutions for which the total cost is only slightly higher than that of the optimal solution. This offers considerable flexibility to the decision maker. These solutions can be evaluated in view of the broader objectives and constraints that are not included in the model. The use of computer graphics and the option of a nonlinear or piecewise linear primary link are also discussed.     

Integrating Neural Networks and Semi-Markov Processes for Automated Knowledge Acquisition: An Application to Real-time Scheduling*

Recently, artificial neural networks (ANN) have gained attention as a promising modeling tool for building intelligent systems. A number of applications have been reported in areas varying from pattern recognition to bankruptcy prediction. In this paper, we present a creative methodology that integrates computer simulation, semi-Markov optimization, and ANN techniques for automated knowledge acquisition in real-time scheduling. The integrated approach focuses on the synergy between operations research and ANN in eliciting human knowledge, filtering inconsistent data, and building competent models capable of performing at the expert level. The new approach includes three main components. First, computer simulation is used to collect expert decisions. This step allows expert knowledge to be obtained in a non-intrusive way and minimizes the difficulties involved in interviewing experts, constructing repertory grids, or using other similar structures required for manual knowledge acquisition. The data collected from computer simulation are then optimized using a semi-Markov decision model to remove data redundancies, inconsistencies, and errors. Finally, the optimized data are used to build ANN-based expert systems. The integrated approach is evaluated by comparing it with the human expert and using ANN alone in the domain of real-time scheduling. The results indicate that ANN-based systems perform worse than human experts from whom the data were collected, but the integrated approach outperforms human experts and ANN models alone.