Ray R. Hashemi

A hybrid intelligent system for predicting bank holding structures

A composite model of neural network and rough sets components was constructed to predict a sample of bank holding patterns. The final model was able to correctly classify 96% of a testing set of four types of bank holding structures. Holding structure is defined as the number of banks under common ownership. For this study, forms of bank holding structure include: banks that are not owned by another company, single banks that are held by another firm, pairs of banks that are held by another enterprise, and three or more banks that are held by another company. Initially, input to the neural network model was 28 financial ratios for more than 200 banks in Arkansas for 1992. The 28 ratios are organized by categories such as liquidity, credit risk, leverage, efficiency, and profitability. The ratios were constructed with 70 bank variables such as net worth, deposits, total assets, net loans, total operating income, etc. The first neural network model correctly classified 84% of the testing set at a tolerance level of 0.20. Another artificial intelligence (AI) procedure known as two-dimensional rough sets was then applied to the dataset. Rough sets reduced the number of input variables from 28 to 18, a drop of 36% in the number of input variables. This version of rough sets also eliminated a number of records, thereby reducing the information system (i.e., matrix) on both vertical and horizontal dimensions. A second neural network was trained with the reduced number of input variables and records. This network correctly classified 96% of the testing set at a tolerance level of 0.20, an increase of 11% in the accuracy of the prediction. By applying two-dimensional reducts to the dataset of financial ratios, the predictive accuracy of the neural network model was improved substantially. Banking institutions that are prime candidates for mergers or acquisitions can then be more accurately identified through the use of this hybrid decision support system (DSS) which combines different types of AI techniques for the purposes of data management and modeling.

A neural network for transportation safety modeling

Abstract Accidents serve as an operational measure of marine safety, and specifically the safety of vessels, crews, and cargoes. The ability to accurately predict the type of vessel accident with such input variables as time, location, weather, river stage, and traffic could significantly reduce marine casualties by alerting port authorities and navigation groups as to the likelihood of a specific kind of casualty. In this paper, three models were developed to predict vessel accidents on the lower Mississippi River. These models are a neural network, multiple discriminant analysis and logistic regression. The predictive capability for vessel accidents of a neural network is compared with multiple discriminant analysis and logistic regression. The percent of “grouped” cases correctly classified is 80% (36 of the 45 cases in the testing set) for the neural network, if nonclassified cases are treated as incorrectly classified by neural network. The percent of “grouped” cases correctly classified by this network is 90% (36 of 40 cases) if nonclassified cases are excluded from the calculation. Discriminant analysis and logistic regression were able to correctly classify only 53% and 56% respectively, of accident cases into three casualty groups: collisions, rammings, or groundings.

BUILDING AN ORGAN-SPECIFIC CARCINOGENIC DATABASE FOR SAR ANALYSES

FDA reviewers need a means to rapidly predict organ-specific carcinogenicity to aid in evaluating new chemicals submitted for approval. This research addressed the building of a database to use in developing a predictive model for such an application based on structure–activity relationships (SAR). The Internet availability of the Carcinogenic Potency Database (CPDB) provided a solid foundation on which to base such a model. The addition of molecular structures to the CPDB provided the extra ingredient necessary for SAR analyses. However, the CPDB had to be compressed from a multirecord to a single record per chemical database; multiple records representing each gender, species, route of administration, and organ-specific toxicity had to be summarized into a single record for each study. Multiple studies on a single chemical had to be further reduced based on a hierarchical scheme. Structural cleanup involved removal of all chemicals that would impede the accurate generation of SAR type descriptors from commercial software programs; that is, inorganic chemicals, mixtures, and organometallics were removed. Counterions such as Na, K, sulfates, hydrates, and salts were also removed for structural consistency. Structural modification sometimes resulted in duplicate records that also had to be reduced to a single record based on the hierarchical scheme. The modified database containing 999 chemicals was evaluated for liver-specific carcinogenicity using a variety of analysis techniques. These preliminary analyses all yielded approximately the same results with an overall predictability of about 63%, which was comprised of a sensitivity of about 30% and a specificity of about 77%.

A FUSION OF ROUGH SETS, MODIFIED ROUGH SETS, AND GENETIC ALGORITHMS FOR HYBRID DIAGNOSTIC SYSTEMS

A hybrid classification system is a system composed of several intelligent techniques such that the inherent limitations of one individual technique be compensated for by the strengths of another technique. In this paper, we investigate the outline of a hybrid diagnostic system for Attention Deficit Disorder (ADD) in children. This system uses Rough Sets (RS) and Modified Rough Sets (MRS) to induce rules from examples and then uses our modified genetic algorithms to globalize the rules. Also, the classification capability of this hybrid system was compared with the behavior of (a) another hybrid classification system using RS, MRS, and the “dropping condition” approach, (b) the Interactive Dichotomizer 3 (ID3) approach, and (c) a basic genetic algorithm.

Pattern development for vessel accidents: a comparison of statistical and neural computing techniques

Abstract This paper describes a sample of over 900 vessel accidents that occurred on the lower Mississippi River. Two different techniques, one statistical and the other based on a neural network model, were used to build logical groups of accidents. The objective in building the groups was to maximize between-group variation and minimize within-group variation. The result was groups whose records were as homogenous as possible. A clustering algorithm (i.e., a non-inferential statistical technique) generated sets of three, four and five groups. A Kohenen neural network model (i.e., a self-organizing map) also generated sets of three, four and five groups. The two sets of parallel groups were radically different as to the relative number of records in each group. In other words, when the two sets of groups were constructed by the respective techniques, the membership of each comparable group within the two different sets was substantially different. Not only was the respective record count in each group substantially different, so were the descriptive statistics describing each comparable set of groups. These results have significant implications for marine policy makers. Important policy variables include safety factors such as weather, speed of current, time of operation, and location of accidents, but mandatory utilization of a voluntary vessel tracking service may be subject to debate.

Development of group's signature for evaluation of skin cancer in mice caused by ultraviolet radiation

In this research effort, the effect of UVC (260 nm) on the skin of one month old Balb/c mice exposed for a total of 100 hours is studied. The goal is to identify those independent variables in the experimental group that have a significant change in their measurements in compare to the measurements of their counterparts in the control group. To meet the goal, we create signatures for both experimental and control groups using the Kohonen self-organizing map. The comparison of signatures to each other delivers the significant changes in the independent variables between the two groups. The findings are compared with another set of findings obtained from using analysis of variance. The results reveal that using signature approach that is created based on the SOM methodology, is a viable tool for this type of analysis.

A Fuzzy Rough Sets Classifier for Database Mining

Object classifiers are important tool for relational database mining. These classifiers are trained by existing relations in the database and then are used for classifying new objects. Rough Set (RS) classifiers have been shown to be useful in classification queries. These classifiers do not, however, perform adequately when the attribute set of a new object does not precisely match the antecedent of classification rules. We combine the capabilities of the Fuzzy Set and Rough Set theories to create a Fuzzy-Rough Set (FRS) classifier that alleviates this shortcoming of the RS classifiers. For a data set we perform a statistical analysis of the classification power of the proposed classifier. In addition, we compare the performance of the new classifier with that of three other classifiers, using eight additional data sets. All data sets were obtained from actual experiments. The results show that the FRS classifier outperforms the other three classifiers.

Identifying and testing of signatures for non-volatile biomolecules using tandem mass spectra

Identification of volatile and semi-volatile molecules using traditional electron ionization mass spectrometry has been successful. The major contributor to this success is the reproduceability of the mass spectra, which allow identification of components based on comparison of fragmentation patterns within very large databases. However, this approach is not useful for the identification of typical nonvolatile biomolecules. Tandem mass spectrometry with collision induced dissociation (CID) has the potential to provide structure-specific fragmentation from non-volatile biomolecules.The recognition of these molecules based on CID is not an easy task, since the spectra generated for a given molecule are not as reproducible as in traditional electron ionization mass spectrometry. Also, the rules governing the formation of CID produced ions are not completely understood.In this study we investigate the use of the Kohonen Self-Organized Mapping (SOM) neural network to generate and test signatures (fragmentation patterns) for a given set of non-volatile biomolecules using spectra generated by tandem mass spectrometry with CID. The signatures then may be used as a discriminator for identifying unknown non-volatile biomolecules.Identification of volatile and semi-volatile molecules using traditional electron ionization mass spectrometry has been successful. The major contributor to this success is the reproduceability of the mass spectra, which allow identification of components based on comparison of fragmentation patterns within very large databases. However, this approach is not useful for the identification of typical nonvolatile biomolecules. Tandem mass spectrometry with collision induced dissociation (CID) has the potential to provide structure-specific fragmentation from non-volatile biomolecules.The recognition of these molecules based on CID is not an easy task, since the spectra generated for a given molecule are not as reproducible as in traditional electron ionization mass spectrometry. Also, the rules governing the formation of CID produced ions are not completely understood.In this study we investigate the use of the Kohonen Self-Organized Mapping (SOM) neural network to generate and test signatures (fragmentation patterns) for a given set of non-volatile biomolecules using spectra generated by tandem mass spectrometry with CID. The signatures then may be used as a discriminator for identifying unknown non-volatile biomolecules.

SASSY: A Petri Net Based Student-Driven Advising Support System

This paper introduces the Course-Petri net, a specialization of Petri net which is used as the foundation for development of an advising system, SASSY. The system suggests courses for an advisee based on (i) Frequency of the course offering (ii) Balancing the course load, (iii) Shortening the path length to graduation, (iv) Preference of advisee, and (v) Entertaining different scenarios of course loads for the entire duration of the advisee’s university life. A prototype of SASSY was informally evaluated by ten advisees. The preliminary findings revealed that SASSY received approval of all advisees and it has a high potential for use in a student-centered learning environment.

Knowledge discovery from sparse pharmacokinetic data

In this research effort, we show that the following hypothesis is true: The independently verified sparse information secured from the scientific literature regarding the effects of methyl mercury on mice enables us to predict the effects of the methyl mercury on humans. The Rough Sets methodology is used in this endeavor.