Recently, Internet based Electronic Commerce is recognized as one of the alternatives for strengthening sales power of small and medium companies. However, small and medium manufacturers can't adjust properly to the new environment because they are in short of money, personnel, and technology. To cope with this problem, this paper deals with the development of virtual manufacturing agent to support sales agent. The sales activity of most of parts manufacturing companies is based on orders of buyers. The process of promotion, receipt and selection of orders of the parts manufacturing is closely coupled with the load status of the production lines. On deciding whether to accept an order or not, as well as negotiating with buyers, sales person needs information such as load and schedule of production lines, manufacturability of the order. Therefore, the functions of virtual manufacturing agents manufacturability analysis, process planning, and scheduling are key features in developing an agent of sales activity for the parts manufacturing business. While most of research on virtual manufacturing system so far is focused on the simulation of each product, this paper deals with the development of agent assisting internet-based product sales by supporting production information promptly. The pilot system of virtual manufacturing agent is implemented using KQML-based agent template and Java-based expert system shell for a small molding company

Tree model is the most popular classification algorithm in data mining due to easy interpretation of the result. In CART(Breiman et al., 1984) and C4.5(Quinlan, 1993) which are representative of tree algorithms, the split fur classification proceeds to attain the homogeneous terminal nodes with respect to the composition of levels in target variable. But, fur instance, in the chum prediction modeling fur CRM(Customer Relationship management), the rate of churn is generally very low although we are interested in mining the churners. Thus it is difficult to get accurate prediction modes using tree model based on the traditional split rule, such as mini or deviance. Buja and Lee(1999) introduced a new split rule, one-sided purity for classifying minor interesting group. In this paper, we compared one-sided purity with traditional split rule, deviance analyzing churning vs. non-churning data of ISP company. Also reviewing the result of tree model based on one-sided purity with some simulated data, we discussed problems and researchable topics.

Sequential discovery from time series data has mainly concerned about events or item sets. Recently, the research has stated to applied to the numerical data. An example is sensor information generated by checking a machine state. The numerical data hardly have the same valuers while making patterns. So, it is important to extract suitable number of pattern features, which can be transformed to events or item sets and be applied to sequential pattern mining tasks. The popular methods to extract the patterns are sliding window and clustering. The results of these methods are sensitive to window sine or clustering parameters; that makes users to apply data mining task repeatedly and to interpret the results. This paper suggests the method to retrieve pattern features making numerical data into vector of an angle and a magnitude. The retrieved pattern features using this method make the result easy to understand and sequential patterns finding fast. We define an inclusion relation among pattern features using angles and magnitudes of vectors. Using this relation, we can fad sequential patterns faster than other methods, which use all data by reducing the data size.

Lazy loaming methods including CBR have relative advantages in comparison with eager loaming methods such as artificial neural networks and decision trees. However, they are very sensitive to irrelevant features. In other words, when there are irrelevant features, larry learning methods have difficulty in comparing cases. Therefore, their performance can be degraded significantly. To overcome this disadvantage, feature weighting methods for lazy loaming methods have been studied. Most of the existing researches, however, were focused on global feature weighting. In this research, we propose a new local feature weighting method, which we shall call CBDFW. CBDFW stores classification performance of randomly generated feature weight vectors. Then, given a new query case, CBDFW retrieves the successful feature weight vectors and designs a feature weight vector fur the query case. In the test on credit evaluation domain, CBDFW showed better classification accuracy when compared to the results of previous researches.

Data Envelopment Analysis(DEA), a non-parametric productivity analysis tool, has become an accepted approach for assessing efficiency in a wide range of fields. Despite of its extensive applications and merits, some features of DEA remain bothersome. DEA offers no guideline about to which direction relatively inefficient DMUs improve since a reference set of an inefficient DMU, several efficient DMUs, hardly provides a stepwise path for improving the efficiency of the inefficient DMU. In this paper, we aim to show that DEA can be used to evaluate the efficiency of life insurance companies while overcoming its limitation with the aids of machine learning methods.

Plant construction projects usually take much higher uncertainty and risks than the projects from other domains. This implies the importance of plant construction project management should be more emphasized than the other domain. Especially, the overall successes of the projects often depend on the performance of process planning and scheduling performed at the initial stage of the project. However, most plant construction projects suffer great difficulties in establishing proper process planning and scheduling timely because of unstructureness and dynamicity of environment of the project itself In this paper, we propose a knowledge-based process planning and scheduling approach in a plant construction domain to cope this problem. First, we modulize process planning knowledge and present the knowledge representation scheme. Second, we propose an inferencing mechanism to build a process planning for plant construction based on the represented process planning knowledge. Since our approach automate the initial process planning, which was usually done by manual way, it can improve the correctness and also completeness of the process plan and schedule by reducing the time to plan and allowing simulations on the various situation. We also design and implement this our approach as a real working system, and it is successfully applied to real plant construction cases from a leading construction company in Korea. Based on this success, we expect our approach can be easily applied to the projects of other areas, while contributing to enhancement in productivity and quality of project management.

Until now, we have believed that one of advantages of cyber market is that it can virtually display and sell goods because it does not necessary maintain expensive physical shops and inventories. But, in a highly competitive environment, business model that does away with goods in stock must be modified. As we know in the case of AMAZON, leading companies already consider merchandise management as a critical success factor in their business model. That is, a solution to compete against one's competitors in a highly competitive environment is merchandise management as in the traditional retail market. Cyber market has not only past sales data but also web log data before sales data that contains information of path that customer search and purchase on cyber market as compared with traditional retail market. So if we can correctly analyze the characteristics of before sales patterns using web log data, we can better prepare for the potential customers and effectively manage inventories and merchandises. We introduce a systematic analysis method to extract useful data for merchandise management - demand forecasting, evaluating & selecting - using web mining that is the application of data mining techniques to the World Wide Web. We use various techniques of web mining such as clustering, mining association rules, mining sequential patterns.

Medical domain is very applicable for multi-agent system because medical information systems need much knowledge and close relationship with medical staff, In this paper, we describe design and implementation of an intelligent medical multi-agent platform that provides medical images'information services. This platform supports a physical environment that medical agents can be deployed following FIPA(Foundation for Intelligent Physical Agent)\`s agent management reference model. To use a variety of components on Windows, COM(Common Object Model) interfaces and XML(extensible Markup Language) for encoding ACL(Agent Communication Language) are used for multi-agent communications. Since many kinds of diverse and close relationships with medical staff) are essential, a medical staff is conceptualized as an agent and integrated with multi-agent systems. Also it provides an infrastructure applicable to share necessary knowledge between human agents and software agents in order to make intelligent medical information services easier.

This study proposes bankruptcy prediction model using fuzzy neural networks. Neural networks offer preeminent learning ability but they are often confronted with the inconsistent and unpredictable performance for noisy financial data. The existence of continuous data and large amounts of records may pose a challenging task to explicit concepts extraction from the raw data due to the huge data space determined by continuous input variables. The attempt to solve this problem is to transform each input variable in a way which may make it easier fur neural network to develop a predictive relationship. One of the methods selected for this is to map each continuous input variable to a series of overlapping fuzzy sets. Appropriately transforming each of the inputs into overlapping fuzzy membership sets provides an isomorphic mapping of the data to properly constructed membership values, and as such, no information is lost. In addition, it is easier far neural network to identify and model high-order interactions when the data is transformed in this way. Experimental results show that fuzzy neural network outperforms conventional neural network for the prediction of corporate bankruptcy.

There have been lots of researches for providing a personalized service to a customer using one-to-one marketing and collaborative filtering techniques in E-Commerce. However, there are technical difficulties for providing the recommendation of products far users, which often involve high complexity of computation. In this paper, we have presented an integrated method of classification problem solving method and constraint based configuration techniques. This method can reduce a complexity of computation by classifying a solution domain space that has a higher complexity of composition. Thereafter, we have modeled customers constraints and the components of products to configure a complete system by passing it to constraint processing module in Constraint Satisfaction Problems. Constraint-based configuration uses the constraint propagation using the constraints of buyers and the constraints among PC components to configure a proper product for a customer. We have transformed and applied vector space modeling method in the field of information retrieval to consider a customer satisfaction in addition to the CSP. Finally, we have applied our system to test data fur evaluating a customers satisfaction and performance of the proposed system.

In this paper, a detection method of contradiction between input informations is proposed when the inference is processed in rule-based systems. The proposed method is accomplished by improving the label representation and the label management scheme in a conventional ATMS(Assumption-based Truth Maintenance System). The Proposed method also can represent and process input informations having uncertainty values.

This paper is aimed at proposing a data mining-driven analysis to manage the customer defection rate in the bank. After 1997 IMF crisis, Korean banks were suffering from hard-pressed restructuring. At the heart of such restructuring effects, there was the need to manage the customer more effectively than ever. So far, many banks in Korea used to a poor management of customers without any highly-skillful techniques. In line with this argument, we propose several data mining techniques to determine more effective technique far managing customer deflection. We applied three data mining techniques such as logit model, neural network, and C5.0. Experiment data were collected from personal deposit account data of a specific bank in Korea. After experiments, we found that C5.0 showed more robust performance compared to other two techniques. On the basis of those experiment results, we proposed customer defection management policy.