Port Customer Credit Risk Evaluation System Based On Semi-supervised Learning And Information Fusion

With the developing trend of economic integration and globalization,the rapid increasements of the national economy and international trade,which not only made rapid growth in cargo transportation,but also developed ports.In order to maintain and strive for more customers,ports expanded the application scope of credit settlement.With several credit policy changes,port managers have been plagued by the customer credit risks.Customer deferments and malicious arrears seriously affected regular port operations and business processes.Traditional customer credit risk evaluation methods based on manual processes,which are difficult to fit in with the needs of port's daily management.Therefore,how to use existing resources to enhance the levels of port information construction and application and improve the ability of port customer credit risk evaluation,thus reduce or avoid losses,improve the flexibility of the port,is a crucial problem.Faced with this problem and based on practical project experience in Operation Management System and General Software Industry of Guangzhou Port Group(2008B090500244),Application Trial Project Based on Motor Ro-Ro Management System of RFID Port(2009B090300467),and Optimization Research of Integration and Scheduling of Logistics Resources(71132008),which is a key program of the national natural science foundation,this dissertation analysed the causes of port customer credit risk,integrated semi-supervised learning,active learning,information fusion,neural networks and genetic algorithms,designed and built the Internal and External Information Fusion based Port Customer Credit Evaluation System.The main research contents and results are as follows:(1)Port customer credit risk evaluation systemBased on the depth analysis of port customer credit risk,a detailed comparison of the differences between port customer credit evaluation system and the existing credit evaluation system was proposed.Existing credit risk evaluation systemsare hard to satisfy the actual requirements,and there also has insufficient researches on port customer credit risk evaluation system.Faced with these problems,this dissertation introduced external information,constituted a port customer credit risk evaluation system with a clear indicator hierarchy.(2)Label-propagation improved tri-training frameworkCause ofinsufficient sample labels in practical applications that affects the performances of text tendency classification,two semi-supervised learning algorithms:label propagation algorithm and tri-training algorithmhave been introduced.Based on the problems that label propagation algorithm can not directly deal with out-of-sample-data and tri-training algorithm is vulnerable to the initial noise problems,this dissertation proposeed a semi-supervised text classification framework LIT2based on combination of these two algorithms.(3)Active learning optimization strategy for LIT2For dealing with the bottleneck during the certain period in learning processes of LIT2,active learning has been introduced.For the different causes of learning bottlenecks inearly and late stage of training,this dissertationcorrespondinglyproposed optimization strategies based on active learning:for the early stage learning bottleneck,used membership query active learning,improved the LIT2 learning ability and enhanced the classification performance of LIT2;for the late stage of training,used pool-based active learning,which selects high training valuesamplesthereby enhanced learning and classification performance of LIT2.(4)Internal and external information fusion based port customer credit evaluationmodelBased on information fusion model and BP neural network,Internal and External Information Fusion based Port Customer Credit Evaluation Model(IEPCCM)was proposed.For the shortages of BP neural network,a Mutil-Improved BP-NN Model Construction Method(M2C)has been proposed to support forbuilding IEPCCM.The results showed that the evaluation resultsfrom IEPCCM predicted more accurately in port customer credit risk evaluation.The introduction of external environmental indicators and M2C have effectively improved the evaluation accuracy in IEPCCM model.(5)Prototyping development of internal and external information fusion based port customer credit evaluation systemBased on the previous research and requirement analysisof Semi-supervised Learning and Information Fusion Based Port Customer Credit Evaluation System(SIPCC),a system frameworkwas proposed,and a prototype system has been developed.