Research On Parallel Data Mining And Application For Business Intelligence

Many enterprises have accumulated much operational data along with using management information system. These data have not been dug out sufficiently and utilized for real commercial. For wining initiative and more commercial opportunity in furious market, business intelligence is needed to guidance business behavior and to assist decision-making. Data mining, a kernel technique in business intelligence, provides possible and technical assurance for realizing business intelligence. Confronted complicated data analysis issue, existed data mining is incapable of adapting and solving all these problems. Besides studying better technologies and theories, how to improve the efficiency of data mining is becoming the focus of academic research.This dissertation emphasizes to enhance the efficiency of data mining and backgrounds business intelligence application. Parallel processing technology and data mining technology are combined tightly in the thesis. A complete set of solution for realizing business intelligence is provided. They are from parallel data mining architecture to expression and storage for mining result pattern, to parallel neural network backward propagation algorithm and business application. The research work of the thesis has higher academic meaning and practical value.The main innovation works can be included to the following.After analyzing and inducing characteristic of PRAM, BSP and phases models, a general formula of time computing for three parallel computing models is given. This expression provides valuable reference for parallel architecture or performance estimation of parallel algorithm.A parallel data mining architecture is put forward, facing to business intelligence and having higher data mining efficiency. Two parallel computing platforms are chosen, COW and PVM, each has higher ratio of performance and cost. Pattern base management system is added, which manages result pattern. With this history patterns can be used available and the efficiency of whole data mining system can be improved.A unified storage and manipulation for three mining patterns is realized, which including association rule, classify and sequential pattern. Relational database is used to store three mining result patterns and relational algebra is used to describe store method. For manipulating result pattern convenience, SPQL(Structured Pattern Query Language) is defined and the detail implementation method is also presented. The unified storage method is a benefic and innovation explore for storing multiple result pattern.A parallel neural network TP-BP algorithm is designed, which having higher convergence speed and settling local extremum preferably. Based on higher performance RPROP algorithm, parallel TP-BP algorithm searches the space using unequal partition weight. The least extremum area is parallel searched before formal training and the second parallel operation is used to train BP network. Local extremum can be avoided, convergence process be speeded and epoch number be reduced. The evidence of experiment is that twice parallels TP-BP has better accelerate performance and applicability.Combining the project supported by Chongqing application foundation, a medicine sale trend prediction model is established and realized. In the model parallel TP-BP algorithm is adopted and carried out in the COW parallel computing environment. The effect of practical application reveals that the forecast value gained from model is accordance with real sale approximately. This fact verified the feasibility and practicality of parallel data mining architecture and parallel TP-BP algorithm presented in dissertation. It has higher reference value for enterprise management and decision-making.