Identifying Global Exceptional Patterns In Multi-database Mining
Chengqi Zhang, Meiling Liu, Wenlong Nie, Shichao Zhang · UTS ePRESS (University of Technology Sydney) · 2004
Abstract- In multi-database mining, there can be many localpatterns (frequent itemsets or association rules) in each database.At the end of multi-database mining, it is necessary to analyzethese local patterns to gain global patterns, when putting all thedata from the databases into a single dataset can destroy impor-tant information that reflect the distribution of global patterns.This paper develops an algorithm for synthesizing local patternsin multi-database is proposed. This approach is particularly fitto find potentially useful exceptions. The proposed method hasbeen evaluated experimentally. The experimental results haveshown that this method is efficient and appropriate to identifyingexceptional patterns. jlldex Terms- multi-database mining; local pattern evaluation;local pattern; global patteru; exceptional patternL INTRODUCTIONWith the increasing development and application of distributeddatabase technique and computer network, there exist many distributeddatabases in a business or financial organization. For example, a largecompany has many subsidiary companies, and each subsidiary com-pany has its own database, all of the databases from each subsidiarycompany are relevant or irrelevant in logic, but they are distributed indifferent places. Different subsidiary company has different functionsin helping the head company to make decisions. To make decisionsfor the development of company, the decision maker of the head com-pany needs to know every database's interesting pattern or regulationand then synthetically evaluate these local patterns to generate globalpatterns.It would appear to be unrealistic to collect data from differentbranches for centralized processing because of the potentially volumeof data [20]. For example, different branches of Wal-Mart collect 20million transactions per day. This is more than the rate at which datacan feasibly be collected and analyzed by using today's computingpower.On the other hand, because of data privacy and related issues, it ispossible that some databases of an organization can share their asso-ciation rules but not their original data. Therefore, mining associationrules from different databases and forwarding the rules (rather thanthe original raw data) to the central company headquarter provides afeasible way dealing with multiple database problems [19].However, current data mining researches focus on mining in mono-database, but mono-database mining is different from multi-databasemining because of their different data structure. So we need to comeup with other solutions to analyze the data in multi-databases insteadof using the technique in mono-database mining. This paper mainlydiscusses the pattern evaluation process at the end of data miningprocess and presents a method for identifying exceptional patterns.