Multi Database Exception Pattern Mining

Data mining is a research area involved in Artificial Intelligence and Database. It has attracted much attention of many researchers and experts. Data mining has many categories, such as association analysis, clustering analysis, exception analysis and so on. Exception analysis is also called exceptional pattern mining, as one of the data mining research topics. Data in a database do not always satisfy the model, which is generated from classification or clustering analysis. Those data objects, which do not satisfy the model, are called Outliers or Exceptions. Some existing algorithms in machine learning and data mining have considered exceptions, but as noises, and excluded them out of analysis. Indeed, from the point of knowledge discovery, rare events are often more interesting and valuable than others. Examples of its applications include the detection of credit card fraud and the monitoring of criminal activities in electronic commerce. Therefore, exceptional pattern mining is an important research work.At present, the research of exceptional pattern mining focuses on single database. With the rapid development of the distributed database technologies and the computer Internet work, the multi-database system has applied in our real world. For example, a large company needs to create databases respectively for its branches distributed in different locations. And so a multi-database system is generated. For the purpose of effective decision-making, many organizations need to mine the multiple databases distributed throughout their branches. Some patterns mined from multiple databases distributed in different branches are only supported strongly by several databases. Such patterns are exceptional patterns. They reflect personality of several branches and decision makers can make special plans for these several branches. The other hand, these exceptional patterns may reflect the trend of the company in the future. Exceptional pattern mining in multi-database is worth researching.In this paper, we first introduce the concepts and main techniques in data mining and data mining in multi-database. Based on the present work, we propose some new ideas and opinions. Experimental results have proved that the proposed methods are feasible and effective. The main works in the thesis are listed as follows:(1) Based on the present work on low-voting exceptional pattern mining, an approach of mining exceptions in multi-database based on the restraint of data is proposed. This method considers users'interest. First, users provide data object of their interest. Then we select relevant data with respect to users'interest. Through analysis of every database locally to obtain local patterns and then synthesis, we can obtain exceptions. As the result of the experiment shown, our algorithm is effective and efficient.(2) A new exceptional pattern in multi-database—high-voting exceptional pattern and an approach of mining the new exceptions in multi-database is proposed. The method uses the technology of clustering rules to merge similar rules or classify the rules. Local patterns are obtained through analysis of every database locally .Then cluster the local patterns and find high-voting patterns in every class. We can obtain exceptions according to square deviation. We also evaluate our algorithm experimentally .It is validated that our algorithm is effective and efficient.(3) Some measures of pattern evaluation are analyzed from both the objective and subjective point of view respectively. Also two objective measures of interestingness to evaluate patterns in multi-database are proposed.