Temporal Data Mining And Its Applications

Due to the large scale and compound contents of modern temporal database, concerning more on the flexibility of the mining tools, most recent mining tools are satisfied with some experiential results. not going to unveil the rules behind these results. After surveying over international and national researches. I find temporal data mining is just at the primary stage. In order to improve the accuracy and reliance of mining techniques, I have done lots of researches, and the main innovations are listed as follows.Firstly, this dissertation initiate the conception of durative event series, and build a formal framework for the temporal data. by which the time structure are kept intact, durative event series has two crucial attributes: (1) Event must not finish instantaneously but keep happening status in a certain time span; (2) The individual is the basic unit in durative event series, and relative variables are recorded as additional information.Secondly, this dissertation advanced a sort of mining tools for temporal data. In rule mining, we re-integrate the conception of association rule, and propose one kind of temporal rule mining tools; In clustering analysis, we invite the identity relationship from Roughset Theory, and advance a creative clustering technique, which is void of the problem of curse of dimensionality; And in model mining, under the mind of hazard model. a model mining technique is developed for further research.Finally, we accomplished three real data analyses. Based on the cell phone consumption dataset. we conduct temporal ruling mining and roughset clustering; and on the compound database of ST events of A-share listed company of China stock. consisting of 2000-2005 yearly financial reports and contemporaneous market performances, the model mining brought up in this dissertation completes the analysis task efficiently.Main results are concluded as followed. (1) Durative event series is currently thebest data format for temporal data mining, which supports a unified data platform for both methodological research and algorithm designment and comparison: (2) Temporal rule mining eclipses static rule mining, for it can extract the sequential relation and parallel relation among transactions; (3) From the view of knowledge classification, identity relationship gives us a new definition of cluster, which contributes to the idea of roughset clustering in high dimensional database; (4) As to temporal data composed of compound contents, model mining excels any other mining tools, and also illuminates the idea of compound temporal data mining.