| A majority of the studies on data mining have focused on efficient algorithms for extracting patterns. In this thesis we consider the less studied issue of interpreting mined patterns. While the statistics-based interestingness measures cannot successfully explain the semantics of the patterns, we attempt to search for additional explanations for the discovered patterns, and thus to improve the effectiveness of predictive and descriptive functionalities. It is realized that explanations may not be found in the original data set, and hence additional information needs to be collected.; As a concrete example of explanation-oriented association mining, the notion of conditional association is introduced, which explicitly expresses the set of conditions under which an association occurs. The addition of conditions to the standard association rules makes it easy to understand, interpret and utilize the associations. A new method is proposed for mining conditional association rules by combining unsupervised and supervised learning algorithms. An unsupervised learning algorithm such as Apriori can be used to identify associations in the transaction table, and a supervised algorithm such as ID3, C4.5 or another inductive algorithm, modified-PRISM, can be used to find conditions in the explanation table.; Not every condition is equally explainable to the desired association. Furthermore, not every explanation table is explainable. The conditional associations are evaluated by local measures and global measures, and by regarding and disregarding the generality of conditions, respectively. Experimental results of a Web log database are presented to support the proposed methods. From the discussion of conditional association mining, we identify and study a whole framework of explanation-oriented data mining. |
