Randomization algorithms for assessing the significance of data mining results

Aalto University, P.O. Box 11000, FI-00076 Aalto www.aalto.fi Author Markus Ojala Name of the doctoral dissertation Randomization Algorithms for Assessing the Significance of Data Mining Results Publisher School of Science Unit Department of Information and Computer Science Series Aalto University publication series DOCTORAL DISSERTATIONS 99/2011 Field of research Computer and Information Science Manuscript submitted 12 April 2011 Manuscript revised Date of the defence 12 November 2011 Language English Monograph Article dissertation (summary + original articles) Abstract Data mining is an interdisciplinary research area that develops general methods for finding interesting and useful knowledge from large collections of data. This thesis addresses from the computational point of view the problem of assessing whether the obtained data mining results are merely random artefacts in the data or something more interesting. In randomization based significance testing, a result is compared with the results obtained on randomized data. The randomized data are assumed to share some basic properties with the original data. To apply the randomization approach, the first step is to define these properties. The next step is to develop algorithms that can produce such randomizations. Results on the real data that clearly differ from the results on the randomized data are not directly explained by the studied properties of the data. In this thesis, new randomization methods are developed for four specific data mining scenarios. First, randomizing matrix data while preserving the distributions of values in rows and columns is studied. Next, a general randomization approach is introduced for iterative data mining. Randomization in multi-relational databases is also considered. Finally, a simple permutation method is given for assessing whether dependencies between features are exploited in classification. The properties of the new randomization methods are analyzed theoretically. Extensive experiments are performed on real and artificial datasets. The randomization methods introduced in this thesis are useful in various data mining applications. The methods work well on different types of data, are easy to use, and provide meaningful information to further improve and understand the data mining results.Data mining is an interdisciplinary research area that develops general methods for finding interesting and useful knowledge from large collections of data. This thesis addresses from the computational point of view the problem of assessing whether the obtained data mining results are merely random artefacts in the data or something more interesting. In randomization based significance testing, a result is compared with the results obtained on randomized data. The randomized data are assumed to share some basic properties with the original data. To apply the randomization approach, the first step is to define these properties. The next step is to develop algorithms that can produce such randomizations. Results on the real data that clearly differ from the results on the randomized data are not directly explained by the studied properties of the data. In this thesis, new randomization methods are developed for four specific data mining scenarios. First, randomizing matrix data while preserving the distributions of values in rows and columns is studied. Next, a general randomization approach is introduced for iterative data mining. Randomization in multi-relational databases is also considered. Finally, a simple permutation method is given for assessing whether dependencies between features are exploited in classification. The properties of the new randomization methods are analyzed theoretically. Extensive experiments are performed on real and artificial datasets. The randomization methods introduced in this thesis are useful in various data mining applications. The methods work well on different types of data, are easy to use, and provide meaningful information to further improve and understand the data mining results.