Data mining methods seek to discover unexpected and interesting regularities, called patterns, in presented data sets. However, the methods often return a collection of patterns for any data set, even a random one. Statistical significance testing can be applied in these scenarios to select the surprising patterns that do not appear as clearly in random data. As each pattern is tested for significance, a set of statistical hypotheses are considered simultaneously. The multiple comparison of several hypotheses simultaneously is called multiple hypothesis testing, and special treatment is required to adequately control the probability of falsely declaring a pattern statistically significant. However, the traditional methods for multiple hypothesis testing can not be used in data mining scenarios, because these methods do not consider the problem of varying set of hypotheses, which is inherent in data mining. This thesis provides an introduction to the problem and reviews some published work on the subject. The focus is in multiple hypothesis testing and specifically in data mining. The problems with traditional multiple hypothesis testing methods in data mining scenarios are discussed, and a solution to these problems is presented. The solution uses randomization, which involves drawing samples of random data sets and using the data mining algorithm with them. The results on the random data sets are then compared with the results on the original data set. Randomization is introduced and discussed in general, and possible randomization schemes in different data mining scenarios are presented. The solution is applied in iterative data mining and biclustering scenarios. Experiments are carried out to display the utility in these applications.
|Translated title of the contribution||Monen hypoteesin testaus tiedonlouhinnassa|
|Publication status||Published - 2012|
|MoE publication type||G5 Doctoral dissertation (article)|
- data mining
- multiple hypothesis testing
- statistical significance testing