This thesis presents an efficient method for mining quantitative association rules for finitely discrete quantitative data. Our main contribution is a new view on what constitutes an unexpected relation, a view that is believed to be more accurate than previous models. To ensure the statistical significance of the results, a non-parametric exact test is developed that does not require data to be normally distributed, works with sample sizes as low as a single sample and has a run time proportional to the sample size. In addition to mining rules, we show how memory-based reasoning can be used to learn and to predict the users’ interests in rules. The method we suggest is compared to two other popular prediction methods, a Naive Bayes classifier and a neural network, and appears to be superior to both, both in terms of learning quickly and being able to reach a high prediction accuracy.

https://www.nada.kth.se/utbildning/grukth/exjobb/rapportlistor/2009/rapporter09/bennedich_max_09007.pdf