Hi board,
a problem I face in nearly every analysis I perform (field of direct marketing) is a huge imbalance of data. Let's say, a marketing campaign was performed. Out of 10'000 customers that were addressed, only 300 reacted (if we were lucky). My job now is to find a pattern, which would help to decide which customers to address in future campaigns, so that - let's say - only 5'000 customers have to be addressed and we would still get 300 responders.
In an approach using a decision tree algorithm, one would use something like "Response: True/False" as a label. Apparently, 300 out of 10'000 is a huge imbalance of data, and so far I didn't manage to get any conclusive results from a decision tree. Mostly there is just one node that says: customer does not react
One would have to use heavy over- or undersampling.
I just stumbled over a recent paper, where they have developed a new decision tree algorithm, that addresses exactly this problem. According to them, it performs better than under-/oversampling. I'll post the abstract below. The entire document can be downloaded freely from:
http://www.nd.edu/~dial/papers/SDM10.pdfIt also contains the algorithms in pseudo-code. Maybe this would be an interesting addition to Rapidminder, alongside with over- and undersampling.
We propose a new decision tree algorithm, Class Confidence Proportion Decision Tree (CCPDT), which is robust and insensitive to size of classes and generates rules which are statistically significant. In order to make decision trees robust, we begin by expressing Information Gain, the metric used in C4.5, in terms of confidence of a rule. This allows us to immediately explain why Information Gain, like confidence, results in rules which are biased towards the majority class. To overcome this bias, we introduce a new measure, Class Confidence Proportion (CCP), which forms the basis of CCPDT. To generate rules which are statistically significant we design a novel and efficient top-down and bottom-up approach which uses Fisher’s exact test to prune branches of the tree which are not statistically significant. Together these two changes yield a classifier that performs statistically better than not only traditional decision trees but also trees learned from data that has been balanced by well known sampling techniques. Our claims are confirmed through extensive experiments and comparisons against C4.5, CART, HDDT and SPARCCC.Thank you very much for considering this and best regards
Hanspeter
