This master thesis investigates how abuse reports can be automatically filtered in order to save time, lower costs and increase safety. Abuse reports are reports that users on a website file when they encounter content they find inappropriate. These reports are then generally handled by the Customer Service who decides if the content should be removed from the website. The reports that potentially can be automatically removed are the ones that do not result in deletion of content.
The study that is presented in this report took place at the community website Stardoll. A decision tree was built to classify reports as either good or bad. The over 200 attributes that were used to train the tree contained information about the user who filed the report, the user who the report was filed against and the report itself. Unfortunately no data could be extracted from the reported content.
On unseen data the decision tree correctly removed 22% of the reports that should be removed while incorrectly removing 8% of the reports that should not. These numbers are not good enough to make it feasible to start using the system without further refinements. Such refinements are outlined in this report together with suggestions for how other automated methods can be used at Stardoll and similar communities.