I'm doing a similar task, with the difference that the analysis preference is for a Dating service. Party application to calculate the spammer offers different services... i.e. whose purpose is NOT to meet you.
Each user has some indicators (we have more than 50). Information is collected at each step of polzovatelya of finding the site. Next, we scrolled through the data statistics, and identify the clusters of users. Abnormal users "pagecity", those whose distance in n-dimensional space with more than two medians of the cluster.