How about a simple Bayesian classifier for legitimacy? Parameters: account open date, date of first comment, date of last comment, karma, std dev of time between comments.
Good for drawing crooked lines between accounts of a similar age, eg. there is a signature for legitimacy that appears on short (two day) and long (two year) time scales. Talk about specific thresholds seems misplaced when it can be computed from a small training sample.
Good for drawing crooked lines between accounts of a similar age, eg. there is a signature for legitimacy that appears on short (two day) and long (two year) time scales. Talk about specific thresholds seems misplaced when it can be computed from a small training sample.