Detecting varying time rates I was thinking about detecting the veracity of user ratings by examining the time-stamp. Basically I have a series of times $(t_1, t_2, \ldots, t_n)$. So assuming that ratings should normally come in at a constant rate, is there a good way to detect whether some time-stamps are unusually clumped up, because a set of consecutive fake ratings was injected over a short period of time?
 A: I'd suggest looking into the Poisson distribution. Of course, the problem is that there may be legitimate reasons for bursts of ratings: a blog article on the product causing a surge in interest and hence a surge in raters, etc, so this doesn't prove fraud.
A: If you assume a stable rate, then you can model the time between ratings as an exponential function with constant parameter. The parameter can be estimated from actual data.
From that point, I guess it is possible to predict when (according to your model) a rating would be expected in new data. Any deviation from the prediction might be interpreted as a fake rating.
I'm not sure of how I would implement the idea, but maybe you can take this approach as it seems rather simple.
A: Your series of times are simply transactions. What you want to do is "bucket" thewe transactions to create a time series e.g. 15 minute buckets or 1 hour buckets. Now that you have a time series you can develop a model that captures/explains the typical expectations. Violations of these expectataions may suggest Pulses or Temporary level shifts or local time trends. The time series may be impacted by day-of-the-week or known events (an ad for example) . Bringing in these possible predictor/exogenous variables can then reduce the number of unusal values providing clarity to the truly unusual.
