Computation isn’t likely prohibitive when using the bedrock of predictive models, linear regression, especially w some optimization. Could also vary observation time so you only need a prediction once every ten minutes or whatever
The real issue is the cost of false positive detection of cheating is negligible since the vast majority of positives are probably true positives—it’s the cost of doing anti-cheat business (minimal)
But yes cheats would be modified to just below thresholds of detection
I think this might be in reply to my first comment about scaling? If so, I just want to clarify that I was thinking more along the lines of scaling the customer service/ban appeal side rather than infrastructure.
If, for example, every ban had a component of someone at Valve taking 10 seconds to review in-game stats at the time of the ban, and then making a determination of whether or not those stats seem reasonably non-cheater-ish (pretty hard policy question in itself), the process would slow to a crawl.
The real issue is the cost of false positive detection of cheating is negligible since the vast majority of positives are probably true positives—it’s the cost of doing anti-cheat business (minimal)
But yes cheats would be modified to just below thresholds of detection