Algorithm update incoming
Added 2018-05-04 17:15:50 +0000 UTCYou might have noticed that starting on April, 27th some games had their Steam Spy numbers jump a lot. This is related to the new algorithm, but not in the way you'd think. Actually, all the data on the site from April, 2nd is already calculated using the new method.
What happened is I fixed some data sources and suddenly the site started to get more information that I'm using for estimates. It caused over-inflation for some games, even though they were in the correct ballpark before.
The reason for this is that my algorithm (as all ML algorithms) doesn't really work well when some data points are missing. And when it got those data points back, it overcorrected for them, causing the inflation.
This is a very exciting problem to solve, because so far there is no singular solution to this in the machine learning community. I will try to use dumb extrapolation to fill the missing data points and then retrain the existing model.
It will take a couple of days (teaching AI takes time), but if it works, you should see the new results on Sunday evening or on Monday morning.
If it doesn't, I have some more ideas in place, of course.
On a side note, I'm talking to several data providers that might supply Steam Spy's algorithm with more food for thought. If any of those work out, I will let you guys know.
