Originally Posted by
Math_Mage
To answer your questions myself...
-Would it work? Your later qualifiers defining "work" are "would it give better teamcomps statistically?" and "would it lead to higher win rates?" I think the human factor leads to the answer "no", for both questions. If you work off tournament games, you don't have enough data to have a working model, and misleading factors like the success of one particular team with a particular strat other players can't perform as well will dominate the data. If you work off a larger sample, you introduce people who don't know how to play (which is most of the playerbase), and their success with a given champion won't reflect the theoretical matchup, but rather the relative inability of the players. Nor will the theoretical matchup be a reliable indicator of what the player USING this tool can do in that matchup.
-Would it behave similarly to human-chosen bans and picks? Not really, because humans choose champions largely based on their personal expertise, and this isn't a factor in the proposed model.