DOTA IDEAS A DOTA 2 INTERACTIVE FORUM

By gathering metadata that Dota servers already have and asking voluntary classification of their teammates after a game, ranking them from best to worst, we train a neural network or decision tree model to do it automatically with high accuracy.
By using that model, the Dota ranked games could reserve a small part of the win/loss MMR variation to punish or reward the players according to how they played. That way, a player that was good enough can be rewarded even when he loses the game. I suggest that small variation to be fixed as 3 MMR points at first.
As for which metadata to use, I suggest to include the following, but not restricted to that:
- Role
- Average distance to teammates
- Participation in battles
- Everything that was typed
- Everything that was said
- Chosen hero
- All metadata from the game summary (with timestamps)
- Revelations from invisible, wards destroyed and used
- Numeric values should always be taken with relation to the total amount in the game in order to make long games comparable to short games.

As for how the MMR variation should work, I suggest the following:
The best in each team gets + 3 MMR. The worst in each team gets - 3 MMR. The second best gets + 2 MMR, the second worst gets - 2 MMR. The third and last one, doesn't get a MMR variation.