I wanted to do this but never figured out where I could get access to the data. I think with setting up the inputs correctly to handle 0-5 heroes chosen per team it could work. Once you have a model you just need to rank remaining heroes by the expected win probability if they are included with the team. If you have a way to get data I would be interested.
There's a free API endpoint, you can get an API Key based on your Steam logon.
The problem is that it has been highly throttled, throwing 429 errors after just a few dozen calls. When I looked at it before it was "soft" throttled and would return data at a pretty decent rate. If I remember correctly I got something like 80 million game results downloaded in about a week.
The "ID" is the game ID, which is 570 for Dota 2. Hence the actual API endpoint is:
GET https://api.steampowered.com/IDOTA2Match_570/GetMatchHistoryBySequenceNum/v1
[1] It would be ever so nice if Steam provided daily batches in gzip files. That would be thousands of times cheaper for them to host, and much more useful for AI researchers.
