Hey I'm a 4.2k player and a developer. I like your app's interface more than dotapicker.com. I noticed both had same suggestions for a given set of heroes. Is the source of data for both of these apps same?
My guess is we've landed on similar algorithms to DotaPicker, it's usually the set of matches we're using that makes the difference so I'm surprised the suggestions were the same.
Feel free to add me on Steam if you'd like: http://steamcommunity.com/id/sklink
I ask because this always seems the biggest weakness of all the existing picking tools. A 1k player and a 5k player throwing the same lineup into a picking tool should ideally get dramatically different suggestions, both because of their own skill and the skill of their teammates.
114 heroes that all influence each other when they play together so if you're looking at matchups for two heroes that don't get played frequently the accuracy is already limited.
That said, my strength is in front end, not statistics. Getting a Normal and High split might be fine as is.
That would limit the available match pool, granted. It depends on how many thousand games you need to make predictions, I suppose. I was able to pull some interesting data a while back from the 2k pool but I wasn't trying to solve as hard a problem as a global hero picker.
I would have to collect my own data and poll their API with each match id to get the MMR estimate.
Although that works in theory, their API rate limit is 3/s and last time I checked there were about 1M matches / day (~12/s) so we couldn't keep up.