Expanding ARAM Predictions

I recently managed to achieve a 67% success rate in predicting the outcomes of ARAM matches. This was done on a dataset from North American players of League of Legends. Since then, I upgraded my data storage and crawling mechanisms which has allowed me to collect data from other regions. Since Riot’s API limits are on a per-region basis, I was able to collect this data in parallel. I now have data on tens of thousands of matches from each new region: Korea, EU West, and EU North.

My original 67% model was trained on matches from patch 7.11. Since then, patch 7.12 was released and almost all of the new ARAM matches I collected where on that patch. Out of curiosity, I wanted to see how well my previously learned model would perform on the new data. Surprisingly, it worked better than expected, achieving the same 67% accuracy.

Test results from a model trained on Korean players on patch 7.12

One hypothesis in my last post was that Korean players would produce better training data since they’re known as being more consistent. To my surprise, this Korean-trained, Korean-tested model performed about the same as its North American counterpart. I also attained about a 67% accuracy on my EU North and EU West datasets. I’m still collecting more data for training (because it’s so easy to do now), but I don’t expect it to move the models accuracy much.

I am still working on gaining a deeper understanding of champions and getting humans to review the mistakes the model made to learn new insights that could lead to better accuracy.