League of Legends Champion Predictor

Zhonghan Li zhonghanli2019@u.northwestern.edu | James Xie jamesxie2019@u.northwestern.edu

League of Legends is one of the most popular games on the planet that even with an incredibly complex gameplay system has managed to captivate millions of players. However, one of the hardest decisions players often face is not during the game, but before- choosing which champion (think street-fighter selection screen) to play. There are over 100 champions, each with their own unique set of “rules” that apply to them in-game, making it tough for a human to know whether or not they’d be good at the specific champion. A champion recommender could have widespread impact on players trying new champions and achieving success within the game. Our task is to make a recommender that players can use by entering their own in-game name. The program will generate a list of which champions the player may have good performance on or start to play often(an indicator of enjoying the playstyle of that champion). A successful recommending tool would garner popularity, resulting in millions of people using it to determine their next champion to try.

The most interesting part of the problem proposed, however, is not the results of a successful tool but perhaps on how it is trained. Looking into what attributes are effective in predicting the best champion success brings to light a bunch of interesting facts about the community. For example, the percentage of accuracy of running nearest neighbor on our dataset peaks at around gold elo and gets skewed as we progress above the bracket. This seems to makes sense because less skilled players tend to have a smaller champion pool to select from due to them only looking to play powerful champions, which would naturally raise the accuracy of any ML algorithm for lower elo players. As we test other methodologies on the same datasets, we eventually saw that that hypothesis was correct. Moreover, as we shaved down our attributes from over 40 to just under 15, we found that the most impactful attribute was not winrate or role played, but actually champion mastery! This means that although the community seems to be extremely obsessed with winning, enjoying playing a champion is ultimately more important to most players than winning a game. As seen in the diagram below, our accuracy peaks around the gold elo, where players follow the meta most successfully.

The conclusion of the project is something to be expected- with over 140 champions to recommend, it’s extremely tough for a computer algorithm to predict which champion a player should play next with any amount of accuracy – a flat guess yields under 1% success rate. However, with a 3.13% accuracy at the gold elo, a nearest neighbor champion predictor that outputs a list of 10 possible champions a player may enjoy would yield almost a 30% success rate, which is something to be excited about.