Driftlon tries to learn how to predict future professional LoL-players, using the public available ranked games on their private profiles.
The data-fetcher fetches and combines data from multiple sources.
From Leaguepedia I get the list of active players. I only use those players to be pro-level who have played more than 20 games. This way I filter for only the best of the best and have a pretty harsh entry requirement. Currently only those players from the best leagues are considered. These leagues are the LCK, LEC, LCS and the LPL (leagues who continue to succeed at international events).
TrackingThePros (TTP) provides the list of soloq account names for each player. Players who are not listed by TTP are ignored.
Leaguepedia also delivers us a list of each the players soloq accounts. These lists are probably not complete, since Leaguepedia is only a wiki. An alternative to be explored could be LOLProps which was still in development at the start of this project.
The official Riot Games API provides the data for the games played. For each account of each player the games from the last 2 patches are retrieved. All data is saved to support multiple ways of preprocessing. For each game saved the number of games played, the timestamp of retrieval and the player ID is saved alongside the game data itself.
The preprocessor prepares the data and filters unnecessary data points to reduce complexity. Currently the filtering is done by the suggestions of an expert (player who was in Diamond 2). In the future I plan to do a correlation analysis and use those variables that seem logical. The filtered attributes for each match get standardized (z-score).
The most logical model would be a Recurrent Neural Network. The experiments using a LSTM and GRU cell have yet to deliver a good model. Old models have been lost, so currently I plan to rewrite the Neural Nets using PyTorch. Conventional losses could reach their limits, as the data is heavily unbalanced. An idea could be to balance by pro games played.
The ability to function well in a team is not at all measured using this approach. LoL does have some indication in the honor level, but the data is not retrievable in the API. Prominent counter examples are the twitch streamer Tyler1 who is known for his extreme behaviour towards his teammates. Another one who is on the opposite side of being extreme are those pros who never reach a higher rank than diamond but excel as a team member.
- correlation analysis
- lstm model in torch
- custom loss?
- gru model in torch