LOL Player Ratings

Inspiration

Followed LOL Esports for many years and works as a data scientist. I have a lot of experience with rating models so this was an interesting challenge.

What it does

Calculates ratings using historical match-player-data and exposing them using API endpoints.

How we built it

Rating Model

• A separate libary for optimizing player-based-rating models was created for this usecase (https://github.com/Hiderdk/player-performance-ratings)
• Player-and-team-stats such as kills, deaths,assists, game_length, champion compositions, damage was adjusted for context and used to calculate ratings.
• The weights were optimized for predictive accuracy (who wins next match).
• To address regional rating bias's (weak isolated regions tend to be have inflated ratings in rating-models), a new player who enters his first match receives a different rating if he plays an LCK match compared to a new LCS Academy player.
• If teams/players from a league outperforms in matches against teams from another league, it will slightly boost the rating of all other players within the league. This ensures that a leagues strength dynamically can change over time

Post Rating Processing

• Ratings transformed into API-ready data format and uploaded to S3 (see https://github.com/Hiderdk/lol-aws-global-rankings/blob/main/src/tournament_rankings_storer.py and https://github.com/Hiderdk/lol-aws-global-rankings/blob/main/src/global_rankings_storer.py )

• Lambda Function loads S3 file in response to GET Request body and returns new body given query and path parameters.

• API Gateway exposes API endpoints.

Challenges we ran into

• Data quality was incosistent. E.g. some games not mapping to tournaments, players not mapping to their names. Incorrect playercounts per game.

• Accurate relative strenghts between different leagues/regions. LFL/Prime definitely overrated. A lot of work has been to try to address it, and the attempts have resulted in noticeable improvements, but clearly room for more.

• Experimentation with ingame state data (every 10 second) --> Took a lot of storage to work with this. Attempts at making win-probability-added metrics did not lead to improvement in predictive accuracy.

Accomplishments that we're proud of

• Generic framework for optimizing any type of rating model was created to support the project.
• Optimization for regional ratings resulted in significant accuracy gains.
• Player performances adjusted for context (e.g. champion composition).

What we learned

• Gained experience with AWS ecosystem.
• Teams winning matches in short game-time is better than in drawn out.

What's next for LOL Player Ratings

• Utilize ingame-state data better. One could imagine evaluating players to measure whether their contribution in a gank/teamfight improved win-probabilities.
• Continue improving regional ratings (still some flaws in the rating values between teams from different leagues).

Built With

• api
• gateway
• lambda
• s3