Abstract

Riot’s API does not provide the position that players selected during each match. Instead, they only provide the role and lane selections of each player in each match. Ideally, one could infer the position from the role and lane. But it is not a clear-cut, one-to-one mapping, with many possible combinations of the role and lane resulting in unknown positions 1. To overcome this, we estimate positions for players in our main dataset by building a classifier that relies on different items and spells utilized by players during a match, reusing match data and methodology from Lee and Ramler [2]. Their dataset, which we refer to as the L&R dataset, contains nearly 5 million matches from the North American Server over a period of 8 months across multiple match queues including 112,301 ranked team matches and 2,044,068 ranked solo matches. Since we are interested in individual player behavior in matches and each match consists of 10 players (5 players in each team), each match corresponds to 10 units of analysis which we call player match instances. Additionally, for each player in every match, this dataset contains the recommended position for the employed champion based on the community metagame at the time. These positions had been scraped from the LoL metagame website https://champion. gg/. We randomly select 20,000 player-match instances from the ranked team queue data and train a one-vs-one SVM model to predict the positions the players had chosen. Following Lee and Ramler [2], we use the set of items and spells that players had acquired by the end of the game as our independent variables with the metagame positions as …

Date

October 18, 2025

Authors

JULIE JIANG, DANAJA MALDENIYA, KRISTINA LERMAN, EMILIO FERRARA