Before the match starts, this model picks the winning lineup: joint player-proficiency and team-congruency modeling reaches 61.33% accuracy
Synopsis
The outcome of a temporary team is often foreshadowed before play begins, yet a single player rating says neither whether that player fits the current role nor whether the five assignments complement one another. This work puts both questions into one deep-learning model: it first reads each player's earlier matches against the current champion and role, then compares the two ten-player lineups as a whole. On a chronologically split validation set, the model predicts the winner before play at 61.33% accuracy, above the TrueSkill, CatBoost, GRU and DraftRec comparisons.
Interpretation
Knowing only the champions and role assignments fixed before play, the model judges which side will win more accurately than existing ratings and lineup measures. Earlier approaches either summarized a player with a single ability rating such as TrueSkill or described a lineup with hand-built counts and win rates, and both compress how a player differs across roles and champions. On the chronologically ordered validation set the model reaches 61.33%±0.05% accuracy and 0.6490±0.0008 binary cross-entropy, against 58.31%±0.06% and 0.6689±0.0007 for DraftRec, the strongest comparison, and 53.88% for TrueSkill.
The model first reinterprets each player's history in light of the current assignment, then evaluates all ten representations jointly instead of averaging separately scored players. DraftRec builds a player representation independently of the current champion and role and then average-pools within each team; this model lets the current champion, role and patch take part in attention over the history and uses a learned composition token to gather all ten positions on both sides. Ablations show accuracy falling to 55.37%±0.03% when histories are removed and only the current composition is used, and to 58.19%±0.07% when the joint composition representation is replaced by independent evaluation with mean aggregation, both below the full model's 61.33%±0.05%.
Model capacity and history length both saturate, and the principal configuration reaches near-larger-model quality with about 51.9 million parameters and 32 prior appearances per player. Accuracy rises steadily as the model grows from about 0.92 million parameters to the principal configuration, then stays essentially flat and does not improve cross-entropy when scaled to 173 million parameters. Extending history from 8 to 32 appearances gives a substantial gain while doubling to 64 gives only a marginal one; the principal configuration scores 61.33%±0.05% and the 173-million-parameter version 61.36%±0.06%.
Perspective
The evaluation covers the moment in League of Legends solo-queue matches when both sides' roles and champions are fixed but play has not begun, and inputs include only records preceding the target match, with nothing generated during play. The author also releases processed, anonymized data and code for replication and further research.
Proficiency and congruency are relative quantities implied by specialized predictive models rather than directly observed team attributes; match outcomes also include execution and chance, so prediction errors do not necessarily mean the initial composition was misjudged. The analysis is observational, identifying associations that generalize chronologically but not the causal effect of changing a player's assignment.