DraftRec: Personalized Draft Recommendation for Winning in Multi-Player Online Battle Arena Game (WWW' 2022): Official Project Webpage
This repository provides the official PyTorch implementation of the corresponding paper: DraftRec
Authors: Hojoon Lee*, Dongyoon Hwang*, Hyunseung Kim, Byungkun Lee, and Jaegul Choo,
Affiliation: KAIST AI
Here we provide a detailed description about our manually constructed League of Legends dataset, which was collected by utilizing the publicly accessible API endpoint provided by Riot Games (https://developer.riotgames.com/). Our League of Legends dataset is composed of 279,893 top 0.1% ranked players' matches. In order to guarantee the quality of our dataset, we strictly restricted the tier (i.e., rank) of the players to Challenger, GrandMaster, Master, Diamond-1, and Diamond-2 ranked players.
Our dataset is mainly divided into two files : the match dataset (e.g., train.csv, val.csv, test.csv) and the user history dataset (e.g., user history array). The match dataset contains information regarding the userID and the match history index respective to the user's personal match history. The user history dataset contains detailed information about each user's past matches.
- Match dataset: Each row in the match dataset (i.e., match_id) indicates a separate match in the dataset. Each column in the match dataset (i.e., user1, user2, ..., user10) is listed according to the champion selection order. For example, user1 indicates the 1st user who selects a champion and user9 indicates the 9th user who selects a champion. Furthermore, each column includes the (i) user_id information and the (ii) history_index information.
- (i) user_id : Information about which user has participated in that particular match.
- (ii) history_index : Information about the respective match history index respective to the user's personal match history.
- Example : As shown in the figure above, for match_id =2 , the user_id and his/her respective match history index for user9 (the 9th user who selects a champion in the match) is (5,31). Thus, the 9th user who selected a champion in match_id =2 has user_id=5 and match_id=2 is recorded as the 31st match within user_id=5's match history.
- User history data : The user history data contains detailed information regarding each match within the user's personal match history. It includes information about the champion selected by the user, designated role and team, banned champions in the match, information about whether the user won or lost the match, and specific personal match performance statistics (e.g., total gold earned, total damage, etc).
- Data structure : The user_history_array is formed as a tensor with shape (number of users) x (number of features) x (max user history length). There are a total of 62,466 users, 57 different type of features (detailed information regarding the semantic of the features can be found in /data/lol/categorical_ids.json), and the max user history length is 1145.
More detailed descriptions considering the dataset statistics are provided in EDA.ipynb.
The pre-trained models can be found below.
| Model checkpoint | ACC | HR@10 |
|---|---|---|
| DraftRec (hidden_dim=64, seq_len=10) | 54.6 | 88.0 |
| DraftRec (hidden_dim=64, seq_len=50) | 55.2 | 89.9 |
| DraftRec (hidden_dim=128, seq_len=10) | 54.8 | 88.7 |
| DraftRec (hidden_dim=128, seq_len=50) | 55.0 | 89.8 |
After investigating through the dataset to publish our code, we found that several matches within the dataset were preprocessed incorrectly. Thus, we discovered a minor decreasement in accuracy and a minor increasement in HR@10 compared to the results of our paper.
Our code can run on a single GPU or on multi-GPUs. See requirements.txt for all prerequisites, and you can also install them using the following command.
pip install -r requirements.txt
To train the draftrec model with a single GPU, try the following command:
python main.py --template draftrec --dataset_path [DATASET_PATH]
To train the draftrec model with multiple GPUs, try the following command:
python main.py --template draftrec --dataset_path [DATASET_PATH] --use_parallel true
@article{lee2022draftrec,
title={DraftRec: Personalized Draft Recommendation for Winning in Multi-Player Online Battle Arena Games},
author={Hojoon Lee and Dongyoon Hwang and Hyunseung Kim and Byungkun Lee and Jaegul Choo},
booktitle = {Proceedings of the Web Conference 2022},
series = {WWW '22},
year={2022}
}
This is not an official Riot Games product.