2022 International Conference on Machine Learning and Knowledge Engineering (MLKE)

A machine learning approach to predict the result of

2022 International Conference on Machine Learning and Knowledge Engineering (MLKE) | 978-1-6654-9567-7/22/$31.00 ©2022 IEEE | DOI: 10.1109/MLKE55170.2022.00013

League of Legends
Qiyuan Shen
Data Science, Duke Kunshan University, Kunshan, Jiangsu Province, 215300
* Corresponding author: [email protected]

Abstract—Nowadays, the MOBA game is the game type with Then some recent works in the application of machine
the most audiences and players around the world. Recently, the learning in League of Legends are checked for reference. In
League of Legends has become an official sport as an e-sport Xia, Wang, and Zhou’s research, they mainly focus on another
among 37 events in the 2022 Asia Games held in Hangzhou. As popular MOBA game: DOTA 2 (Defense of the Ancients 2).
the development in the e-sport, analytical skills are also involved They choose R to analyze the data and they find out that the
in this field. The topic of this research is to use the machine tactical awareness affects the multiplayer killing indicator and
learning approach to analyze the data of the League of Legends the result of the game more than the operations [2]. Another
and make a prediction about the result of the game. In this project also research data of the game DOTA 2. They present
research, the method of machine learning is applied to the
live win predictions in MOBA games by using standard
dataset which records the first 10 minutes in diamond-ranked
games. Several popular machine learning (AdaBoost,
machine learning models [3]. The third research mainly
GradientBoost, RandomForest, ExtraTree, SVM, Naïve Bayes, compares the results of 3 models: Linear Regression, Neural
KNN, LogisticRegression, and DecisionTree) are applied to test Network, and LSTM [4]. Another research mainly uses 3
the performance by cross-validation. Then several algorithms learning models Naïve Bayes classifier, Logistic Regression,
that outperform others are selected to make a voting classifier to and Gradient Boost. And their result shows that the prediction
predict the game result. The accuracy of the voting classifier is oscillates depending on players’ operation skills [6]. Then
72.68%. researches in League of Legends was cited to see the
commonly used method in League of Legends’ field, in a
Keywords-League of Legends; machine learning; Voting recent research, the researchers mainly use a deep neural
classifier; MOBA network based on the players’ previous experience on the
champions they choose [5]. While another research mainly
I.INTRODUCTION focuses on the pre-game factors’ influence on the result of the
Compared to the last decade, the views towards games have game [7]. There is one research that carries recurrent neural
changed a lot in these years. Parents hold an open opinion networks on the data about League of Legends. And their
towards the games rather than the spirit opium which is results show that the accuracy of the model trained by 0-5
commonly believed in the past. As the development in the minutes data is worse than the data from 20-25 minutes [8]. In
game industry, there are more and more types of games to Lin’s research, the researcher conducts several popular
choose like: RPG (Role-Playing Game), FPS (First-person machine learning models which give some inspiration to this
shooting game), strategy game, and so on. And the game research [9].
competition has also become an official sports event which is
called e-sport. And the e-sport will also be one of the 37 sports II.METHODS
played in the 2022 Asia Games held in Hangzhou. Among all A. Game map introduction
these e-sports, the MOBA (Multiplayer Online Battle Arena)
game is the most popular one. According to Mora-Cantallops Figure 1 below is the classical map for the MOBA game
and Sicilia’s article, about 205 million people participate in including League of Legends. There are three lanes on the map
the MOBA games (2018). The research topic League of which are the top, middle, bottom lane. Each lane is separated
Legends is the hottest MOBA game. In 2013, the League of into 2 parts for each team. Each lane has 2 defend towers. At
Legends Championship had drawn 27 million online audiences the end of each lane, it is the base of each team and consists of
to watch the competition. Correspondingly, money and 3 defend towers and a Nexus. Between each lane, it is the
attention have been spent on players’ performance during the Jungle where there are jungle minions in it which can provide
game. Data analysis by machine learning is one of the topics of experience and gold. There are also some elite monsters in the
this research. jungle. They are dragons and heralds which can bring great
benefits to the team who kills them.
The research topic is the application of machine learning
skills in competitive games: predicting the result of the
competition by game indicators and analyzing which indicator
affects the game most. The research provides a relatively
objective method to analyze and rate the game to avoid
subjective influence. In addition, this research can also be
applied to other e-sport games and sports games to help players
and athletes improve their performance.

978-1-6654-9567-7/22/$31.00 ©2022 IEEE 38

DOI 10.1109/MLKE55170.2022.00013
Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 B. Dataset introduction
Normally, one game of League of Legends will last for
30-45 minutes. To ensure the accuracy of the model and make
the prediction available to be applied in e-sport games, the
dataset used in this research records the data of the first 10
minutes of different games. All the games in the dataset are
diamond-ranked. Moreover, to ensure enough data to train,
there are nearly 10000 data in the set. In the dataset, 40
columns record the detailed game data in the first 10 minutes.
C. data processing and visualization
The first thing in the data processing part is to clean the
missing and duplicate values. Fortunately, there are no data
missing or duplicating in this dataset. Because there is so much
data in the dataset. Choosing proper features is especially
important in this research. Table 1 below shows some features
that are useful in this project.
Figure 1. The general map for the MOBA games

TABLE I. SOME USEFUL INFORMATION IN THE DATASET

blueWins The final goal for the research. The feature blue wins is used as the determent (1 means blue wins, 0 means red wins).

blueWardsPlaced/ The warding totems are an important item in League of Legends. When it is placed in an invisible area, it can reveal an
blueWardsDestroied area for the team that places the warding totems. This feature shows the wards placed/destroyed by the blue team.
blueKills/
blueDeath/ These 3 features are the basic data for the game which record the kills and death between different teams.
blueAssists
blueEliteMonsters/ blueDragons/ These 3 features record the data about the elite monsters: Dragon and Heralds which can give great benefit to the team
blueHeralds killing them in a certain period.
This feature records the defending tower destroyed by the blue team which is also an important factor deciding the
blueTowerDestroyed
result of the game.
blueTotalGold/ blueAvgLevel/
The total gold and experience earned by the team can represent the development of the champion in that period.
blueTotalExperience
blueTotalMinionsKilled/
This feature is closely related to the gold and experience earned by the team.
blueTotalJungleMinionsKilled

Different teams have different game paces. Moreover, some blueWardsDestroyed, blueTotalGold, blueTotalExperience,
champions can dominant the game at the early stage which blueAvgLevel, blueTotalJungleMinionsKilled,
may affect the accuracy of the model. So, several features are blueTotalMinionsKilled, redKills, redAssists, redWardsPlaced,
standardized to make it easier to compare their differences. The redWardsDestroyed, redTotalGold, redTotalExperience,
main standardized function is shown below: redAvgLevel, redTotalJungleMinionsKilled,
redTotalMinionsKilled. Because the features of elite monsters,
ܾ݈‫ ܽݐܽ݀݁ݑ‬െ ‫ܽݐܽ݀݀݁ݎ‬ dragons, heralds and first blood are so close to each other. So,
–ƒ†ƒ”†‹œ‡†”ƒ–‹‘ ൌ they haven’t been standardized it.
ܾ݈‫ ܽݐܽ݀݁ݑ‬൅ ‫ܽݐܽ݀݀݁ݎ‬
In this research, this function is used to standardize the
Then the processed data is visualized in the following
following features: blueKills, blueAssists, blueWardsPlaced,
figures.

39

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 Figure 2. The distribution of features

Figure 3. The distribution of features

Figure 4. The distribution of features

40

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 Figure 5. The distribution of features

Figure 6. The distribution of features

Figure 7. The distribution of features

41

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 Figure 8. The distribution of features

Figure 9. The distribution of features

From the figures above, the conclusion is that the level, are some features less related to the result of the game which
experience, gold and kills ratio have a great influence on the may decrease the accuracy of the model. So, the first thing
result of the game. While for the none-standardized feature, the before the machine learning part is dimensionality reduction.
dragons, elite monsters, and first blood can influence the result Firstly, a heatmap is made based on the correlation map
of the game to some extent. between features and features shown below.
D. Machine learning
However, there are still 15 columns in the dataset which is
quite a lot of data. In addition, from the visualizing result, there

42

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 Figure 10. Heatmap of the correlation matrix among all the features

From the heat map above, the feature gold ratio, experience several models will be chosen to make a voting classifier to
ratio, kills ratio, level ratio, and minions ratios have a stronger make the result more precise. The result will be shown in the
relationship with the feature blue wins. Then, the dataset is next part.
separated into train dataset and test dataset for cross-validation.
After that, several popular machine learning models (AdaBoost, III.RESULT AND DISCUSSION
GradientBoost, RandomForest, ExtraTree, SVM, Naïve Bayes, After training the model, the accuracy of the training is
KNN, LogisticRegression, and DecisionTree) were chosen to shown below as well as the confusion matrix.
test the accuracy of the training result. Based on the result,

Figure 11. The training result of the chosen algorithm

43

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 Figure 12. The training result of the chosen algorithm

Figure 13. The training result of the chosen algorithm

From the figure above, KNN, Naïve Bayes, GradientBoost, test and train accuracy of the whole model which is 0.72638
and AdaBoost were chosen to make the Voting classifier. and 0.79831 respectively.
Moreover, to make the training result of the algorithm more
precise, RandomForest is added into the Voting classifier. The However, there are lots of champions in League of Legends
training and test result is shown below. who are weak in the early stage and dominant in the further
stage. So, if only the first 10 minutes of data the result is
analyzed, some errors may occur. So, this is the reason why the
testing and training results can’t be more precise. In addition,
since the champions will not be fixed in one lane in League of
Legends, this attribute will confuse the positions of opponents.
Although the data about the pick and ban in diamond ranked
games were used, based on this property, this data can’t be
used properly. So, there is no choice but to drop these features.
IV.CONCLUSION
The test accuracy of the voting classifier is 0.72638 which
has spaces for improvement. For further research, a wider
dataset that clearly records the data with every stage of the
game will be applied in the research. A comparison is
Figure 14. Training and test accuracy of the Voting Classifier conducted on every stage to determine which factor plays the
most important role in that stage. To analyze the relationship
Judging from the confusion matrix about the voting between different champions, the dataset will dig deep into the
classifier and the 9 algorithms above, it can conclude that the games which record the performance when facing another
algorithm related to decision trees always performs well in the champion. An algorithm will then be developed to provide
training dataset, but not so good in the test dataset which advice at any stage using the current data of the game.
results in the situation of overfitting. In the algorithms related
to decision trees, the random forest performs best. While in ACKNOWLEDGMENT
other algorithms the Naïve Bayes performs best in the test Firstly, I am extremely grateful to professors in my
dataset. The Voting Classifier in the last step does improve the university who give me helpful advice at the algorithm

44

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.
 selecting stage and help me solve some coding problems.
Secondly, I am very grateful to our teaching assistant Kunyu
An who helps me decide the topic of this research: a machine
learning approach to analyze League of Legends. Last but not
least, I must also thank my family members and friends who
cheer me up when I met troubles and problems without their
help I will not complete this research.
REFERENCES
[1] Mora-Cantallops, M., & Sicilia, M.-Á. (2018). MOBA games: A
literature review. Entertainment Computing, 26, 128–138.
https://doi.org/10.1016/j.entcom.2018.02.005
[2] Xia, B., Wang, H., & Zhou, R. (2019). What contributes to success in
MOBA games? an empirical study of defense of the ancients 2. Games
and Culture, 14(5), 498-522. https://doi.org/10.1177/1555412017710599
[3] Hodge, V. J., Devlin, S., Sephton, N., Block, F., Cowling, P. I., &
Drachen, A. (2021). Win prediction in multiplayer esports: Live
professional match prediction. IEEE Transactions on Games, 13(4),
368-379. https://doi.org/10.1109/TG.2019.2948469
[4] Akhmedov, K., & Phan, A. H. (2021, June 3). Machine learning models
for DOTA 2 outcomes prediction. arXiv.org. Retrieved January 9, 2022,
from https://arxiv.org/abs/2106.01782v1
[5] Do, T. D., Wang, S. I., Yu, D. S., McMillian, M. G., & McMahan, R. P.
(2021). Using machine learning to predict game outcomes based on
player-champion experience in League of Legends. The 16th
International Conference on the Foundations of Digital Games (FDG)
2021. https://doi.org/10.1145/3472538.3472579
[6] Semenov, A., Romov, P., Korolev, S., Yashkov, D., & Neklyudov, K.
(2017). Performance of machine learning algorithms in predicting game
outcome from drafts in Dota 2. Communications in Computer and
Information Science, 26–37.
https://doi.org/10.1007/978-3-319-52920-2_3
[7] Costa, L. M., Mantovani, R. G., Monteiro Souza, F. C., & Xexeo, G.
(2021). Feature analysis to league of legends victory prediction on the
picks and bans phase. 2021 IEEE Conference on Games (CoG).
https://doi.org/10.1109/cog52621.2021.9619019
[8] Silva, A. L. C., Pappa, G., & Chaimowicz, L. (1970, January 1).
continuous outcome prediction of League of Legends Competitive
matches using recurrent neural networks: Semantic scholar. Semantic
Scholar. Retrieved January 10, 2022, from
https://www.semanticscholar.org/paper/Continuous-Outcome-Prediction
-of-League-of-Legends-Silva-Pappa/85a507426e3df16c0e5624fba0ad1f
8572ce1874
[9] Lin, L. (2016). League of Legends Match Outcome Prediction.
Retrieved January 10, 2022, from
http://cs229.stanford.edu/proj2016/poster/Lin-LeagueOfLegendsMatchO
utcomePrediction-poster.pdf
[10] Michel's fanboi, “League of Legends Diamond Ranked Games (10
min)”, Kaggle, April 13, 2020,
https://www.kaggle.com/bobbyscience/league-of-legends-diamond-rank
ed-games-10-min

45

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on June 09,2023 at 08:43:11 UTC from IEEE Xplore. Restrictions apply.