Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends

Abstract

League of Legends (LoL) is the most widely played multiplayer online battle arena (MOBA) game in the world. An important aspect of LoL is competitive ranked play, which utilizes a skill-based matchmaking system to form fair teams. However, players' skill levels vary widely depending on which champion, or hero, that they choose to play as. In this paper, we propose a method for predicting game outcomes in ranked LoL games based on players' experience with their selected champion. Using a deep neural network, we found that game outcomes can be predicted with 75.1% accuracy after all players have selected champions, which occurs before gameplay begins. Our results have important implications for playing LoL and matchmaking. Firstly, individual champion skill plays a significant role in the outcome of a match, regardless of team composition. Secondly, even after the skill-based matchmaking, there is still a wide variance in team skill before gameplay begins. Finally, players should only play champions that they have mastered, if they want to win games.

Full text

Using Machine Learning to Predict Game Outcomes Based on
Player-Champion Experience in League of Legends
Tiany D. Do
University of Central Florida
Orlando, Florida, United States
tianydo@knights.ucf.edu
Seong Ioi Wang
University of Texas at Dallas
Richardson, Texas, United States
seongiwang@gmail.com
Dylan S. Yu
University of Texas at Dallas
Richardson, Texas, United States
dylanyu461@gmail.com
Matthew G. McMillian
University of Texas at Dallas
Richardson, Texas, United States
matthewgmcmillian@gmail.com
Ryan P. McMahan
University of Central Florida
Orlando, Florida, United States
rpm@ucf.edu
ABSTRACT
League of Legends (LoL) is the most widely played multiplayer
online battle arena (MOBA) game in the world. An important as-
pect of LoL is competitive ranked play, which utilizes a skill-based
matchmaking system to form fair teams. However, players’ skill
levels vary widely depending on which champion, or hero, that
they choose to play as. In this paper, we propose a method for
predicting game outcomes in ranked LoL games based on players’
experience with their selected champion. Using a deep neural net-
work, we found that game outcomes can be predicted with 75.1%
accuracy after all players have selected champions, which occurs
before gameplay begins. Our results have important implications
for playing LoL and matchmaking. Firstly, individual champion
skill plays a signicant role in the outcome of a match, regardless
of team composition. Secondly, even after the skill-based match-
making, there is still a wide variance in team skill before gameplay
begins. Finally, players should only play champions that they have
mastered, if they want to win games.
CCS CONCEPTS
•Information systems →Massively multiplayer online games
;
•Applied computing →Computer games.
KEYWORDS
game outcome prediction, League of Legends, multiplayer online
battle arenas, player experience, machine learning, matchmaking
ACM Reference Format:
Tiany D. Do, Seong Ioi Wang, Dylan S. Yu, Matthew G. McMillian, and Ryan
P. McMahan. 2021. Using Machine Learning to Predict Game Outcomes
Based on Player-Champion Experience in League of Legends. In The 16th
International Conference on the Foundations of Digital Games (FDG) 2021
(FDG’21), August 3–6, 2021, Montreal, QC, Canada. ACM, New York, NY,
USA, 5pages. https://doi.org/10.1145/3472538.3472579
FDG’21, August 3–6, 2021, Montreal, QC, Canada
©2021 Association for Computing Machinery.
This is the author’s version of the work. It is posted here for your personal use. Not for
redistribution. The denitive Version of Record was published in The 16th International
Conference on the Foundations of Digital Games (FDG) 2021 (FDG’21), August 3–6, 2021,
Montreal, QC, Canada,https://doi.org/10.1145/3472538.3472579.
1 INTRODUCTION
League of Legends (LoL), a popular computer game developed by
Riot Games, is currently the most widely played Multiplayer Online
Battle Arena (MOBA) [
3
] game in the world. In 2019, there were
eight million concurrent players daily [
14
], and the player base has
continued to grow since its release in 2009. A core aspect of LoL
is competitive ranked gameplay. In typical ranked gameplay, ten
human players are matched together to form two teams of approxi-
mately equal skill. These two teams, consisting of ve players each,
battle against each other to destroy the opposing team’s base.
Fair matchmaking is crucial for player experience [31]. In 2019,
Riot Games stated that ranked matchmaking should be as fair as
possible [
11
]. This goal has persisted throughout the history of the
game. In 2020, Riot Games stated that some of their main goals for
the year were to preserve competitive integrity [
12
] and improve
matchmaking quality [
13
] for ranked games. In order to create fair
matches between players of approximately equivalent skill level,
matchmaking is determined using an Elo rating system, similar to
the one originally used by chess players [
26
]. Although this match-
making system has improved in recent years (c.f., [
10
,
11
,
13
]),
it does not consider players’ champion selections when forming
matches. LoL has over 150 playable characters, known as champi-
ons, that have their own unique playstyles and abilities [
7
]. Players
select a champion at the start of every match after the matchmaking
algorithm has formed teams. However, players will often perform
better on some champions than on others due to their diering
levels of mechanical expertise, which is dened as a player’s knowl-
edge of their champion’s abilities and interactions [
8
]. Higher levels
of mechanical expertise on particular champions allow players to
make quicker and better judgments, which are essential in the
game’s fast paced environment. Since mechanical expertise plays
such a large impact on a player’s own performance, it can therefore
cause a similar impact on the match’s outcome.
In this paper, we introduce a machine learning model based
on a deep neural network (DNN) that can predict ranked match
outcomes based on players’ experience on their selected champion
(i.e., player-champion experience). We show that the outcome of
a ranked match can be predicted with over 75% accuracy after all
players have selected their champions. This occurs before the match
actually begins. Our results indicate that individual champion skill
plays a signicant role in the outcome of a match, regardless of
arXiv:2108.02799v1 [cs.LG] 5 Aug 2021

FDG’21, August 3–6, 2021, Montreal, QC, Canada Do, et al.
team composition. Additionally, current matchmaking may not
form teams of approximately equal skill level in game because
matchmaking is done before players select their champions. Our
results also imply that players should only play champions that
they have mastered in order to win ranked games. This can decrease
gameplay variety as players typically master only a few champions
out of the 150+ champions available in the roster [5,7].
2 BACKGROUND AND RELATED WORK
2.1 Ranked Gameplay in League of Legends
Ranked gameplay in LoL is based on seasonal gameplay, similar to
traditional sports. During each season, LoL separates players into
six main divisions known as Leagues based on skill level. These
Leagues are known as Iron, Bronze, Silver, Gold, Platinum, and
Diamond [
23
]. There are a few distinctions past Diamond. However,
we do not consider these players for the purposes of this player
as they make up a minuscule portion of the playerbase. In order
to rank up and enter higher Leagues, players must win ranked
matches. In ranked matches, two teams of ve players each are
formed using a matchmaking algorithm.
For ranked matches, LoL utilizes a matchmaking algorithm based
on a hidden Elo rating system, similar to the one originally used
by chess players [
26
]. This matchmaking algorithm has been con-
tinuously improved upon since the release of the game and is used
to create matches consisting of players of approximately equal
skill level [
10
,
11
,
13
]. In 2020, Riot released a new matchmaking
algorithm that better identies player skill level [13].
When a player decides to play a ranked match, they will be
matched with nine other players (via the matchmaking algorithm)
who are then separated into two teams. Ranked matches are played
with a minimum of eight strangers. Players may choose to enter
matchmaking alone or with a similarly ranked partner, in which
case they will be placed on the same team as their partner [26].
Since mechanical expertise is important to a player’s perfor-
mance, it seems logical that players will select champions that
they have high expertise on. However, this may not always be the
case. An important factor of champion selection is team composi-
tion. An optimal team has a combination of dierent roles in the
game [
21
]. Certain champions have dierent roles, with some ex-
amples being "Tank" (champions with high defense), "Mage" (long
ranged champions with burst damage and crowd control), and "Sup-
port"(champions with support ability) [
15
]. However, since players
are playing with strangers, they cannot predetermine optimal team
compositions. For instance, if no players on the team typically play
"Tank" champions, one player may be forced to play a "Tank" cham-
pion, even though they may not have high mechanical expertise
on it.
2.2 LoL Game Outcome Predictors
In this section, we describe previous work on LoL game outcome
predictors. Most MOBA game outcome predictors are focused on
two popular MOBAs: LoL and Dota 2. For the purposes of this paper,
we focus only on LoL predictors due to the dierences in gameplay
between LoL and Dota 2. For instance, Chen et al. [
5
] indicated that
LoL has more diverse skill compositions than Dota 2.
There are several predictors that use pre-match data (i.e., be-
fore the match begins) like our predictor. However, the majority
of these pre-match predictors have an accuracy around 55-60%
(e.g., [
5
,
20
,
24
]). Similar to our work, Chen et al. [
5
] used player-
champion experience to predict game outcomes. Using a combi-
nation of player-champion experience, champion experience, and
player experience, they predicted game outcomes with an accu-
racy of 60.24% using logistic regression. They found that player-
champion experience was more inuential than both champion
and player experience on game outcomes. We further expand on
this work by investigating multiple features of player-champion
experience, since Chen et al. [
5
] used only one feature to describe
player-champion experience.
Within-match predictors predict the winner while the match is
ongoing. Typically, as the game time elapses, the accuracy increases.
Data such as kill dierence, gold dierence, and towers destroyed
can help determine which team has the lead and will ultimately win
the game. Lee et al. [
22
] found that within-match data can predict
the winner with 62.26% accuracy at 5 minutes elapsed and 73%
accuracy at 15 minutes elapsed using Random Forest (RF). Similarly,
Silva et al. [
25
] predicted the winner with an accuracy of 63.91% at 5
minutes elapsed and 83.54% at 25 minutes elapsed using a recursive
neural network. Unlike these predictors, our model does not rely
on within-match data and instead predicts the match outcome
before the match has begun, deviating entirely in functionality and
application.
Ani et al. [
1
] found that pre-match data can predict the game
outcome of professional LoL games with very high accuracy (>90%)
using RF trees and ban data (each team is allowed to ban up to
ve champions). However, we focus on ranked gameplay rather
than organized professional gameplay. Unlike professional games,
ranked games are designed to have fair teams consisting of strangers
and support all varieties of skill levels. Additionally, professional
games contain a much dierent playerbase from ranked games, as
they are only played among a couple dozen players.
3 METHODOLOGY
3.1 Dataset
Riot Games provides a public API [
9
] containing several types of
data pertaining to players and their matches. This data contains
features such as champion mastery, games played per season, and
win rates. Since our methodology involves training machine learn-
ing models with data from ranked matches, we used this API to
query random players and lter their most recent ranked match.
Using this method, we pulled a total of 5000 unique ranked matches
for our dataset. Any duplicate matches were removed. From these
matches, we recorded general match information, as well as in-
formation about each player, specically because they describe a
player’s experience on a given champion:
•Champion mastery points
- An integer
∈ [
0
,∞]
approxi-
mating the player’s lifetime experience on the given cham-
pion [
7
]. Champion mastery points are accrued by playing
matches on the champion. The mean champion mastery
points of players in our dataset was 122,368.44, while the
max value was 9,364,624.

Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends FDG’21, August 3–6, 2021, Montreal, QC, Canada
•Player-champion win rate
- A percentage (0.0 to 1.0) in-
dicating the player’s victory ratio for ranked games played
on the given champion during the 2020 ranked season.
•Season total number of games played on the cham-
pion
- An integer
∈ [
0
,∞]
signifying the total number of
ranked games the player has played on the given champion
during the 2020 ranked season. The mean season total num-
ber of games of players in our dataset was 58.95, while the
max value was 1,819.
•Number of recent games played on the champion
- An
integer
∈ [
0
,
20
]
representing the number of ranked games
that the player has played on the champion within their last
20 ranked games during the 2020 ranked season.
These features represented each player’s experience levels on
their chosen champion and helped us determine whether player-
champion experience can predict match outcomes. Overall, our
methodology allowed us a large pool of data to build upon and use
for our model, while also stripping away personally identiable
data from each player.
3.2 Feature Selection
Using Pearson’s correlation test, we discovered that player-champion
win rate and champion mastery points were the only features that
correlated with the outcome of a match, and that the total number
of games played and number of recent games played on a particular
champion had no eect on the outcome of the match. Furthermore,
during preliminary analysis, our models performed worse when
the other features were included in the training dataset.
For each match, we derived additional features to better reect
each team’s overall player-champion experience. For both teams,
we added the team’s average, median, coecient of excess kurtosis
[
32
] (i.e., how normal the distribution is), coecient of skewness
[
32
], standard deviation, and variance of the players’ champion
win rate and of their champion mastery points. The nal training
dataset contained 44 features per match (2 features per player for
10 players and 12 features per team).
3.3 Models
We chose to explore several machine learning models based on
the success of previous work that also investigated MOBA game
outcomes [
1
,
2
,
22
,
24
,
29
]. We investigated Support Vector Classi-
ers (SVC) [
19
], k-Nearest Neighbors (kNN) [
18
], Random Forest
(RF) trees [
17
], Gradient Boosting (GBOOST), and Deep Neural Net-
works (DNN) [
17
]. In this section, we briey explain each model,
as well as any parameters that we used for our implementations.
3.3.1 Support Vector Classifier. An SVC is a form of Maximal Mar-
gin Classier that attempts to form a
(𝑝−
1
)
dimensional separable
hyperplane on a
𝑝
dimensional feature space [
18
]. SVCs intention-
ally misclassify training observations in an attempt to improve the
classication of remaining observations [
18
]. We used scikit-learn’s
[
28
] implementation of SVC with the following modied param-
eters:
C=8
,
coef0=1
,
kernel=’poly’
, and
tol=1e-2
. We opted to
use a polynomial kernel over a radial-basis-function kernel [
18
] be-
cause the polynomial kernel showed better classication accuracy.
3.3.2 k-Nearest Neighbors. The kNN [
27
] algorithm takes a given
record
𝑥
and compares it with the
𝑘
closest records from its corre-
sponding dataset, with closeness dened using some distance mea-
sure. We used scikit-learn’s [
28
] implementation of kNN with the
following modied parameters:
leaf_size=5
,
n_neighbors=600
,
p=1
and
weights=’distance’
. For this model in particular, we no-
ticed that including more features caused our overall prediction
accuracy to drop dramatically. In order to circumvent the curse of
dimensionality [
18
], we reduced our feature space to only include
the team’s average win rate.
3.3.3 Random Forest Trees. RF trees [
17
] use bagging to create
multiple base decision tree classiers and aggregate them into a
single model. This method can yield a model with less variance and
overtting in comparison to classic decision trees. We used scikit-
learn’s [
28
] implementation of RF trees [
17
] with the following
modied parameters:
n_estimators=350
,
min_samples_leaf=3
,
min_samples_split=10, and max_features=7.
3.3.4 Gradient Boosting. GBOOST is another ensemble method
that sequentially adds predictors to its ensemble and tries to t each
new predictor based on the residual errors of the previous predictor
[
16
]. We used scikit-learn’s [
28
] implementation of GBOOST, whose
default estimator is a Decison Tree [
16
], with the following modied
parameters: learning_rate=0.14 and n_estimators=55.
3.3.5 Deep Neural Networks. We utilized Keras [
6
], a Python deep
learning API, to build our model. We decided to go with a pyramid
network architecture based on examples described in [
16
], as we
noticed that it performed better than networks with hidden layers
with the same number of neurons in our preliminary experiments.
We used the following network topology to dene our model:
•A attened input layer resulting in a 1×44 output.
•
Alternating dropout, normalization, and dense layers for a
total of 15 layers (5 dropout, 5 normalization, and 5 dense
layers). Each group of alternating layers had 160, 128, 64, 32,
and 16 neurons, in that order.
–Each dropout layer had a dropout rate of 0.69%.
–Each normalization layer utilized batch normalization.
–
Each dense layer used Exponential Linear Unit (ELU) acti-
vation [16], He initialization [16].
•A1×1dense layer with Sigmoid activation
Notably, we used ELU activation, He initialization and batch nor-
malization to deal with unstable gradients [
16
], which led to more
accurate model predictions. We also added dropout layers to prevent
overtting [30].
4 EXPERIMENTS AND RESULTS
4.1 Experimental Settings
For our experiments, we split our dataset into a training and testing
dataset. For the training dataset, 80% of the matches (4000 matches)
were randomly chosen, while the other 20% (1000 matches) were
chosen for the testing dataset. We utilized stratied k-fold validation
[
4
] with
𝑘=
10 to evaluate our models. Stratied k-fold validation
tries to keep the class distribution of the testing dataset as similar
to the actual class distribution as possible, which can lead to better

FDG’21, August 3–6, 2021, Montreal, QC, Canada Do, et al.
Table 1: The performance of each machine learning model,
including the accuracy, 95% condence interval (CI), stan-
dard deviation, and standard error.
Model Mean Accuracy ±CI Std. Dev. Std. Error
SVC 74.3% ±1.21% 1.7% 0.54%
KNN 72.7% ±1.23% 1.2% 0.38%
RF 74.7% ±1.2% 2.0% 0.63%
GBOOST 75.4% ±1.19% 5.25% 1.66%
DNN 75.1% ±1.2% 1.9% 0.60%
Figure 1: The loss and accuracy of our deep neural network
while training and validating.
accuracy. Each model’s average accuracy, its standard deviation,
and its standard error are shown in Table 1.
4.2 Results
From our experiment results (see Table 1), we can see that all models
predicted the winner with relatively high accuracy (>70%). GBOOST
had the highest accuracy with 75.4%, which was marginally higher
than our DNN model’s of 75.1%. However, the standard error of
GBOOST was much greater than all of our other models. Therefore,
we consider our DNN model to be most suitable for game outcome
prediction since its standard error was low. Interestingly, from
Figure 1, we can see that the validation accuracy is higher than
the training accuracy of our DNN model. This is a byproduct of
using dropout layers [
16
]. Using dropout layers articially increases
the diculty of learning for the network during training, with the
intent to increase test/validation accuracy.
5 LIMITATIONS AND FUTURE WORK
It is important to note that our work is limited to our specic do-
main. We only analyzed ranked matches from the North American
(NA) server that occurred in 2020 within ranks Iron to Diamond.
These results may not be similar for other servers such as the Ocea-
nia (OCE) server, due to dierences between the playerbases, and
may not reect highly skilled play (top 1%). Additionally, the player-
base may change over time. In the future, it would be interesting
to investigate other regional servers (e.g., EUNE, EUW, OCE) to
determine if player demographics have an eect on prediction ac-
curacy. Furthermore, it may be of value to determine player-role
experience and determine if this could serve as a similar predictor
of a player’s skill.
6 CONCLUSION
In this paper, we introduced a DNN machine learning model that
can predict the accuracy of ranked LoL matches with an accuracy
of 75.1% using player-champion experience. To reect each player’s
champion experience in a ranked match, we extracted multiple
features regarding player statistics on their chosen champion. How-
ever, we found that only champion mastery points and ranked win
rate on the selected champion were suitable for predicting game
outcomes. We used this data to create summary features that repre-
sented the overall player-champion experience of the team. Using
these summary features, we also found high accuracy (73%-75%)
using other machine learning models, such as random forest trees
and support vector classiers, showing that player-champion expe-
rience can be a strong indicator of team performance in general.
Our results oer insight into the importance of champion se-
lection and matchmaking design. Since we are able to determine
the winner with reasonable accuracy after champion selection, this
implies that even after the skill-based matchmaking, there is still a
wide variance in team skill before gameplay begins. Fair matchmak-
ing is important for player experience [
31
], and ranked matches
in League of Legends are designed to be as fair as possible [
11
].
However, the current skill-based matchmaking system does not
consider a player’s champion selection, which can widely vary
their skill level and thus cause the match to feel imbalanced. We
show that the outcome of a ranked match can be determined with
relatively high accuracy before gameplay begins based solely on
player-champion experience. Although this problem cannot be re-
solved with the current matchmaking system, our results imply
that players should limit their choices to champions that they have
already mastered, regardless of team composition. However, most
players can only master a handful of champions [
5
]. This can be an
issue for gameplay variety as players may choose not to experience
the large roster of champions, in favor of playing the same few
champions in order to win games.
REFERENCES
[1]
R. Ani, Vishnu Harikumar, Arjun K. Devan, and O. S. Deepa. 2019. Victory
prediction in league of legends using feature selection and ensemble methods.
2019 International Conference on Intelligent Computing and Control Systems, ICCS
2019 Iciccs (2019), 74–77. https://doi.org/10.1109/ICCS45141.2019.9065758
[2]
Mochammad Anshori, Farhanna Mar’i, Mukhammad Wildan Alauddin, and
Fitra Abdurrahman Bachtiar. 2018. Prediction Result of Dota 2 Games Using
Improved SVM Classier Based on Particle Swarm Optimization. 3rd International
Conference on Sustainable Information Engineering and Technology, SIET 2018
(2018), 121–126. https://doi.org/10.1109/SIET.2018.8693204
[3]
M. Aung, V. Bonometti, A. Drachen, P. Cowling, A. V. Kokkinakis, C. Yoder, and
A. Wade. 2018. Predicting Skill Learning in a Large, Longitudinal MOBA Dataset.
IEEE Conference on Computatonal Intelligence and Games, CIG 2018-August (2018).
https://doi.org/10.1109/CIG.2018.8490431
[4]
José Ramón Cano, Francisco Herrera, and Manuel Lozano. 2007. Evolutionary
stratied training set selection for extracting classication rules with trade o
precision-interpretability. Data & Knowledge Engineering 60, 1 (2007), 90–108.
https://doi.org/10.1016/j.datak.2006.01.008 Intelligent Data Mining.
[5]
Zhengxing Chen, Yizhou Sun, Magy Seif El-Nasr, and Truong Huy D. Nguyen.
2017. Player skill decomposition in multiplayer online battle arenas. arXiv (2017).
arXiv:1702.06253

Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends FDG’21, August 3–6, 2021, Montreal, QC, Canada
[6] François Chollet et al. 2015. Keras. https://github.com/fchollet/keras.
[7]
Tiany D Do, Dylan S Yu, Salman Anwer, and Seong Ioi Wang. 2020. Using
Collaborative Filtering to Recommend Champions in League of Legends. In 2020
IEEE Conference on Games (CoG). 650–653. https://doi.org/10.1109/CoG47356.
2020.9231735
[8]
Scott Donaldson. 2017. Mechanics and Metagame: Exploring Binary Expertise in
League of Legends. Games and Culture 12, 5 (2017), 426–444. https://doi.org/10.
1177/1555412015590063
[9]
Riot Games. [n.d.]. Riot Developer Portal. https://developer.riotgames.com/apis
[10]
Riot Games. 2019. /dev: Making Matchmaking Better. https://nexus.
leagueoegends.com/en-us/2018/03/dev- making-matchmaking-better/
[11]
Riot Games. 2019. /dev: Updates to Ranked for 2019. https://nexus.
leagueoegends.com/en-us/2018/04/dev- updates-to-ranked- for-2019/
[12]
Riot Games. 2020. /dev: A look back at League in 2020. https://na.leagueoegends.
com/en-us/news/dev/dev-a-look-back-at-league-in-2020/
[13]
Riot Games. 2020. /dev: Updates on 2020 Ranked & Matchmak-
ing. https://na.leagueoegends.com/en-us/news/dev/dev-updates-on-2020-
ranked-matchmaking/
[14]
Riot Games. 2020. Game Updates. https://na.leagueoegends.com/en-us/news/
game-updates/
[15]
Gege Gao, Aehong Min, and Patrick C. Shih. 2017. Gendered design bias: Gender
dierences of in-game character choice and playing style in league of legends.
Proceedings of the 29th Australian Conference on Computer-Human Interactio
(2017), 307–317. https://doi.org/10.1145/3152771.3152804
[16]
Aurélien Géronen. 2017. Hands-on machine learning with Scikit-Learn and Tensor-
Flow : concepts, tools, and techniques to build intelligent systems. O’Reilly Media,
Sebastopol, CA.
[17]
Trevor Hastie, Robert Tibshirani, and Jerome Friedman. 2001. The Elements of
Statistical Learning. Springer New York Inc., New York, NY, USA.
[18]
Gareth James, Daniela Witten, Trevor Hastie, and Robert Tibshirani. 2014. An
Introduction to Statistical Learning: With Applications in R. Springer Publishing
Company, Incorporated.
[19]
V. Kecman. 2005. Support Vector Machines: Theory and Applications. Vol. 177. 1–47
pages.
[20] Cheolgi Kim and Soowon Lee. 2019. Predicting Win-Loss of League of Legends
at the Start of the Game Using Machine Learning Techniques. In Proceedings of
ISSAT International Conference on Data Science in Business, Finance and Industry.
[21]
Yubo Kou and Xinning Gui. 2014. Playing with strangers: Understanding tem-
porary teams in league of legends. CHI PLAY 2014 - Proceedings of the 2014
Annual Symposium on Computer-Human Interaction in Play (2014), 161–169.
https://doi.org/10.1145/2658537.2658538
[22]
Sang Kwang Lee, Seung Jin Hong, and Seong Il Yang. 2020. Predicting Game
Outcome in Multiplayer Online Battle Arena Games. International Conference
on ICT Convergence 2020-October (2020), 1261–1263. https://doi.org/10.1109/
ICTC49870.2020.9289254
[23]
Xiangqian Li, Liang Huang, Bingxin Li, Haoran Wang, and Chengyang Han. 2020.
Time for a True Display of Skill: Top Players in League of Legends Have Better
Executive Control. Acta Psychologica 204 (2020). https://doi.org/10.1016/j.actpsy.
2020.103007
[24] Lucas Lin. 2016. League of Legends Match Outcome Prediction. (2016).
[25]
Antonio Luis, Cardoso Silva, Gisele Lobo Pappa, and Luiz Chaimowicz. 2018.
Continuous Outcome Prediction of League of Legends Competitive Matches
Using Recurrent Neural Networks. Proceedings of SBGames 2018 (2018), 639–642.
[26]
Marçal Mora-Cantallops and Miguel Ángel Sicilia. 2018. Exploring player experi-
ence in ranked League of Legends. Behaviour and Information Technology 37, 12
(2018), 1224–1236. https://doi.org/10.1080/0144929X.2018.1492631
[27]
Mahmoud Parsian. 2015. Data Algorithms: Recipes for Scaling Up with Hadoop
and Spark (1st ed.). O’Reilly Media, Inc.
[28]
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour-
napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine
Learning in Python. Journal of Machine Learning Research 12 (2011), 2825–2830.
[29]
Aleksandr Semenov, Peter Romov, Sergey Korolev, Daniil Yashkov, and Kirill
Neklyudov. 2017. Performance of machine learning algorithms in predicting game
outcome from drafts in Dota 2. Communications in Computer and Information
Science 661 (2017), 26–37. https://doi.org/10.1007/978-3- 319-52920-2_3
[30]
Nitish Srivastava, Georey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan
Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from
overtting. The journal of machine learning research 15, 1 (2014), 1929–1958.
[31]
Maxime Véron, Olivier Marin, and Sébastien Monnet. 2014. Matchmaking in
multi-player on-line games: Studying user traces to improve the user experience.
Proceedings of the 24th ACM Workshop on Network and Operating Systems Support
for Digital Audio and Video, NOSSDAV 2014 (2014), 7–12. https://doi.org/10.1145/
2578260.2578265
[32]
Daniel Zwillinger and Stephen Kokoska. 1999. CRC standard probability and
statistics tables and formulae. Crc Press.