Classifying Dota 2 Hero Characters Based on Play Style and Performance

Classifying Dota 2 Hero Characters Based on Play Style and
Performance
Lynn Gao, James Judd, Dave Wong, and Jamie Lowder
Abstract— In this paper, heroes from the video game Dota 2
are classified by hero identifier and the positional role most
commonly assigned to that hero. Classification is based on
play style and performance data. Dota 2 is a team-based,
competitive video game with complex rules and strategy. There
is an active community of professional teams and broadcasters,
which make Dota 2 an integral part of the eSports community.
We collect and label data from professional and public Dota 2
matches. Using this data and supervised learning techniques, we
attempt to predict hero identifiers and positional roles. Logistic
regression and random forest classification methods are used to
train classifiers on labeled Dota 2 player data. These classifiers
are used for prediction. We are able to classify Dota 2 heroes
using data from professional matches with about 90% accuracy
for hero positional role and 75% accuracy for hero identifier.
Data from public matches is classified with about 85% accuracy
for hero positional role and 75% accuracy for hero identifier.
We conclude that classification of Dota 2 heroes by positional
role and hero identifier is relatively accurate.
I. INTRODUCTION
In this paper, player controlled heroes from the video game
Dota 2 are classified. Dota 2 is an action real-time strategy
video game developed by the Valve Corporation and released
in July 2013. In Dota 2, discrete matches take place featuring
two teams of five players. Each player controls a single hero
character throughout the course of a match, with the goal of
defeating the opposing team. The performance and actions
of heroes in a match tend to vary by hero. Accordingly, the
performance of a hero in a match and the actions taken by
a hero during a match are used for classification.
A. Motivation
In team sports, such as basketball or American football,
players are commonly assigned to a particular position on
the field of play. For example, a quarterback or lineman in
American football, or a point guard in basketball. In some
sports, physiological attributes and performance characteristics of players tend to vary based on position [1], [2], [3], [4],
[5], [6]. It is not unreasonable to assume that a quarterback
in American football might complete more passes than a
lineman, or that a center in basketball might be the tallest
player on the team.
Much like traditional team sports, heroes in Dota 2 are
assigned to positions that define the role of a hero in a
match. Although Dota 2 is not a traditional team sport like
basketball, it is still a team sport. Accordingly, we assume
that performance characteristics of heroes will tend to vary
based on position.
It may seem odd to call Dota 2 a team sport. However, in
recent years, team based professional video game competi-
Fig. 1.
Prize pools in millions of USD for major eSports events in 2013.
tions, also known as eSports, have become prevalent due to
the popularity of games like Dota 2 and League of Legends.
Major international tournaments featuring large prize pools
attract a vast number of viewers and have given rise to
professional eSports teams. As shown in Figure 1, prize
pools of major eSports events in 2013 have exceeded two
million dollars. Recent major eSports events have attracted
millions of viewers. The League of Legends Season 3 World
Championship attracted over 32 million viewers and more
than 8.5 million concurrent viewers. The Dota 2 International
3 Championships attracted over 1 million concurrent viewers.
We now present a basic overview of the rules and strategy
of Dota 2. We do this in order to give the reader a better
understanding of our feature selection for classification. To
clarify, in American football the number of pass completions
might be a relevant feature to classify between a quarterback
and a lineman. However, if the reader is unfamiliar with
American football, this feature will be unintuitive.
B. Dota 2 Basic Rules
Dota 2 is a game involving two teams of five players. The
game is played on a rectangular map, as shown in Figure 2.
Each team occupies a fortified stronghold located at opposite
corners of the map. In the center of each team’s stronghold
there is a building known as the Ancient. A match is won
when either team’s Ancient is destroyed. A team’s ancient
and experience priority is the support. Heroes used for this
position are usually strong early and weak later in the game.
Supports usually help carries acquire gold and experience in
the early game, and are later ‘carried’ to victory by the hero
in the carry position. Solo lane heroes are given an average
amount of gold and experience priority and perform well by
themselves. They tend to be strong in the midgame.
II. R ELATED W ORK
Fig. 2. Dota 2 map. Fortified strongholds, ancients, and lanes (top, middle,
bottom) are highlighted. In the bottom left corner resides the Radiant
stronghold. In the top right corner is the Dire stronghold. The circles in
the center of either stronghold are the Ancients.
is invulnerable until a series of structures in at least one lane
(top, middle, or bottom) is destroyed.
In Dota 2 each player controls one hero. At the time of
writing there are over 100 heroes available in the game. Each
hero has a different set of attributes, meaning different heroes
have different amounts of health, do different amounts of
damage, etc.
Every hero has a unique set of abilities. These abilities
allow the hero to interact with the game world. An ability
can be used to damage enemy heroes, heal allied heroes,
etc. Heroes have six inventory slots. Items can be placed
in inventory slots. An item somehow makes a hero more
valuable. For example, items can give a hero more health,
or cause a hero to deal more damage, etc.
Throughout the course of a Dota 2 match, a hero will
acquire gold and experience points. When a hero acquires
enough experience points, the hero will level up. Upon level
up, a hero will gain one skill point, which can be used to
upgrade one of the hero’s abilities. Upgrading an ability will
somehow make the ability better. The gold a hero acquires
can be used to purchase items, which can be placed in the
hero’s inventory slots. As mentioned earlier, items somehow
make a hero more valuable.
C. Dota 2 Basic Strategy
In Dota 2 there are three positions: carry, solo lane, and
support. As seen in Figure 3, positions are determined by
the gold and experience priority of that position. Carry is
the position with the highest gold and experience priority.
The heroes used for this position are usually weak early and
strong later in the game. The position with the lowest gold
There is research which suggests that performance characteristics and physiological attributes of athletes vary based
on position. In this section, a brief review of the literature
surrounding the impact of positional role is conducted. Ziv
and Lidor [5] studied the physical attributes, physiological
characteristics, and on-court performances of female and
male elite basketball players. They found that differences
in physical attributes, physiological characteristics, and performance existed among positions and skill level. Ziv and
Lidor [6] also conducted a similar study of handball players.
They found that differences exist among handball players
of different positions and skill levels in physical attributes,
physiological characteristics, and performance.
Salvo et al. [1] used a computerized match analysis system
to study the motion characteristics of top class soccer players.
They found that the distance traveled at various intensities
varied by positional role and skill level. Nicholas [4] studied
rugby union football players and found that the anthropometric and physiological characteristics of the players varied by
positional role and playing standard. K¨okl¨u et al. [3] studied
fitness characteristics of Turkish professional basketball players. They found that physical fitness characteristics tend to
vary by position. Gocentas et al. [2] investigated the positionrelated differences in cardio respiratory functional capacity
(physiological attribute) of elite basketball players. They
found that there exist significant physiological differences
between players of different positions.
Overall, the work referenced above suggests that performance characteristics, and physical and physiological attributes tend to vary based on positional role in a variety
of traditional team sports. Although the number of sports
studied in the referenced work is limited, the research
seems to suggest that that positional role has an impact
on performance characteristics, and physical/physiological
attributes regardless of the team sport. Accordingly, our
assumption that these findings extend to Dota 2 does not
seem unreasonable. Whether our assumption is reasonable or
not will be further explored when we attempt classification
of Dota 2 heroes by play style and performance.
III. P RELIMINARIES AND P ROBLEM D EFINITION
In this paper we classify Dota 2 heroes. In each game of
Dota 2, a player controls one hero. Throughout the course of
the match, the player performs various actions. These various
actions are recorded and stored by the Valve Corporation and
are accessible after a match has been completed. We define
the actions a player takes during the game to be the play style
and performance data associated with a player in a match.
Fig. 3.
Dota 2 positional roles sorted by gold and experience priority.
Accordingly, for every game of Dota 2 there is performance
and play style data associated with each hero in a match. This
data is analogous to performance data in traditional team
sports. For example, points scored in basketball or running
yards in American football.
To classify Dota 2 heroes we rely heavily on supervised
learning techniques. In general, these techniques involve a
data set Q of size n elements. This data set is comprised
of label, data pairs {yi , xi }ni=1 . yi is the label associated
with a d dimensional vector xi . The elements of xi are
data {xj }dj=1 . In our case the elements of xi are Dota 2
performance and play style data. The collection of labels yi
and data vectors xi in data set Q are used to train a classifier.
After being trained, this classifier can be used to predict the
label yz of a data vector xz .
In our problem, we have two labels associated with a data
vector, yi and zi , representing the numerical identifier of
a hero and the role the hero is most commonly used for,
respectively. We collect a data vector xi for each player in
a Dota 2 match, each player controlling one hero. We use
supervised learning techniques to classify either hero role or
hero identifier. Essentially, we have two data sets: {yi , xi }ni=1
and {zi , xi }ni=1 . We train a classifier using the data vector
and either the hero identifier or the hero role. After training,
we use the classifier to predict either the hero identifier yi
or the hero role zi . The accuracy of the classifier is defined
to be the percentage of labels which the classifier accurately
predicts.
In essence, we collect play style and performance data
from Dota 2 matches. We use this data and labels associated
with this data to train a classifier. We then use this classifier
to predict the labels associated with play style and performance data. We define the accuracy of a classifier to be the
percentage of data the classifier accurately predicts labels for.
IV. A PPROACH
We classify Dota 2 heroes by hero identifier and the
positional role most commonly assigned to a hero. We perform this classification based on play style and performance
data obtained from Dota 2 matches. A single data point is
defined as the performance and play style data for one player
controlling one hero during the course of one match. The
specific performance and play style data utilized is explained
later in this section. Each match features 10 players each
controlling one hero, hence we collect ten data points per
match.
The data we collect is labelled using two labels: the
numerical identifier (ID) of the hero and the positional role
most commonly assigned to the hero. The numerical ID of
the hero is a number which uniquely identifies a particular
hero. At the time of writing, there are 108 heroes in Dota 2,
hence this number is a number from 1-108. The positional
role is a number which uniquely identifies the position a hero
is most commonly assigned to. There are three roles, hence
the role identifiers are
1) Carry
2) Solo lane
3) Support
Classification is performed separately using either of these
two labels, but not using both labels. This fits the problem
description where we classify heroes by ID and hero positional role separately.
We collect two data sets: a professional data set and public
data set. The professional data set is comprised of matches
played which took place during The International 3 Dota
2 Championships in August 2013. The International is an
annual invitational Dota 2 tournament. The Valve Corporation invites the top professional Dota 2 teams from around
the world to participate in this tournament. Accordingly,
the professional data set is comprised of data from the
top professional Dota 2 players in the world. This data is
equivalent to basketball data collected from the National
Basketball Associated Playoffs.
The data collected from The International 3 are assumed
to be high quality and are not filtered by match duration. We
include only normal Dota 2 matches from The International
3. To clarify, we exclude the solo championship matches and
the all-star show match.
The public data set is comprised of matches which were
created by Valve’s public matchmaking service. Valve uses
a matchmaking service to match players of appropriate
skill levels together from the pool of all players currently
searching for a match. There is no restriction on player skill
to participate in public matchmaking. Accordingly, matches
in the public data set consist of very low skill matches to
matches of players with near professional skill levels.
Because the matches in public matchmaking are not guaranteed to be high quality, we restrict the matches we collect
by balance patch, duration, lobby type, and game mode. In
addition, we only use player data from players where the
player was connected to the game for the entire match. It is
possible for a player to disconnect from and abandon a Dota
2 match. This behavior is heavily discouraged by Valve, but
it does happen. In an attempt to avoid data from a player
which played in only part of a match, we filter players who
disconnected or abandoned from a game.
In the public data set we collect matches which took place
during June, 2013. All the matches collected took place
during Dota 2 balance patch 6.78. Balance patches in Dota
2 are periodically released and adjust settings in the game in
an attempt to prevent heroes or items in the game from being
too powerful or weak. Sometimes balance patches change the
positional role of a hero or remove a hero from the game.
Accordingly, we collect matches which took place during a
single balance patch.
Matches in the public data are between 20 and 80 minutes
in duration. This filtering is done in an attempt to avoid
low quality matches. Matches in Dota 2 are typically 25-50
minutes in duration. Occasionally matches last longer than
60 minutes, and extremely rarely matches last less than 20
minutes. By filtering very short or long matches, we attempt
to eliminate outliers from the public match data.
The public data set is filtered by lobby type (matchmaking
type) and game mode. We only collect data from matches
featuring ten human players created using public matchmaking lobbies, where players queue solo or in groups of one to
five players. We restrict the game modes to normal, team
based Dota 2 game modes. We exclude special seasonal
game modes and game modes dedicated for new players.
The allowable game modes are
•
•
•
•
•
•
•
All Pick
Captains Mode
Captains Draft
Single Draft
Random Draft
All Random
Compendium Matchmaking
The remainder of this section describes our methods. We
first discuss the methods by which we collect both our data
sets. Then we discuss feature selection. Finally we discuss
the supervised learning methods we use to classify our data.
A. Data Collection
In this section we introduce the methods by which data is
collected and prepared for classification.
We obtain match data for both data sets via the Valve
Dota 2 match history web application programming interface
(WebAPI). Data obtained from the match history WebAPI
includes detailed information about a Dota 2 match and
each of the players in that match. We use the data available
from the WebAPI to collect data from matches which meet
the criteria stated above. Data regarding player performance,
items a player had in their inventory at the end of a match,
and the order a player upgraded the abilities of their hero is
obtained from the WebAPI.
The professional data set consists of 124 matches that took
place during The International 3. This data is assumed to be
high quality, so this data set contains 1,240 data points. The
public data set consists of 12,000 matches that took place
during June 2013 and balance patch 6.78. This data is filtered
to avoid low quality data, so this data set contains 110,979
data points.
B. Feature Selection
In this section we discuss the features we extract from the
data. These features are used for hero classification.
For the purpose of classification, we transform the hero
data collected into features. We use 275 features per data
point. These features can be divided into three categories:
1) Performance (10 features)
2) Items (240 features)
3) Ability upgrades (25 features)
1) Performance Features: Performance features are data
regarding hero performance during a match normalized based
on the duration of a match. Dota 2 matches have no guaranteed length. The duration of a match may affect the amount
of gold a hero acquires during a match, the number of assists
a hero has during a match, etc. These features are analogous
to performance data in other sports. For example, points
scored in basketball or passing yards in American football.
There are ten performance features for each data point.
These features are per minute:
1) Kills
2) Deaths
3) Assists
4) Gold earned
5) Experience points acquired
6) Damage done to enemy heroes
7) Damage done to enemy towers
8) Health points of allied heroes healed
Per ten minutes:
9) Last hits (creeps killed)
10) Denies (creeps denied)
2) Item Features: At the time of writing, there are 240
unique items in Dota 2. We generate 240 features per
player, one for each item. Each item has a unique numerical
identifier from 1-240. We collect the ID’s of the items which
a player had in their inventory at the end of a match. Items
are expensive and not usually replaced throughout the course
of a match. Accordingly, the items in a player’s inventory at
the end of a match are a relatively decent indicator of the
items a player purchased across the course of a match.
We transform the item into features by creating a length
240 array filled with 0’s. Values in the array corresponding to
the ID’s of the items the player had in their inventory at the
end of the match are set to 1. The values at all other indices
remain 0. Accordingly, there can be at most 6 1’s set in this
array because a player has 6 inventory slots. For example, if
a player had items 1, 100, and 132 in their inventory at the
end of a match. Indices 1, 100, and 132 in the array would
be set to 1.
4.2.3 Ability Upgrade Features: When a hero acquires
enough experience points, the hero gains a level. The maximum level a hero can obtain is level 25. When a hero gains
a level, they gain one skill point that they can use to upgrade
one of their abilities. This makes the ability somehow better.
We acquire data for the ability a hero upgrades at each
level. Each hero has n abilities. We will label these abilities
from 1 to n based on the position from the left of the ability.
Fig. 4. Positional labeling of Dota 2 hero abilities. The ability furthest
left is 1. The ability furthest right is n. The abilities shown are for the hero
Elder Titan, who has 4 abilities.
The ability furthest left is ability 1, the ability furthest right
is ability n. Figure 4 gives an example of this numbering for
the hero Elder Titan.
We transform ability upgrade data features by creating
a length 25 array. The values in this array represent the
position of the ability the hero upgrades each level. Each
index corresponds to a hero level in order from level 1 to
level 25. To clarify, index 1 corresponds to level 1 and index
25 corresponds to level 25. We store the position of the
ability the hero upgrades at each level in the corresponding
index. Value 0 is used if a hero does not reach a particular
level or does not spend their skill points before a match
concludes. The value -1 is used if a hero uses their skill
point to upgrade their hero attributes (health, damage, mana,
etc.).
C. Classification Methods
In this section we discuss the methods used to classify
heroes by performance and play style into hero id or hero
positional role.
To classify heroes we use supervised learning techniques.
Specifically we use logistic regression and random forest
classifiers. Each classifier is trained using a subset of one
of the datasets (public or professional). The classifier is then
used to predict either hero role or hero id of the remaining
data in the dataset. The value predicted depends on the label
used to train the classifier. If we train using hero role, we
predict hero role. If we train using hero ID, we predict hero
ID.
We first partition a dataset into a training set and a testing
set. A random 90% of the data is selected for training. The
remaining 10% of the data is used for testing. Then the
label associated with this data and the features generated
from this data are used to train the classifier. After training,
the classifier is used to predict the hero role or hero ID
of the testing data. This process is repeated 20 times. A
random 90% is selected for training each time. The average
percentage accurately classified across all 20 runs is used to
determine the quality of the classifier.
The International 3 (TI3) competition. Our worst results for
classifying hero roles was 84% using logistic regression on
data obtained from public matches. The first key observation
that was made is that the TI3 data proves to be significantly
easier to classify - that is, based on information about a
player and his/her style, the roles and IDs of professional
players are much more predictable. As expected, the heroes
of the players in these matches are much easier to classify
because at a high skill level, the players act much more predictably and maximize the abilities of each hero much more
consistently. Furthermore, these matches are in a competitive
setting, free from any sort of distraction or perhaps apathy. In
contrast, players in the public realm are generally expected
to have a less consistent playing style, as their strategies are
less defined, and they may be more open to experimental
playing styles. Additionally, the range of player skill in the
public data is likely much larger than that of the professional
players, which can introduce varying common playing styles
at various levels of expertise. Comparing the data, we find
that the classification accuracy is consistently better for hero
roles than it is for hero IDs, which is certainly expected,
since there are only 3 roles, but over 100 actual heroes.
TABLE I
C LASSIFICATION R ESULTS FOR TI3 (T HE I NTERNATIONAL 3) M ATCHES
Hero Roles
Hero IDs
Logistic Regression
89%
72%
Random Forest
88%
80%
While we are able to reliably classify hero roles with high
accuracy, the most surprising results are the accuracy rates
for hero IDs. Given that there are over 100 unique hero
types, we would initially expect the accuracy of a random
guess to be lower than 1%. In contrast, there are only three
hero roles, which means a random guess should have 33%
accuracy if we are classifying hero roles. Obtaining a 72%
accuracy rate for hero IDs in professional matches is quite a
leap from the expected accuracy of ¡1% for a random guess.
However, it is important to note that the usage of heroes is
not distributed evenly, and only a smaller subset of heroes
are used the majority of the time. In professional matches
especially, where we only have 124 matches of data, this
can mean that we are generally only classifying a portion
of the heroes available. To be exact, there were only 77
unique heroes used in the professional matches we classified.
Furthermore, the top 28 most-used heroes made up for 76%
of all the heroes used in the matches. For limited data sets
like competitions, it may be in our future interests to only
include player data whose hero is used frequently enough to
provide enough training data. However, this issue is likely
V. E XPERIMENTAL R ESULTS
Through our tests, we were able to produce accurate
and fairly desirable results consistent with our expectations.
The highest classification accuracy obtained was 89% while
classifying amongst the three hero roles using logistic regression and data from professional matches obtained from
TABLE II
C LASSIFICATION R ESULTS FOR P UBLIC M ATCHES
Hero Roles
Hero IDs
Logistic Regression
84%
77%
Random Forest
85%
72%
Fig. 6. Confusion matrices for hero positional role classification of public
and professional data sets using logistic regression. The horizontal axis
shows predicted roles and the vertical axis shows actual roles. Elements
are shaded based upon frequency of classification.
Fig. 5. Confusion matrix for hero ID classification of public data set using
logistic regression. Predicted hero ID’s are shown along the horizontal axis.
Actual hero IDs are shown along the vertical axis. Elements of the matrix
are shaded based upon the frequency of the classification.
much less of a concern for the public matches in our data,
since we are able to train our classifier with 12,000 matches.
To better illustrate the accuracy of our classifications on
the testing dataset, information about our classifications are
displayed in the confusion matrix in Figure 5. The confusion
matrix illustrates three dimensions of information - the actual
role or hero ID of the player (vertical axis), the predicted
role or hero ID of the player (horizontal axis), and the
number/frequency of predictions for the combination of the
two (illustrated by color). As a reference, a 100% accuracy
rate would be illustrated by a single dark diagonal line
illustrating 100% predictions where the actual role is equal
to the predicted role.
In Figure 5 is the confusion matrix for classifying hero
IDs in public matches. The confusion matrix is not only
another way to view accuracy, but it also serves as a way
for us to identify common mis-predictions, indicated by dark
squares that do not fall on the diagonal line. Although the
confusion matrix above may be hard to read because of the
100+ hero IDs that need to be shown, it should be clear to see
by the varying darkness of the diagonal line that some hero
IDs are classified correctly much more often than others.
Additionally, in Figure 6, we can see two more confusion
matrices, this time showing hero positional role information
instead of hero IDs.
In Figure 6, not only can we see that the professional
matches have a much higher classification accuracy, but we
can also observe what kind of mis-classifications are more
common than others. In both graphs, there is noticeable
shading where the “Solo Lane” positional role meets the
“Carry” positional role. This is expected, because the player
style and traits of a “Solo Lane” hero often overlap with
those of a “Carry” hero. We also observe a slightly lighter
shading for where “Support” meets “Solo Lane”, and most
notably, we see that “Support” is rarely ever misclassified as
“Carry”, and vice versa.
In addition to comparing hero positional roles vs. hero
IDs and public matches vs. professional matches, we also can
compare our results for various subsets of our final feature set
to reveal more information about the game. In Table III, we
show classification accuracies for only professional matches,
using the random forest classifier.
In Table 3, we observe a few interesting details. First,
it appears that classification of hero roles works best with
items in particular, if we have to choose only one from
items, statistics, or abilities. Furthermore, we notice that the
classification accuracy doesn’t suffer greatly when limited to
only one of the three subsets of features. Second, we notice
that hero ID classification accuracy is significantly reduced
if we are forced to choose one of these three subsets of
features. Making these observations, we can deduce a few
things. Given that the classification accuracy for hero roles
only decreases by 0.04 from using all features to using just
the hero items as features, it is clear that these features
are the most important subset of features observed. We can
also draw similarities between a hero’s items and a sports
player’s physical abilities, since items can play a key role
in determining the key strengths of a hero in Dota 2, and
it is often the key strengths of a player in any team sport
that determine what kind of role or position he/she fills.
Additionally, we find that the classification of a hero ID
is very much dependent on all subsets of our feature set.
Since many hero IDs may fill the same role, or act and
perform similarly, we need as much data as possible to
distinguish between hero IDs. It is also worth noting that
using the ability upgrade information about a hero as the
only feature produces a much higher classification accuracy
than the rest of the feature subsets. Since each hero has a
completely unique set of abilities, and particular abilities for
TABLE III
C LASSIFICATION ACCURACIES FOR VARYING F EATURE S ETS
All Features
Performance Only
Items Only
Abilities Only
Hero Roles
0.88
0.73
0.83
0.76
Hero IDs
0.80
0.35
0.48
0.63
each hero are desirable earlier on in the game than other
abilities, this feature makes for a great starting point for hero
ID classification.
VI. D ISCUSSION AND C ONCLUSION
Our classification of Dota 2 heroes based on hero playing
styles is relatively accurate. Specifically, we observe that
the role of a hero can be predicted with 89% accuracy,
using logistic regression over various player features. Even
more surprising, we are able to obtain a 80% classification
accuracy for hero IDs - a massive improvement over a random guess, which would be expected to yield less than 1%
classification accuracy. The results of our study illuminate
the true team-playing nature of Dota 2, and help validate our
hypothesis that there are many parallels between traditional
team-playing games and certain eSports like Dota 2 that
make it possible to identify team player roles based on the
characteristics and actions of a player. Since eSports may
typically be considered completely separate and independent
from traditional sports, it is interesting to be able to reveal
some of these similarities. While our results primarily serve
to present interesting information to the public, they also
offer some applications for the Dota 2 community. Using a
trained classifier model built from professional match data,
we could evaluate the playing style of a more amateur player,
and provide useful feedback to that player about the way they
are using a particular hero, and how the player can improve
their playing style with that hero. Furthermore, an application
might classify the playing type of an amateur player in order
to provide the player with suggestions for heroes that they
would enjoy using and excel with in the future.
Overall we are pleased with our results, but we are still
limited by the data provided to us by Steam. For example,
information about player position on the game map over
time would likely be a useful addition to our feature set
for classification, but this data is not available to us from the
matches. Furthermore, this kind of study could be extended
to other games to open up new possibilities for applications,
such as online analysis for predicting the outcome of a game
or guiding a player to a desired outcome, or even using
classification data to improve the realism of AI (artificial
intelligence) players. While performing in-depth analyses on
eSports data may seem unimportant on the outset, it is clear
that there is plenty of interesting information to learn and
potentially useful applications for these analyses.
R EFERENCES
[1] V. Di Salvo, R. Baron, H. Tschan, F. J. Calderon Montero, N. Bachl, and
F. Pigozzi, “Performance characteristics according to playing position
in elite soccer,” International journal of sports medicine, vol. 28, no. 3,
p. 222, 2007.
[2] A. Gocentas, N. Jascaniniene, S. Poprzecki, J. Jaszczanin, and A.
Juozulynas, Position-related differences in cardiorespiratory functional
capacity of elite basketball players, J Hum Kinet, vol. 30, pp. 145152,
Dec. 2011.
[3] Y. Koklu, U. Alemdaroglu, F. U. Kocak, A. E. Erol, and G. Fındıko˘glu,
Comparison of Chosen Physical Fitness Characteristics of Turkish
Professional Basketball Players by Division and Playing Position, J
Hum Kinet, vol. 30, pp. 99106, Dec. 2011.
[4] D. C. W. Nicholas, “Anthropometric and Physiological Characteristics
of Rugby Union Football Players,” Sports Med, vol. 23, no. 6, pp.
375–396, Jun. 1997.
[5] G. Ziv and D. R. Lidor, “Physical Attributes, Physiological Characteristics, On-Court Performances and Nutritional Strategies of Female and
Male Basketball Players,” Sports Med, vol. 39, no. 7, pp. 547–568, Jul.
2009.
[6] G. Ziv and R. Lidor, “Physical characteristics, physiological attributes,
and on-court performances of handball players: A review,” European
Journal of Sport Science, vol. 9, no. 6, pp. 375–386, 2009.
`