Prediction of Euroleague Games based on Supervised Classification
Algorithm k-Nearest Neighbours
Tomislav Horvat
1
, Josip Job
2
and Vladimir Medved
3
1
Department of Electrical Engineering, University North, 104. brigade 3, Varaždin, Croatia
2
Chair of Visual Computing, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek,
Josip Juraj Strossmayer University of Osijek, Osijek, Croatia
3
Department of General and Applied Kinesiology, Faculty of Kinesiology, University of Zagreb, Zagreb, Croatia
Keywords: Basketball, Classification, Database Analysis, Euroleague, Information System, k-Nearest Neighbours,
Machine Learning, Prediction.
Abstract: Machine learning becomes one of the top fields in world of ICT (Information and communications
technology). In the last few decades machine learning has taken a big boost in sports. Sport fans are using
machine learning in sport betting and sport managers are using it for player selection, performance
evaluation and even in outcome prediction. This paper gives basketball outcome prediction algorithms using
k-nearest neighbours (k-nn). Few methods of preparing data calculated with different k and time period for
k-nn algorithm will be presented and closely explained. Feature selection will also be discussed and results
based on feature selection will be presented.
1 INTRODUCTION
The main goal of a machine learning algorithm is to
predict an outcome of the observed process.
Machine learning is a field of computer science
closely related to computational statistics with the
background in probability theory, linear algebra,
information theory and cognitive sciences. The
definition of machine learning comprises that goal of
machine learning is to build a model that can serve
as a good and useful approximation of data
(Alpydin, 2010). The area of machine learning can
be divided into supervised, unsupervised and
reinforcement learning (Jacobs, 2017). The main
difference between supervised and unsupervised
learning lies in the fact that in supervised learning
target variable should be specified. Based on the
type of target variable, supervised learning can be
divided into classification and regression. In
classification, the target variable is discrete value
while in regression target variable can take infinitive
number of numeric values (Harrington, 2012). On
the other hand, in unsupervised learning, there’s no
label or target value given for the data. The main
goal of unsupervised learning is to find regularity in
given data. Sports prediction is usually treated as a
classification problem by which one class is
predicted (victory or defeat), and rare cases are
predicted by numerical values. In this paper data
with a known outcome will be used, so problem will
be classified as a supervised classification problem.
In this paper, prediction model based on k-
nearest neighbours (later k-nn), one of supervised
classification method, will be presented. K-nn
method is easy to grasp and thus very effective,
usually gives high prediction accuracy, it is
insensitive to outliers and with no assumptions on
data, but also computationally very expensive and
requires a lot of memory (Harrington, 2012).
Because of its simplicity, k-nn method gives a
researcher many options and opportunities. Thus,
two algorithms for data preparation, with and
without feature selection, will be presented in this
paper, as well as various possibilities for defining
coefficient k and different time periods. Feature
selection methods are used to achieve
dimensionality reduction and exploration of model
simplification and acceleration execution algorithms.
For the purposes of research, an information
system called Basketball Coach Assistant (later
BCA) was built. BCA is a web application based on
database, built with PHP and MySQL technologies.
The application is supported by JavaScript and
Horvat, T., Job, J. and Medved, V.
Prediction of Euroleague Games based on Supervised Classification Algorithm k-Nearest Neighbours.
DOI: 10.5220/0006893502030207
In Proceedings of the 6th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2018), pages 203-207
ISBN: 978-989-758-325-4
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
203
jQuery on the client side. The first version of the
BCA information system, called AssistantCoach,
was presented at the international conference
icSports 2015 in Lisbon (Horvat et al., 2015).
The second chapter will give an insight into
previous research primarily intended for basketball,
not necessarily based only on predicting game
outcome. The data preparation and methods used for
predicting the outcomes will be discussed in the
third chapter. In the chapter Results, the results of
the research will be presented and explained more
closely. In the last chapter the conclusions are given
as well as the guidance for future research.
2 RELATED RESERCH
This chapter will provide some basic information on
current research and use of machine learning in
basketball. Some interesting conclusions and actual
results of prediction outcome from prior research
will be presented. Sports prediction has gained in
popularity because of availability of various
datasets, statistics, tools and applications. Different
researchers use various machine learning methods to
predict the outcomes. Most research in basketball is
based on the NBA league, while this research
focuses on the strongest European basketball league
called Euroleague.
Methods for basketball games prediction started
to develop in the second half of the 20th century.
According to research from 1979, if team scores are
0.7 points per possession, the one will win the game
(Shanahan, 1979). In 1997, the application named
Advanced Scout was introduced with the goal of
revealing interesting patterns in NBA game data
(Bhandari et al., 1997). At the symposium in
Budapest, a study was presented where authors were
using neural networks for analysing five seasons of
the second Serbian league and concluded that the
most influence on the final outcome have defensive
rebounds, then percentage for two point shots,
percentage for three point shots, stolen balls, lost
balls, offensive rebounds, percentage of free throws,
blocks and the slightest influence have assists
(Ivanković et al., 2010).
One of the best results in an outcome prediction
is shown in (Loeffelholz et al., 2009) where a
prediction model of the basketball games outcome is
presented in the NBA league using various types of
neural networks. The research also included methods
for feature selection. Compared to the experts’
predictions whose accuracy was 68.67%, the best
modelled neural network gave an accuracy of
74.33%. In the paper (Cheng et al., 2016), the
authors presented a classification model with
accuracy of 74.4% for prediction of NBA league
games by using Maximum Entropy.
From other results it is worth mentioning a paper
(Amorim Torres, 2013.) in which various machine
learning methods were used. The best prediction rate
was achieved using the Multilayer Perceptron
Method that achieved 68.44% prediction rate. The
Linear Regression has achieved a performance of
67.89% which was better than the likelihood method
that achieved a performance of 66.81%. In paper
(Miljković et al., 2010) Naive Bayes method is used.
Besides actual outcome prediction, for each game,
system calculates the spread, by using multivariate
linear regression. The system was evaluated on the
dataset of 778 games and it correctly predicted the
winners of about 67% of matches. As for the
prediction of spread, the system has successfully
provided predictions for 78 matches (10%).
3 DATA AND METHODS
Sufficient amount of relevant data is a basic
condition for building a good prediction model. It is
very important to well define the methods that will
be used. Two different data preparation methods and
two ways of using the k-nn algorithm with different
k coefficient will be explained in detail in this
chapter.
3.1 Data Acquisition
For the purposes of research, public statistics of
Euroleague seasons 2012/2013 to 2016/2017 were
used. 80% of games are used for model training and
20% for result evaluation, regardless of the time
period. Using web scraping process, structured data
from Euroleague official Web were extracted,
transformed and loaded into relational database
suitable for further analysis. Process of extracting,
transforming and loading is called ETL process. Due
to specificity of data retrieval, Web scraper in
scripting programming language PHP has been
programmed. The web scraper passes through the
domain www.euroleague.net, extracts data from
page, transforms them if necessary and stores them
into relational database. As was previously stated,
sufficient amount of relevant data is a basic
condition for building a good prediction model and
that is exactly the task of the web scraper.
K-BioS 2018 - Special Session on Kinesiology in Sport and Medicine: from Biomechanics to Sociodynamics
204
3.2 The k-nn Method
As mentioned above, the k-nn method is selected
because it is easy to grasp and thus very effective,
but also computationally very expensive. This is a
non-parametric method used for classification and
regression. The method fundamentally relies on a
metric distance value. The most common metric,
used in this research also, is Euclidian distance, but
there are other metrics that can be used. Euclidian
distance is shown in equation (1).
(1)
In this paper, two data preparation algorithms, both
Euclidian distance based, will be compared and
tested for different k values as well as for different
dataset lengths, i.e. time periods. The main goal is to
find the optimum k and the time period for
predicting games outcome.
3.3 Data Preparation Variants
As mentioned above, two variants of data
preparation for the k-nn method were used. All
variants of data preparing are based on basic
basketball statistics which are shown in Table 1. Due
to the specificity of each team, the outcome
predication is based only on the observed team
games.
Table 1: Basic basketball statistics elements.
Name/Abbrev.
Full name / Explanation
2fgm, 2fga
2 field goals made / attempts
3fgm, 3fga
3 field goals made / attempts
ftm, fta
free throws made / attempts
dr, or
defensive / offensive rebounds
as, st, to
assists, steals (stolen balls),
turnovers
bl, bl_against
blocks made by player / against
player
f, f_drawn
fouls committed / drawn
madeFG_FT
made field goals and free throws
missFG_FT
missed field goals and free throws
First variant is called DefenseOfense. The goal of
that variant is to group defensive and attacking
elements, thus separating the defensive from the
attacking part. Two components, I
ofense
and I
defense
,
were calculated and later used in the outcome
prediction. Equations (2) and (3) show how the
components are calculated. Index r (rival) refers to
opponent / rival team.
(2)
(3)
Euclidian distance based on every game in test
period and game in evaluation period is calculated
and then, based on user defined parameter k
predicted outcome is calculated. Parameter k is
always defined as odd number to avoid the inability
of defining game outcome. The same principle of
calculation is used in another data preparation
variant, but unlike in DefenseOfense more elements
are used.
The purpose of the second variant, called
Components. is to group the logical elements of the
game (defensive and offensive rebounds, stolen balls
and turnovers…) and thus obtain a multidimensional
problem. List of logical groups and their calculation
is shown in Table 2.
Table 2: Logical components groups.
Name
Explanation
points
madeFG_FT – missFG_FT
rebounds
dr + or
steals_turnovers
st - to
assists
as
blocks
bl – bl_against
fouls
f – f_drawn
Both algorithms (DefenseOfense and Components)
were also compared after applying the feature
selection process. Dimensionality reduction (van der
Maaten et al., 2009) or in computing feature
selection (Jović et al., 2015) is a very important step
in estimating output value because significantly
determines the algorithm predictability. The
selection process can be applied in several ways
depending on the goal, the available resources, and
the desired level of optimization. In this research,
the feature selection is performed by measuring the
information gain for each feature (Hoque et al.,
2014). Information gain is a measure of an amount
of information whose specific feature contributes to
Prediction of Euroleague Games based on Supervised Classification Algorithm k-Nearest Neighbours
205
the class. In this case specifically, it is a measure
that shows how much each element of basketball
statistics (Table 1) contributes to the final game
outcome. Feature selection will be defined based on
the average information gain and on zero based
information gain. If the information gain of
statistical component is greater or equal to the
average value of the information gain, component
will be used in the outcome prediction process. Zero
based information gain means that features will be
included into prediction model only if their
information gain is greater than zero. Final feature
subset which will be used in the testing phase of the
algorithm will be generated by a union of feature
subsets defined for the observed and rival team.
4 RESULTS
The research results, based on two different data
preparation variants, each of them defined with
different coefficients k, different number of analysed
seasons and different subset of features are presented
in Table 3.
Table 3: Prediction algorithm results (DefenseOfense).
Feature
selection
Seasons
Accuracy (%)
k=3
k=5
k=7
k=9
1
80,19
83,96
83,96
83,02
3
83,39
79,55
80,51
80,83
5
80,66
79,88
78,53
79,11
inf. gain >
0
1
80,19
83,96
83,96
83,02
3
83,39
79,55
80,51
80,83
5
80,66
79,88
78,53
79,11
inf. gain >
average
value
1
75,47
72,64
71,7
74,53
3
72,2
70,29
72,2
70,61
5
74,27
73,89
75,63
75,63
Table 4: Prediction algorithm results (Components).
Feature
selection
Seasons
Accuracy (%)
k=3
k=5
k=7
k=9
1
70,75
71,7
73,58
68,87
3
76,36
77,32
75,4
74,44
5
71,95
73,11
73,11
76,6
inf. gain >
0
1
70,75
71,7
73,58
68,87
3
76,68
77,32
75,4
74,44
5
71,57
72,92
73,11
76,6
inf. gain >
average
value
1
61,32
61,32
61,32
57,55
3
57,51
61,02
59,11
61,34
5
60,93
62,67
60,93
62,09
As can be seen from result in Table 3 and Table
4 better results are achieved with using
DefenseOfense data preparation variant. It is
noticeable that coefficient k does not have
significant impact on model accuracy while feature
selection process and grouping of input features has.
Much more impact on model accuracy has the
feature selection and defining of logical subgroups
of input features. Both algorithm variants use a set
of 15 features. The first variant, DefenseOfense,
divides features into two logical subgroups, while
the second variant, called Components, divides input
features into six logical subgroups. From all this, it
is easy to notice that dividing no excessive large
input set of features on even smaller subgroups does
not yield better results (speaking about k-nn
classification method). This is an example where all
the features are significant and each of them
contributes to the model accuracy. Excluding
features with information gain equal to 0 does not
produce better results, but it reduces the algorithm
execution time. Feature selection variant that
excludes all features with information gain less than
the average information gain for observed team gave
worse results. The conclusion is that there are still
too few features that would lead to a good
description of the process (in this case process is a
basketball game). The problem of lower model
accuracy, in case where features are selected based
on the average information gain, occurs in both
cases, but still gives good results. Better results of
the prediction would certainly be provided with a
greater number of input features. In that case, feature
selection would get greater significance and would
also accelerate algorithm execution speed.
Compared with prior research, the proposed data
preparation variant DefenseOfense gives improved
results. Most of the prior research is focused on
predicting NBA league outcomes, and typical
dataset is consisted of 5 seasons, but additional
difference between two leagues is the fact that the
NBA league has as twice as many games per season
than in Euroleague. Despite the fact that in this
calculation dataset with less records is used, this
approach achieved better results, but the method
should be tested on the same dataset before drawing
conclusions.
5 CONCLUSIONS
In this paper, traditional k-nn algorithm on game
prediction is presented. Based on the analysis, two
data preparation methods have been proposed.
K-BioS 2018 - Special Session on Kinesiology in Sport and Medicine: from Biomechanics to Sociodynamics
206
Sport predictions present challenging and
interesting field because there is no universal
algorithm that solves prediction problem for every
sport. In addition to expert knowledge related to ICT
technology, there is also a need for expert
knowledge related to the analyzed sport.
According to the presented results, it can be
concluded that the k-nn is a very convenient method
for solving outcome prediction problems in sport.
The combination of k-nn classification method and
feature selection is also an interesting field of
research. Various approaches have been tried up to
this research. An approach that might give better
results is to define the feature weight based on the
information gain when calculating Euclidean
distance (Sun and Huan, 2010), with or without
feature selection. A possible formula for such
research to calculate Euclidian distance based on
feature information gain is:
(4)
Although there is no universal algorithm for
predicting game outcomes, there is no reason not to
try to find the sport specific one, hence in addition to
knowledge of ICT technologies and machine
learning research, advanced knowledge of the
observed process, in this case basketball, is also
required.
REFERENCES
Alpydin, E., 2010. Introduction to Machine Learning:
Second Edition, Cambridge, Massachusetts, London,
England.
Amorim Torres, R., 2013. Prediction of NBA games based
on Machine Learning Methods, University of
Wisconsin Madison.
Bhandari, I., Colet E., Parker, J., Pines, Z., Pratap, R.,
Ramanujam, K., 1997. Advanced Scout: Data Mining
and Knowledge Discovery in NBA Data. Data Mining
and Knowledge Discovery, 1, 121-125.
Cheng, G., Zhang, Z., Kyebambe, M.N., Kimbugwe, N.,
2016. Predicting the Outcome of NBA Playoffs Based
on the Maximum Entropy Principle. Entropy, 18 (12),
1-15.
Harrington, P., 2012. Machine Learning in Action,
Manning Publications, Shelter Island.
Hoque, N., Bhattacharyya, D.K., and Kalita, J.K., 2014.
MIFS-ND: A mutual information-based feature
selection method, Expert Systems with Applications,
vol. 41, issue 14, pp. 6371–6385.
Horvat, T., Havaš, L., Medved, V., 2015. Web Application
for Support in Basketball Game Analysis. icSports
2015, Lisboa, 225-231.
Ivanković, Z., Racković, M., Markoski, B., Radosav, D.,
Ivković, M., 2010. Analysis of basketball games using
neural networks. Proceedings of the 11th International
Symposium on Computational Intelligence and
Informatics, Budapest, 251-256.
Jacobs, S., 2017. The Ultimate Guide to Machine Learning
(Neural Networks, Random Forests and Decision
Trees), Amazon Distribution.
Jović, A., Brkić, K., Bogunović, N., 2015. A review of
feature selection methods with applications.
Information and Communication Technology,
Electronics and Microelectronics (MIPRO).
Loeffelholz, B., Bednar, E., Bauer, K.W., 2009. Predicting
NBA Games Using Neural Networks, Journal of
Quantitative Analysis in Sport, 5(1), article 7.
Miljković D., Gajić Lj., Kovačević A., Konjović Z., 2010.
The use of data mining for basketball matches
outcomes prediction. IEEE 8
th
International
Symposium on Intelligent Systems and Informatics,
Subotica, Serbia.
Shanahan, K.J., 1979. An analysis of game statistics in
basketball, Masters thesis. Faculty George Williams.
Sun, S., Huan, R., 2010. An Adaptive k-Nearest Neighbor
Algorithm. 7th International Conference on Fuzzy
Systems and Knowledge Discovery (FSKD 2010).
Van der Maaten, L., Postma, E., van den Herik, J., 2009.
Dimensionality Reduction: A Comparative Review,
Tilburg Centre for Creative Technology, Tilburg
University.
Prediction of Euroleague Games based on Supervised Classification Algorithm k-Nearest Neighbours
207
