Journal of Physics: Conference Series

PAPER • OPEN ACCESS Related content

- Application of Data Mining techniques to
A Comparative Study of Data Mining Techniques relate Cardiovascular Risk and Coronary
Calcium
on Football Match Prediction F N Lujan, L J Cymberknop, M Alfonso et
al.

- A Comparative Study to Predict Student’s

To cite this article: Che Mohamad Firdaus Che Mohd Rosli et al 2018 J. Phys.: Conf. Ser. 1020 Performance Using Educational Data
012003 Mining Techniques
Annisa Uswatun Khasanah and Harwati

- The comparative study for the isotropic

and orthotropic circular plates
C Popa and G Tomescu
View the article online for updates and enhancements.

This content was downloaded from IP address 156.209.143.16 on 13/07/2021 at 04:03

ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

A Comparative Study of Data Mining Techniques on

Football Match Prediction
Che Mohamad Firdaus Che Mohd Rosli1 , Mohd Zainuri Saringat1 ,
Nazim Razali1 , Aida Mustapha1
1
Faculty of Computer Science and Information Technology, Universiti Tun Hussein Onn
Malaysia, 86400 Parit Raja, Batu Pahat, Johor, Malaysia
E-mail: [email protected], [email protected], [email protected],
[email protected]

Abstract. Data prediction have become a trend in today’s business or organization. This
paper is set to predict match outcomes for association football from the perspective of football
club managers and coaches. This paper explored different data mining techniques used for
predicting the match outcomes where the target class is win, draw and lose. The main objective
of this research is to find the most accurate data mining technique that fits the nature of football
data. The techniques tested are Decision Trees, Neural Networks, Bayesian Network, and k -
Nearest Neighbors. The results from the comparative experiments showed that Decision Trees
produced the highest average prediction accuracy in the domain of football match prediction
by 99.56%.

1. Introduction
In data generalization, there are two major topics that are concentrated on this subject; Data
Mining and Statistics. According to [1], the reason that prediction approaches have been widely
used in many companies around the world today is because prediction is the heart of remarkable
disciplines in Science. This approach enables companies and organizations to predict and prepare
reasonable future plans. One particular applications of prediction is in sport. In order to forecast
the future of a particular organization, the logic of prediction is employed. In sport analytic for
association football, match prediction is used to predict the matches before it started using a
particular technique so that the managers and players able to forecast the possibilities of either
win, draw, or loss of the game before it started. Positive outcome will definitely boost up player’s
spirit but if the outcome is negative, it does not mean to lowered player’s sportsmanship but as
a guide for them to play extra cautious and provide an appropriate counter-attack.
There are already many researches on the prediction of match outcome in sports, in basketball
and football particularly. However, most of those researchers analyzed and predict using a single
technique only. There are research that predicts using different techniques but the result might
be biased since there are techniques that using a particular technique to increase the accuracy
of the prediction. For example, [2] predicted football and basketball using Least Square while
[3] predicted football outcome using Neural Tuning and Genetic Tuning fuzzy model. Both
researches showed how to match using different techniques with a specific technique create by
researchers.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

To address this gap, a comparative analysis of several data mining techniques using only raw
data is essential to find the most accurate techniques using only raw data without any elements
that could affect the result like derivation of data or particular technique. This paper is set to
achieved three objectives; (1) prediction of football match result (win, lose and draw) using raw
data of previous match outcome, (2) model generation of those results using several techniques
in data mining, and finally (3) comparative analysis to observe the result of accuracy metrics.
This research is limited to previous match’s data of three seasons of English Premier League
from 2013-2014, 2014-2015, and 2015-2016 totaling 1,140 matches of both home and away.
The rest of the paper is organized as follows: Section II describes the related work on
profiling and classification models. Section III presents the method uses in this paper to produce
comparative analysis results. Section IV presents the experimental results. Finally, Section V
concludes the work and highlights a direction for future research.

2. Related Work
Numerical prediction is a method where a continuous-valued function or ordered-value are
predicts by the model constructed. This particular model is also known as predictor. The
accuracy of the predictor depends on how well the provided predictor able to predicts the value
of predicted attribute for a new data. When compared to classification methodology in data
mining, predicted values in prediction methods are usually continuous whereas classifications are
discreet, making a classification problem seen as a predictor of classes. Prediction are normally
about predicting the future and classification are about classifying the present. Machine learning
approaches have been used in predicting match outcome in sports especially for basketball and
football association. This prediction is carried out by generating past match results. In this
work, four standard prediction algorithms are used for comparative purposes, which are Decision
Trees, Neural networks, Bayesian networks, and k -Nearest neighbors.
A decision tree is a structure that includes a leaf nodes, branch nodes, and root nodes. Each
leaf node holds a class label, each internal node denotes a test on an attribute, and branch
denotes the outcome of a test. The topmost node in the tree is the root node. Decision tree
model was used in predicting football matches outcome by [4]. In their experiment, the decision
model was developed using the Machine Learning package in order to generate MC4 decision
tree. MC4 decision tree is different than the original because it used the tree pruning technique
in order to reduce the size and the complexity of the tree. However, MC4 decision tree only
managed to predict 45.77% and 41.72% in accuracy for both expert model data and general
model data.
Neural Network is a parallel distributed processor that has a propensity for storing
experiential knowledge and making it available for users [5]. Neural computing is the study
of networks of adaptable nodes which through a process of learning from task. According to [6],
neural network model was used to predict the match outcome between two teams during World
Cup in 2006. Technically, they proposed a supervised multi-layer perceptron neural network
(MLP) with error propagation learning rule (BP) to predict the winning rate for the teams.
Then, they compared the result in order to determine the relationship between win and loss.
The accuracy for this model was 76.9% excluding the tied games due to limitation in predicting
draw outcome.
Belief networks or Bayes nets for short, were the other name for Bayesian Network. It belong
to the probabilistic graphical models. These models can be used to represent knowledge about
uncertain amount. In particular, the graph contains nodes and edges where nodes represent
a random variables and the edges represent the relation of probabilistic dependencies among
the corresponding random variables. This Bayesian Network model was used by [7] to predict
football outcomes using the Skellam’s distribution. This paper focuses on the goal differences
instead of the number of goals like most of researchers. The research was made to a specific

2
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

team with observed number point greater than 1.95 or probability of observed outcome lower
than 20%. For instance, the prediction accuracy of win for Bolton team is 53%.
The k-Nearest Neighbors method begins in the early 1950s. The method is labor-intensive
when given large training sets, and did not gain popularity until the 1960s when increased
computing power became available. Nearest Neighbors are based by comparing a given test tuple
with training tuples that are similar to it. [4] used the IB classifier from the family of k -Nearest
Neighbors (KNN) in the MLC++ library for predicting matches outcome for Tottenham Hotspur
Football Club for 2 seasons. The KNN technique used likeness approach, where a Euclidean
distance is measured and plotted on and n-dimensional graph where each dimensional is one of
the supplied attributes. KNN technique produced prediction accuracy at 50.58% and 47.21%
for two seasons.

3. Methodology
In this section, the dataset, the classification techniques, and the performance measures
are presented and discussed. Experiments for Decision Tree, Bayesian Network, and k -
Nearest Neighbors are implemented using Weka 3.8 [8] while the Neural Networks algoritm
is implemented using the GMDH Shell DS (https://www.gmdhshell.com/). In Weka 3.8, the
tool uses 10-fold cross-validation method to validate the data. As for classifier, Decision Tree
uses J48 classifier while Bayesian Network uses BayesNet classifier. IBk will be the classifier
appointed for k-Nearest Neighbors technique. In GMDH Shell DS, the Neural Network technique
is implemented using the cross-validation criterion PRR to validate the data.

3.1. Dataset
The dataset chosen to be implement in this research was the data of the English Premier League
of 2013 to 2016 season. English Premier League consisted of 20 teams played both home and
away totaled 380 matches throughout the season. This is mean there are approximately 1,140
matches for three seasons of English Premier League. The data consists details of the number
of full time home and away goal, full time result (win, draw or lose), the number of home and
away shots, the number of home and away shots on target and the number of home and away
corner. The data was retrieved from the website http://www.football-data.co.uk. There are
eight features extracted from the datasets, which are Full Time Home Goal, Full Time Away
Goal, Home Team Shot, Away Team Shot, Home Shot On Target, Away Shot On Target, Home
Corner, and Away Corner while Full Time Results which are win, draw or lose as target class
for prediction.

3.2. Evaluation metrics

The performance measures used in this paper is the accuracy. Overall, accuracy measures how
often is the classifier or the classification techniques (Decision Trees, Neural Networks, Bayesian
Networks, k -Nearest Neighbor) getting correct target class of classification. Equation 1 shows
how accuracy is measured.
TP + TN
accuracy = (1)
P +N
From the equation, T P represents true positives or equivalent to hit, T N represents true
negatives or equivalent to correct rejection, P is the total number of real positive cases in the
data, and finally N is the total number of real negative cases in the data. In context of football
match results, its refer to win, draw or lose target class. For example, if the predicted result
of a match hit win but the actual result is lose the result will not be counted in T P and T N .
However, it will counted in P which it is actually false positve. Literally, its can be explained

3
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

that the total number of correct predicted match results (win, draw or lose) are divided with
total number of matches.

4. Results
The experiment generated the percentage of accuracy for each technique. All four techniques
have been implemented using the specific tool to acquire prediction data. The result were
recorded and analyzed. A comparison of the results collected have been made to observe the
most accurate technique to predict football match outcome. In order to validate the data, 10-fold
cross-validation techniques have been choosen.

4.1. Decision trees

The data retrieved have been imported into Weka 3.8. Then, the data are run using Decision
Tree technique. In this case, Weka 3.8 uses Decision Tree J48 Classifier which are also called
C4.5 Decision Tree to generate the result. There is no need to discretize the data because C4.5
Decision Tree can handle numeric attributes. The summary of the result is shown in Figure
1. The summary report shows that for 15/16 seasons of English Premier League matches, the
accuracy percentage generated was 99.47%.

Figure 1. Summary Report of J48 Decision Tree of 15/16 Seasons

4.2. Neural networks

Different tool called the Group Method of Data Handling (GMDH) has been used for running the
football prediction using Neural Networks. The same validation techniques have used in order
to generate prediction accuracy for Neural Networks technique to be compared to accuracy
of Decision Tree, Bayesian Network, and k -Nearest Neighbors, which were all implemented in
Weka 3.8. The result of Neural Networks is surprisingly accurate with 100% accuracy throughout
380 matches. Nevertheless, this result does not make Neural Networks as the best technique
because the the preduction accuracy is averaged out to take into account performance across
three seasons. Figure 2 shows the result for 2015-2016 season.

4
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

Figure 2. Twice-Multilayered Neural Network of 15/16 Seasons

4.3. Bayesian networks

Weka 3.8 was used to generate model using Bayesian Network technique as well as Decision Tree.
The data set of English Premier League of 13/14 seasons to 15/16 seasons were then implemented
into Weka 3.8 to generate prediction data. The summary report of Bayesian Network of 15/16
Seasons in Figure 3 shows 75.53% of accuracy from 380 matches throughout the seasons.

Figure 3. Summary Report of Bayesian Network of 15/16 Seasons

4.4. k-Nearest neighbors

The k -Nearest Neighbors technique was also applied to the football dataset by using Weka
3.8. The experiment used IBk classifier with 10-fold cross-validation technique similar to J48
Decision Tree and BayesNet Bayesian Network algorithms. Figure 4 shows the summary report
generated by Weka 3.8 of IBk k-Nearest Neighbors technique. The result produced accuracy of
78.16% for 2015-2016 season.

5
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

Figure 4. IBk k-Nearest Neighbors of 15/16 Seasons

4.5. Comparative results

The comparative results are shown in Table 1. The results using Neural Network technique
generated by GMDH is surprisingly accurate with 100% accuracy throughout 380 matches for
season 2014-2015 and 2015-2016. However, the technique still does not suggested as the best
since the average accuracy of three seasons been taken into account. The comparative results
shown that Desicion Trees have the highest average accuracy by 99.56% follow by Neural Net-
works and k-Nearest Neighbors technique by 96.83% and 77.54% while Bayesian Networks have
the worst average accuracy by 76.41%.

Table 1. Comparison of Accuracy Percentage

Decision Neural Bayesian k -Nearest
Accuracy Trees Networks Networks Neighbors
Season 2013-2014 99.74% 90.50% 77.37% 78.68%
Season 2014-2015 99.47% 100.00% 76.32% 75.79%
Season 2015-2016 99.47% 100.00% 75.53% 78.16%
Average 99.56% 96.83% 76.41% 77.54%

5. Conclusions and Future Works

This study compared four standard data mining algorithms in football match outcome
prediction. The experimental results proved that Decision Tree is the most accurate technique
to predict football match outcome with 99.56% accuracy as compared to other techniques, which
are Neural Network (96.83%), Bayesian Network (76.41%), and k -Nearest Neighbors (77.54%),
respectively. However, the research can be improved by studying and creating the same technique
for each technique to increase the accuracy of results by using other attributes that correlated to
match outcome besides the attributes that have been used in this research such as the distance
travelled by away team or current performance of competing teams in the league or the number
of yellow or red cards fouled by player which offcourse influece the football match outcome.
Moreover, the techiques used in this research is static over time because it depend on past

6
ICoTSM IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1020 (2018)
1234567890 ‘’“” 012003 doi:10.1088/1742-6596/1020/1/012003

match data while the real time football match statistic during match also need to be measured
and taking account for prediction (dynamic).
In any sports, knowing the possibilities of winning a game is essential. Result of outcome
prediction can be used to elevate player’s sportsmanship or to prepare a counter-attack against
their opponents. Match outcome prediction should become a trend in our national sports in
order to produce a profitable and quality organization of a particular sports while increases
athletics’ performances. Management of a particular sports also can used this prediction method
to forecast the future of an organization.

Acknowledgements
This project is sponsored by the Malaysian Ministry of Higher Education (MOHE) under the
Research Acculturation Grant Scheme vot 1513.

References
[1] Dhar V 2013 Communications of the ACM 56 64–73
[2] Stefani R T 1977 IEEE Transactions on systems, man, and cybernetics 7 117–121
[3] Rotshtein A P, Posner M and Rakityanskaya A 2005 Cybernetics and Systems Analysis 41
619–630
[4] Joseph A, Fenton N E and Neil M 2006 Knowledge-Based Systems 19 544–553
[5] Rumelhart D E, Hinton G E and Williams R J 1988 Cognitive modeling 5 1
[6] Huang K Y and Chang W L 2010 A neural network method for prediction of 2006 world
cup football game Neural Networks (IJCNN), The 2010 International Joint Conference on
(IEEE) pp 1–8
[7] Karlis D and Ntzoufras I 2009 IMA Journal of Management Mathematics 20 133–145
[8] Witten I H, Frank E, Hall M A and Pal C J 2016 Data Mining: Practical machine learning
tools and techniques (Morgan Kaufmann)