I.J.

Intelligent Systems and Applications, 2019, 7, 20-28

Published Online July 2019 in MECS ([Link]
DOI: 10.5815/ijisa.2019.07.03

Football Match Prediction with Tree Based

Model Classification
Yoel F. Alfredo
Computer Science Department, BINUS Graduate Program-Master in Computer Science,
Bina Nusantara University Jakarta, 11480, Indonesia
E-mail: yoel.alfredo001@[Link]

Sani M. Isa
Computer Science Department, BINUS Graduate Program-Master in Computer Science,
Bina Nusantara University Jakarta, 11480, Indonesia
E-mail: [Link]@[Link]

Received: 26 January 2019; Revised: 27 February 2019; Accepted: 19 March 2019; Published: 08 July 2019

Abstract—This paper presents the football match the opposing team. The involvement of an expert is the
prediction using a tree-based model algorithm (C5.0, common solution, however that does not mean it is
Random Forest, and Extreme Gradient Boosting). unflawed. Human assistance as a tool in making decisions
Backward wrapper model was applied as a feature is not always reliable. Some psyche factors can influence
selection methodology to help select the best feature that the analysis of the outcome and cause the decisions that
will improve the accuracy of the model. This study used are made to be inappropriate.
10 seasons of football data match history (2007/2008 – In addition to relying on experts as a solution in
2016/2017) in the English Premier League with 15 initial helping decision making, several studies have been
features to predict the match results. With the tuning carried out in predicting the results of football matches
process, each model showed improvement in accuracy. [2,3,4,5,6]. Some studies have output limitation of
Random Forest algorithm generated the best accuracy football matches [2,4], the algorithm used is logistic
with 68,55% while the C5.0 algorithm had the lowest regression which only gives 2 output ‘home win’ or
accuracy at 64,87% and Extreme Gradient Boosting ‘away win’ while football match has 3 result possibilities,
algorithm produced accuracy of 67,89%. With the output ‘home win’, ‘away win’, or ‘draw’. Another study has
produced in this study, the Decision Tree based algorithm been conducted using the random forest algorithm [3],
is concluded as not good enough in predicting a football but the prediction accuracy is only 63,4%.
match history. Based on several studies conducted in making
predictions in other fields, in this study the authors
Index Terms—Football match prediction, supervised proposed a decision tree as the method of choice in
machine learning, decision tree, feature selection, predicting the end result of a football match. A sample
classification. research that uses a decision tree is the initial prediction
of a heart disease. The algorithms used in this study are
CART, ID3 and decision table, with accuracy values of
I. INTRODUCTION 83.49% using CART algorithm, 72.93% using ID3
algorithm, and 82.50% using decision table algorithm [7].
Football is a very popular sport among people of
Other studies were conducted to detect bad data using a
various ages and genders. This sport has been enjoyed
decision tree with an accuracy rate of 94% [8]. The use of
since the times of ancient Egypt, where it was played by
the XGBoost method carried out in generating predictions
kicking a ball formed from a collection of linen fabrics
of bank failures in America can produce up to 94.74%
[1]. The English Premier League is one of the leagues accuracy [9].
most favored by the public. Based on statistics from This study will use the C5.0, Random Forest and
ESPN on season 2017/2018, there are as many as 14,56
XGBoost algorithms from the decision tree method to
million viewers who attend these football matches. With
predict the outcome of a football match in the English
the high enthusiasm from the community towards football,
Premier League. With larger datasets and feature
the media and the community often try to predict of the selection method using backward wrapper method, it is
outcome of a football matches. expected to improve the prediction accuracy from
Presently, football is not only a means of a sport but
previous studies that have been done.
also a form of entertainment and investment. In investing,
This paper will be divided into five sections, including
an investor needs help from an expert to determine the
the introduction section. Section II describes the related
right decision to invest. Even so with a football coach,
works. Section III describes the research methodology
they need need guidance in developing a strategy to face

This work is open access and licensed under the Creative Commons CC BY 4.0 License. Volume 11 (2019), Issue 7
 Football Match Prediction with Tree Based Model Classification

used in this research. Result and analysis will be artificial neural network and logistic regression with
described in section IV. Conclusions of this research knowledge discovery in database framework with 18
drawn in section V. features which consists of goals, shots, corner, odds,
attack strength, player’s performance index, manager
performance index, managers win and streak for each
II. RELATED WORKS home and away team. Artificial neural network method
has a prediction percentage of 75.04% and it improved to
Since football match prediction have become 85% by giving weight to the features. On another hand,
commonplace for people of all ages, several studies were the logistic regression method has a high accuracy
conducted to find out the criteria associated with football
percentage, 93%, but limited due to the results only
match result and the model that can generate the highest
obtain 2 values. Igiri [5] have another research for
accuracy in predicting football match result. These are
football match prediction using the SVM method with 15
following studies that have been conducted in order to features and reaching prediction percentage of 53.3%.
find the optimal model for football match predictions. Bayesian network method also used by Razali [Link]. [6] to
One study was done by Prasetio and Harlili [2] using
predict football match. They use shots, shots on target,
the logistic regression method and home offense, home
corners, fouls committed, yellow cards, red cards, half-
defense, away offense and away defense as their model
time goals, and full-time goals as features to build the
features with prediction numbers reaching 69.5%. A
model. In this research, the model prediction accuracy is
weakness in predicting the results of a soccer match using 75.09%.
the logistic regression method is that the results obtained
only have 2 values, where in reality, the results of a
soccer match can produce 3 final scores, namely win,
III. RESEARCH METHOD
draw or lose. Yezus [Link]. [3] produces predictions of
football matches using machine learning with K-nearest Fig. 1 will display a design experiment for this
neighbors and random forest algorithms, which use form, research. To predict football match results, this research
concentration, motivation, goal difference, score will have two major steps: data preprocessing and
difference, and history as key features to develop classification, with each step broken down into more
classifier and had accuracy values of 55.8% and 63.4%, detailed steps. This research will use R as a programming
respectively. Igiri and Nwachukwu [4] also conducted language to develop the model.
another research on football match predictions by using

Fig.1. Design experiment.

into two categories: football match statistics and

A. Data Collection
bookmaker odds prediction. This research will only use
Dataset used in this research comes from football- football match statistics to develop a model to predict
[Link] which is a common dataset to be used in football match results. In addition, the feature selection
conducting research in football match predictions. The process will be done by using the backward wrapper
data used comes from 10 seasons of English Premier model to determine which attributes will be used for
League matches from the 2007/2008 season to the creating model.
2016/2017 season. The total number data used for these
B. Preprocessing
entire study is 3800 historical match data consisting of
380 matches per season. Each season, 20 football team Dataset used in this research still needs to be cleaned.
participated in the English Premier League and each team There are some attributes which have no value for several
acted as a host or a guest. seasons. In this process, the bookmaker odds and
The dataset has 71 attributes that will be cleaned in the irrelevant data which have no impact in the model
preprocessing step. From 71 attributes, we can divide it development will be removed such as match date, referee

Volume 11 (2019), Issue 7 21

 Football Match Prediction with Tree Based Model Classification

name, and football team name. Attribute FTR (Full Time 1. Feature Selection
Result) will be used as the label of the model.
Feature selection is a process that is commonly used by
Table 1. Initial features
researchers to get a smaller part of the entire data. This
smaller set of data is chosen based on relevance, to
Features Data Scale Description improve the performance of the model created in the
Full-time result. Used as the label of training process [10]. This feature selection process will
FTR Nominal
the model
Number of goal for the home team in be very influential in the model training stage. From
HTHG Ratio initial feature candidates, the feature selection process
half time
HTAG Ratio
Number of goal for away team in will select some features that have relevance and process
half time data with the selected features to form the model.
Number of shot done by the home
HS Ratio
team
In this research, backward wrapper model will be used
AS Ratio Number of shot done by away team as a feature selection methodology. The classifier will act
Number of shot on target is done by as a black box and will use random forest algorithm as
HST Ratio
the home team the algorithm for the classifier. Features that have been
Number of shot on target is done by evaluated will be assessed for their performance by
AST Ratio
the home team
Number of fouls done by the home
utilizing the classifier. This will be done until all the
HF Ratio features have been tested. Feature with the highest
team
AF Ratio Number of fouls done by away team estimated value will be used in the development of the
Number of corners obtained by the classifier model. On this method, Random Forest will be
HC Ratio
home team used as a classifier method. Fig. 2 display the framework
Number of corners obtained by away of wrapper models [11].
AC Ratio
team
Number of yellow cards obtained by C. Classification
HY Ratio
the home team
Number of yellow cards obtained by Classification is one of the predictive methods of data
AY Ratio
away team mining. The purpose of classification is to get a
Number of red cards obtained by the
HR Ratio
home team prediction of future value based on other variables
Number of red cards obtained by contained in the dataset. The classification has four
AR Ratio
away team fundamental components: class, predictors, training
datasets and testing datasets [12]. This research will use
From a total 14 attributes in the initial feature, feature supervised machine learning (classification) technique as
selection will be done to only select the best attributes. a prediction method.
Only those that have the potential to have a good impact
on prediction and result accuracy, will be included in the
model development.

Fig.2. Wrapper model framework.

model. Data set will be divided with composition 80:20.

1. Data Partition
80% of the data will be used as training data and the other
Data that have been processed through feature 20% will be used as testing data.
selection will be divided into training data and testing
2. Cross-Validation
data. Training data will be used to build the model and
testing data will be used to test the performance of the Cross-validation is a statistical method to evaluate and

22 Volume 11 (2019), Issue 7

 Football Match Prediction with Tree Based Model Classification

compare algorithms in the learning process by dividing each node randomly sample mtry of the predictors
the data into two parts. The first part is used for the and pick the best split among other variables.
learning process of the model and the second part is used • Predict new data by aggregating the predictions of
to validate the model [13]. the ntree tree.
This research uses K-fold cross-validation method,
with 10 as the number of K variable. This method will be With an ensemble of classifiers ℎ1 (𝑋) , ℎ2 (𝑋) , … ,
applied to model development using only training data. ℎ𝐾 (𝑋) and training dataset is drawn in the random forest
Fig. 3 explains how K-fold cross-validation work. from the distribution of vector X, Y, the margin function
can be expressed as formula (3).

𝑚𝑔(𝑋, 𝑌) =
𝑎𝑣𝐾 𝐼(ℎ𝐾 (𝑋) = 𝑌) − max 𝑎𝑣𝐾 𝐼(ℎ𝐾 (𝑋) = 𝑗) (3)
𝑗≠𝑌

Estimation of the error rate on the training data can be

done by predicting data not in the bootstrap sample, using
tree grown with the bootstrap sample on each bootstrap
iteration. The next step is to aggregate the OBB (out-of-
bag) and calculate the error rate. We can call it the OOB
estimate of error rate.
5. Extreme Gradient Boosting Algorithm
Extreme Gradient Boosting or known as XGBoost is a
decision tree algorithm introduced by Tianqi Chen and
Fig.3. K-fold cross-validation illustration. Tong He. This method was introduced to solve problems
in the Higgs boson machine learning competition. This
3. C5.0 Algorithm method is a development of the gradient boosting
approach by studying ensembles from boosted trees. This
C5.0 is a decision tree algorithm that has been method is able to offer speed in the training process and
developed based on C4.5 algorithm [14]. It has all C4.5 good accuracy values [18].
algorithm functionality with improvements on the XGBoost has been used in many competitions in the
technology. Techniques introduced in C5.0 as an machine learning field. One of them was in the
improvement from it ancestor is [15] boosting, variable competition held by Kaggle in 2015, where 17 solutions
misclassification, new attributes, values can be marked as used XGBoost from a total of 29 who entered and
missing or not applicable on a particular case, support succeeded as the competition's winning candidate [19]. In
sampling and cross-validation. To get entropy value, we the 2015 KKDCup competition, the top 10 winning team
can use (1) formula. all used XGBoost method [19]. The XGBoost
methodology works by combining all predictions of a set
𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝑘) = ∑𝑖 − 𝑝𝑖 𝑙𝑜𝑔 𝑝𝑖 (1) of weak learners to develop a strong learning with
additive training strategy. Equation (4) is the general
where 𝑝𝑖 is the probability of class i within node k used to formula to make predictions:
split node. To find splitting criterion from a root that has
been defined before can be expressed with formula (2). 𝑓𝑖 (𝑡) = ∑𝑡𝑘=1 𝑓𝑘 (𝑥𝑖 ) = 𝑓𝑖 (𝑡−1) + 𝑓𝑡 (𝑥𝑖 ) (4)
𝐺𝑎𝑖𝑛(𝑘)
𝐺𝑎𝑖𝑛 𝑅𝑎𝑡𝑖𝑜(𝑘) =
𝑆𝑝𝑙𝑖𝑡𝐼𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛(𝑘)
(2) where 𝑓𝑡 (𝑥𝑖 ) is a learner at step t, 𝑓𝑖 (𝑡) and 𝑓𝑖 (𝑡−1) is the
prediction in step t and t-1 and 𝑥𝑖 as the input variable.
4. Random Forest Algorithm To improve the performance of the model, this method
Random forest is a classification of tree that can be optimize computation resources. In the XGBoost method,
used to make a prediction. [16] proposed this combining predictive and regularization is used to
methodology which changes how the classification tree simplify the objective function and to maintain the
and regression tree are constructed. In this method, each optimal computational speed expressed in the formula (5).
node is split based on the best among subsets of
predictors randomly chosen at that node. Ref. [16] stated 𝑂𝑏𝑗𝑡 = ∑𝑛𝑘=1 𝑙(ỹ𝑖 , 𝑦𝑖 ) + ∑𝑡𝑘=1 Ω(𝑓𝑖 ) (5)
this technique will make the tree constructed robustly
against overfitting. The algorithm for this random forest
is explained below [17]. where l is the loss function, n is the number of
observations and Ω is the regularization terms. The
• Draw ntree bootstrap sample from original data regularization terms can be expressed in (6) formula.
• For each of the bootstrap samples, make an
1
unpruned classification tree with modification at Ω(𝑓) = 𝛾𝑇 + 𝜆||𝜔||2 (6)
2

Volume 11 (2019), Issue 7 23

 Football Match Prediction with Tree Based Model Classification

where 𝜔 is the vector score in the leaves, 𝜆 is the generated the following combinations of HTHG, HTAG,
regularization parameters, and 𝛾 is the minimum loss HS, AS, HST, AST, HF, AF, HC, and HY. It is the best
needed to expand partition of the leaf node. feature combinations with prediction accuracy reaching
69.21%. Selected feature combinations will be used in
6. Evaluation Methodology
further model predictions using a tree-based model
The evaluation will use a confusion matrix to calculate algorithm.
each model’s prediction accuracy, kappa, F1, sensitivity,
and specificity. Each model will have one confusion Table 3. Features selection process
matrix that consists of each class from the chosen label. No. of Features Accuracy Kappa OOB Error Rate
The class will have 3 nominal value, H for home win, D 14 67,5% 0.4761 30,99%
for draw and A for away win. 13 68,16% 0.4882 31,84%
12 68,03% 0.4865 31,71%
11 66,97% 0.4708 31,48%
IV. RESULT AND ANALYSIS 10 69,21% 0.5080 31,71%
9 67,37% 0.4808 32,04%
A. Feature Selection 8 68,03% 0.4909 31,48%
Feature selection method used in this research is the 7 67,63% 0.4845 32,27%
wrapper method with a backward approach. From a total 6 66,45% 0.4675 33,59%
of 14 features available, features combination will be 5 64,61% 0.4375 36,02%
tested against the training model. Each combination will 4 63,55% 0.4266 36,38%
be measured by the accuracy percentage of the training 3 63,68% 0.4208 39,38%
model. Data partition in this feature selection will be 2 58,82% 0.3824 38,98%
80:20 which 80% of the available data used as training 1 55,26% 0.2694 44,54%
data. The model will use a random forest algorithm to B. Analysis of Tree-Based Classification Model
check the prediction accuracy of the model with the
parameter ntree is 500 and mtry is 2. ntree parameter used in The experiment process started by dividing the data
the Random Forest algorithm is the number of trees to into two partitions, training data, and testing data, with
grow while the model developed and parameter mtry is the composition of 80:20. Training data consists of 3,040
number of variables available for splitting at each tree data with 10 features and testing data consist of 760 data
node while the model developed. with 10 features.
From all number of feature combination available, 10
features combination has the highest prediction accuracy Table 4. Evaluation of training data using C5.0 algorithm
which is 69.21% and an out-of-bag estimate of error rate Rules
is 31.71%. The confusion matrix of 10 features Trial
Number Errors
combination is shown in table 2. 0 48 870(28,6%)
1 29 980(32,2%)
Table 2. Confusion matrix of 10 features combination 2 26 1054(34,7%)
Output Class 3 34 1072(35,3%)
Class
A D H 4 24 1075(35,4%)
Target A 153 50 22 5 26 1123(36,9%)
D 24 74 32 6 30 1045(34,4%)
H 36 70 299 7 33 1027(33,8%)
8 40 906(29,8%)
Table 3 shows the results of each feature combination 9 59 896(29,5%)
using the backward wrapper method. The process will Boost 741(24,4%)
have 14 iterations based on the number of features used
1. C5.0 Model
in this research. Feature deletion is done based on the
lowest feature importance on the generated training The first algorithm used in this experiment is C5.0. The
model. The table displaying accuracy, kappa and out-of- model trained with 10 trials and a rule-based tree.
bag error rate as a measurement on each model. Training data evaluation can be seen in table 4. The
Although not having the lowest out-of-bag error rate, model was evaluated using a confusion matrix to see the
the backward wrapper method concluded that 10 feature classification performance, this can be seen in table 5
combinations give the highest accuracy and kappa with accuracy reaching 75,534%.
number. To ensure the best feature combination, the The trained model is tested against testing data using a
training model re-generated the 10 feature combinations confusion matrix. The prediction accuracy result is 63,
by reducing 4 feature combinations for each iteration. 29% with Kappa 0.4095. The confusion matrix result can
The experiment result of 10 feature combinations be seen in table 6.

24 Volume 11 (2019), Issue 7

 Football Match Prediction with Tree Based Model Classification

Table 9. Experiment result on best ntree parameter

Table 5. Confusion matrix of C5.0 training model
Accuracy
ntree
Output Class Min Mean Max
Class
A D H 250 0.6053 0.6273 0.6447
Target A 691 53 130 300 0.6086 0.6280 0.6480
D 155 370 271 350 0.6086 0.6293 0.6513
H 81 51 1238 400 0.6066 0.6290 0.6546
450 0.6151 0.6309 0.6546
Table 6. Confusion matrix of C5.0 validation model 500 0.6131 0.6303 0.6579
550 0.6066 0.6309 0.6546
Output Class
Class 600 0.6086 0.6303 0.6579
A D H
800 0.6066 0.6283 0.6579
Target A 142 63 28
1000 0.6033 0.6290 0.6546
D 26 48 34
2000 0.6066 0.6286 0.6513
H 45 83 291

2. Random Forest Model The experiment to improve Random Forest model

prediction continues by testing on mtry parameter. The
The next algorithm used in this experiment is Random
tuning experiment using a grid search method with
Forest. The model trained with 10 fold cross-validation
number 1 to 10 on a mtry parameter in the training model.
and mtry parameter with value 2, 6 and 10. With
The experiment result is mtry = 3 gives the highest
prediction accuracy, 67,89% and Kappa 0.493, the best
prediction accuracy 65,36%. The experiment result on
Random Forest model used the value of 2 as mtry variable.
mtry parameter can be seen in table 10 and fig. 4.
Trained model tested using testing data to measure
prediction accuracy. The accuracy of Random Forest
Table 10. Experiment result on best mtry parameter
model is 62,76% with Kappa 0.392. The confusion matrix
and statistics of Random Forest model validation can be mtry Accuracy Kappa
seen in table 7 and table 8. 1 63,91% 0.4100
2 65,06% 0.4469
Table 7. Confusion matrix of Random Forest training model 3 65,36% 0.4543
4 65,33% 0.4556
Output Class
Class 5 65,03% 0.4507
A D H
6 64,77% 0.4477
Target A 144 63 25
7 64,54% 0.4440
D 20 30 25
8 64,80% 0.4483
H 49 101 303
9 64,80% 0.4485
10 64,83% 0.4493
Table 8. Random Forest validation statistics

Output Class
Statistics
A D H
Sensitivity 0.6761 0.1546 0.8584
Specificity 0.8391 0.9205 0.6314
Kappa 0.3920
Accuracy 0.6276

With accuracy prediction lower than the C5.0

algorithm, the Random Forest algorithm needs to be
improved. Accuracy improvement will be done in this
experiment by testing on several parameters such as ntree
and mtry. Using the brute force method, the model was
tested using several values which are 250, 300, 350, 400, Fig.4. mtry parameter plot on Random Forest.
450, 500, 550, 600, 800, 1000 and 2000. The result of the
experiment displayed ntree = 500 with the highest Based on the tuned parameters experiment, the
accuracy. The experiment results to find the best ntree can Random Forest model was re-trained and tested with
be seen in table 9. testing data again. The prediction accuracy with tuned
parameters using Random Forest model is 68,55% with
Kappa value 0.5005.

Volume 11 (2019), Issue 7 25

 Football Match Prediction with Tree Based Model Classification

tuning.
3. Extreme Gradient Boosting Model
The last algorithm used in this experiment is Extreme Table 12. Extreme Gradient Boosting Tuning Parameters
Gradient Boosting. With 10 fold cross-validation, the first Learner Initial Parameters
step in this experiment is to get the best iteration with the objective multi:softprob
lowest m-error. Based on the experiment run, 28 iteration eval_metric merror
is the best with the lowest testing m-error with value nrounds 250
0,3526. The m-error plot on each iteration of training and Parameters for Tuning
testing data can be seen on fig. 5 and fig. 6. Parameters Lower Upper
nrounds 200 600
max_depth 3 20
eta 0.001 0.5
lambda 0.55 0.6
subsample 0.1 0.8
min_child_weight 1 5
colsample_bytree 0.2 0.8

Tuning process will use 10 fold cross-validation and

iterated as much as 100 times. The result of this tuning
process displayed in table 13.

Fig.5. Training m-error plot on Extreme Gradient Boosting. Table 13. Extreme Gradient Boosting Tuning Result

Parameters
nrounds 204
max_depth 17
Eta 0.0178
lambda 0.5750
subsample 0.6610
min_child_weight 1,05
colsample_bytree 0.7580
test_mean_accuracy 67,0395%

Tuned model then tested and generated prediction

accuracy as high as 67,89% with Kappa value 0.4867.
Fig.6. Testing m-error plot on Extreme Gradient Boosting. The confusion matrix of tuned Extreme Gradient
Boosting model can be seen in table 14.
The trained model then tested using testing data to
generate the prediction accuracy of Extreme Gradient Table 14. Confusion matrix of Extreme Gradient Boosting tuned model
Boosting algorithm. Using a confusion matrix, the Output Class
accuracy of Extreme Gradient Boosting model is 64,74% Class
A D H
with Kappa value 0.4375. The confusion matrix result Target A 145 50 23
can be seen in table 11. D 28 74 33
H 40 70 297
Table 11. Confusion matrix of Extreme Gradient Boosting
validation model
Table 15. Model performance evaluation
Output Class
Class Evaluation Random
A D H C5.0 XGBoost
Metrics Forest
Target A 149 58 27 Sensitivity 0.5837 0.6545 0.6345
D 23 60 43 Specificity 0.8078 0.8381 0.8295
H 41 76 283 Precision 0.6053 0.6652 0.6477
Recall 0.5837 0.6545 0.6345
Extreme Gradient Boosting algorithm has the lowest F1 0.5506 0.6546 0.6348
accuracy between the three algorithms used in this Kappa 0.4258 0.5123 0.4867
research. Therefore the algorithm needs to be tuned. The Accuracy 0.6487 0.6855 0.6789
tuning process will use mlr package from R library by
creating a learner on each model training. Initial To summarize the performance of each final model
parameter defined in the tuning process is objective, algorithm, table 15 and fig. 7 display the comparison with
evaluation metric, and nrounds. Table 12 displays the detail performance evaluation.
description of Extreme Gradient Boosting algorithm

26 Volume 11 (2019), Issue 7

 Football Match Prediction with Tree Based Model Classification

[6] N. Razali, A. Mustapha, F. A. Yatim, and R. Ab Aziz,

“Predicting Football Matches Results using Bayesian
Networks for English Premier League (EPL),” in IOP
Conference Series: Materials Science and Engineering,
2017, vol. 226, no. 1.
[7] V. Chaurasia, “Early Prediction of Heart Diseases Using
Data Mining,” Caribbean. J. Sci. Technol., vol. 1, pp.
208–217, 2013.
[8] A. Zakerian, A. Maleki, Y. Mohammadnian, and T.
Amraee, “Bad data detection in state estimation using
Decision Tree technique,” in 2017 25th Iranian
Conference on Electrical Engineering, ICEE 2017, 2017.
[9] P. Carmona, F. Climent, and A. Momparler, “Predicting
Fig.7. Evaluation Metrics Chart. failure in the U.S. banking sector: An extreme gradient
boosting approach,” International Review of Economics
and Finance, 2018.
[10] J. Tang, S. Alelyani, and H. Liu, “Feature Selection for
V. CONCLUSION Classification: A Review,” Data Classif. Algorithms Appl.,
2014.
In this research, we developed a prediction model [11] R. Kohavi and G. H. John, “Wrappers for feature subset
using the Decision Tree based machine learning to selection,” Artif. Intell., 1997.
predict the output of the English Premier League using [12] F. Gorunescu, “Data mining: Concepts, models and
historical match statistic data. Feature HTHG, HTAG, techniques,” Intell. Syst. Ref. Libr., 2011.
HS, AS, HST, AST, HF, AF, HC, and HY is the best [13] P. Refaeilzadeh, L. Tang, and H. Liu., “‘Cross-
feature combinations to optimize the model prediction Validation.,’” in Encyclopedia of database systems, 2009.
accuracy. The accuracy of C5.0, Random Forest, and [14] S. PANG and J. GONG, “C5.0 Classification Algorithm
and Application on Individual Credit Evaluation of Banks,”
Extreme Gradient Boosting consecutively is 64,87%,
Syst. Eng. - Theory Pract., 2009.
68,55%, and 67,89%. Although the accuracy of the model [15] T. Bujlow, T. Riaz, and J. M. Pedersen, “A method for
using decision tree based is not good enough compared to classification of network traffic based on C5.0 machine
the model prediction developed by Bayesian network learning algorithm,” in 2012 International Conference on
algorithm or artificial neural network algorithm, it still Computing, Networking, and Communications, ICNC’12,
has a better performance compared to SVM with 2012.
accuracy 53,3% or logistic regression which only [16] L. Breiman, “Random Forrest,” Mach. Learn., 2001.
generated 2 output class with accuracy 69,5%. We [17] A. Liaw and M. Wiener, “Classification and Regression
concluded that the tuning method can be used on the by randomForest,” R news, 2002.
[18] T. Chen, G. Cowan, C. Germain, I. Guyon, B. Azs Kégl,
model to improve the accuracy of prediction.
and D. Rousseau, “Higgs Boson Discovery with Boosted
In this research, the random forest algorithm and C5.0 Trees,” in NIPS 2014 Workshop on High-energy Physics
performed better than the extreme gradient boosting and Machine Learning, 2014.
algorithm before the parameters were tuned. It happened [19] T. Chen and C. Guestrin, “XGBoost: Reliable Large-scale
due to the algorithm of extreme gradient boosting being Tree Boosting System,” in LearningSys, 2016.
overfit and the trained model only had a small dataset
with few features to learn. In the future, we could further
develop this research by combining datasets to improve
the accuracy with more features that have significant Authors’ Profiles
relevance, also by improving the feature selection method
to maximize the feature selection process. Yoel F. Alfredo is a student of the graduate
program for Master in Computer Science
REFERENCES BINUS University. He received his
bachelor degree from BINUS University in
[1] W. J. Murray and B. Murray, The worlds game: a history 2016 with System Information major. He
of soccer, vol. 14. Urbana: University of Illinois Press, has worked as a Product Manager in the
1998. startup industry for 3 years. His research
[2] D. Prasetio and Harlili, “Predicting football match results interests are machine learning, big data, and
with logistic regression,” in 4th IGNITE Conference and artificial intelligence.
2016 International Conference on Advanced Informatics:
Concepts, Theory, and Application, ICAICTA 2016, 2016.
[3] M. Faculty, A. Yezus, and A. Igoshkin, “Predicting Sani M. Isa is a lecturer and researcher in
outcome of soccer matches using machine learning,” the Computer Science Department, BINUS
Saint-petersbg. Univ., 2014. Graduate Program - Master of Computer
[4] C. P. Igiri and E. O. Nwachukwu, “An Improved Science. He has numerous experience in
Prediction System for Football a Match Result,” IOSR J. teaching and research in remote sensing and
Eng., vol. 04, no. 12, pp. 12–20, 2014. biomedical engineering areas. He got his
[5] C. P. Igiri, “Support Vector Machine–Based Prediction doctoral degree in Computer Science from
System for a Football Match Result,” IOSR J. Comput. the University of Indonesia. He is also
Eng. Ver. III, vol. 17, no. 3, pp. 2278–661, 2015. received his master degree in Computer Science from the

Volume 11 (2019), Issue 7 27

 Football Match Prediction with Tree Based Model Classification

University of Indonesia as well as a bachelor degree from

Padjadjaran University, Bandung, Indonesia.

How to cite this paper: Yoel F. Alfredo, Sani M. Isa, "Football

Match Prediction with Tree Based Model Classification",
International Journal of Intelligent Systems and
Applications(IJISA), Vol.11, No.7, pp.20-28, 2019. DOI:
10.5815/ijisa.2019.07.03

28 Volume 11 (2019), Issue 7