Transcript Alberto Trindade Tavares - University of Wisconsin–Madison

ECE/CS/ME 539 - Introduction to Artificial Neural Network and Fuzzy Systems
Alberto Trindade Tavares
Brazilian Soccer League
 Since 2003, the Brazilian Soccer League has the following format:
 20 participating clubs
 Each club faces every other club twice in the season, once at their home
stadium, and once at that of their opponents
 380 matches divided into two parts:
First half: May-August
 Second half: September-December

 A match has three possible results:
 Win of the home team
 Draw
 Loss of the home team
Goal of this Work
 Predict the outcome (win of home team, draw, or loss of home team) of
every game of the second half for the current season (2013)
 Using as training data the game results of the first half of 2013 season
 Develop two classifiers, using MATLAB, for performing these
predictions :

Maximum Likelihood Classifier
 Multi-Layer Perceptron
 Compare their accuracy between themselves and to other works
Feature Vector
 The feature vector for representing a match instance has six features,
the first three for the first team (home), and last three for the second
team (visiting):
# wins as
home team
# draws as
home team
First Team
# losses as
home team
# wins as
# draws as
# losses as
visiting team visiting team visiting team
Second Team
 Results from 2003 season to the last match of current season
Data Extraction
 Extraction of results of every match since 2003
 Two different sources:
 2003-2004 seasons: http://www.bolanaarea.com/gal_brasileirao.htm
 2005-2013 seasons: http://www.campeoesdofutebol.com.br
 Python program for parsing the HTML pages, and storing the results
into text files, which can be read via MATLAB function load
Maximum Likelihood Classifier
 Gaussian Distribution
P(x)
Win
Loss
Draw
x
Maximum Likelihood Classifier (Results)
 Classification rate per round:
Average Classification Rate = 53.1579%
Maximum Likelihood Classifier (Results)
 Total Confusion Matrix:
Predicted Wins
Predicted Draws
Predicted Losses
Actual Wins
71
8
18
Actual Draws
25
9
15
Actual Losses
19
4
21
Multi-Layer Perceptron
 # Hidden Layers = 3
 # Neurons in First Hidden Layer = 3
 # Neurons in First Hidden Layer = 20
 # Neurons in First Hidden Layer = 3
 Learning rate (α) = 0.1
 Momentum = 0
 Hidden layers use hyperbolic tangent activation function, and output
layer uses sigmoid activation function
Multi-Layer Perceptron(Results)
 Classification rate per round (10 runs):
Average Classification Rate = 55.7895%
Multi-Layer Perceptron(Results)
 Total Confusion Matrix:
Predicted Wins
Predicted Draws
Predicted Losses
Actual Wins
78
7
12
Actual Draws
27
11
11
Actual Losses
23
4
17
Comparison with other work
 A. Joseph, N.E. Fenton, M. Neil. Predicting football results using
Bayesian nets and other machine learning techniques (2006)
 Published in the Journal Knowledge-Based Systems
 Their results:
 Naïve BN: 47.86%
 kNN: 50.58%
 Expert BN: 59.21%
My Results:
 Maximum Likelihood: 53.1579%
 Multi-Layer Perceptron: 55.7895%
Questions?