Football-Dataset-Analysis-Kaggle

Objective

• Build a model to predict the number of goals in a match according to match events.

Kaggle Data-Set

• The dataset provides a granular view of 9,074 games, totaling 941,009 events from the biggest 5 European football (soccer) leagues: England, Spain, Germany, Italy, France from 2011/2012 season to 2016/2017 season.

The dataset is organized in 3 files:

• events.csv contains event data about each game.

• ginf.csv contains metadata and market odds about each game

• dictionary.txt contains a dictionary with the textual description of each categorical variable coded with integers.

Data Understanding

• Event type: Corner, Foul, Substitution, Red card, Yellow card, Hand ball, Offside, etc.

• Location: Centre of the box, Outside the box , Left side of the six yard box, Long range, etc.

• Shoot place: Too high, Bit Too high, Bottom left corner, Top centre of the goal , etc.

• shoot outcome: On target, OFF target, Blocked or Hit the bar.

Data pre-processing

• Separate each event into home or away match according to data-set.

• Missing data: we put any missing data equal -1, because data have integers and zeroes values.

• One hot encoding Values.

• Labels: labels will be number of goals per match (home and away matches).

• Vectored: 5 features have been selected (6 features have been selected but I use side as anther ID to separate matches have the same ID match), In every match there 180 events (maximum number of events per match). so input is a 2-d array size (5 X 180), by vectorized it into 1-d array (1 X 900).

Neural Network

• Recurrent Neural Network (RNN) as a classifier.

• Inputs 2d-array has size (number of matches X Vectored features events).

• Labels: numbers of goal in a match.

• Output: prediction number of goals in this match (Float number we round it).

Network structure

• Training data: Events of 2000 matches.

• Test data: Events of Events100 matches.

• Accuracy: round(prediction)== label.

• Loss function: Mean Square Error (MSE).

• Optimizer: Adam.

• Device: CPU.

Hyper-parameters

• Batch size: 100

• Learning rate: 1e-3

• Weight decay: 1e-3

Results

	Home with one-hot	Away with one-hot	Home without one-hot	Away without one-hot
Accurcy	1.726 %	2.27 %	56.017 %	60.937 %
Training time (minutes)	84.06	83.046	83.076	80.479

Discussion

• One-hot encoding increases the training time and reduces accuracy significantly because when we use one-hot encoding we increase the variance between values which change the correlation between data.

• Recommend to use LSTM.