Fifa 2017 Player position Classifiers
At the Codenation website, there are several Data Science challenges to apply the Machine Learning knowledge in a practical way. One of these challenges is to discover the position of the players of FIFA© 2017 based on the characteristics of these players.
Development
Primeiramente, devemos importar as bibliotecas utilizadas:
Importing things
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import KFold
from sklearn.model_selection import cross_val_score
from sklearn.model_selection import GridSearchCV
from sklearn import preprocessing
from sklearn.neural_network import MLPClassifier
from sklearn.naive_bayes import BernoulliNB
from sklearn.ensemble import RandomForestClassifier
from xgboost import XGBClassifier
import collections as cll
import warnings
warnings.filterwarnings('ignore')
%matplotlib inline
Reading the data
train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')
ans = pd.DataFrame()
# Check if the test data is in the training data
print(set(test.columns).issubset(set(train.columns)))
True
Quick data exploration
# What attributes do we have? What type are they?
train.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 4829 entries, 0 to 4828
Data columns (total 55 columns):
ID 4829 non-null object
height_cm 4829 non-null float64
weight_kg 4829 non-null float64
eur_wage 4829 non-null float64
skill_moves 4829 non-null int64
weak_foot 4829 non-null int64
crossing 4829 non-null int64
finishing 4829 non-null int64
heading_accuracy 4829 non-null int64
short_passing 4829 non-null int64
volleys 4829 non-null int64
dribbling 4829 non-null int64
curve 4829 non-null int64
free_kick_accuracy 4829 non-null int64
long_passing 4829 non-null int64
ball_control 4829 non-null int64
acceleration 4829 non-null int64
sprint_speed 4829 non-null int64
agility 4829 non-null int64
reactions 4829 non-null int64
balance 4829 non-null int64
shot_power 4829 non-null int64
jumping 4829 non-null int64
stamina 4829 non-null int64
strength 4829 non-null int64
long_shots 4829 non-null int64
aggression 4829 non-null int64
interceptions 4829 non-null int64
positioning 4829 non-null int64
vision 4829 non-null int64
penalties 4829 non-null int64
composure 4829 non-null int64
marking 4829 non-null int64
standing_tackle 4829 non-null int64
body_type 4829 non-null object
nationality 4829 non-null object
preferred_foot 4829 non-null object
outside_foot_shot_trait 4829 non-null bool
playmaker_trait 4829 non-null bool
power_free_kick_trait 4829 non-null bool
power_header_trait 4829 non-null bool
puncher_trait 4829 non-null bool
rushes_out_of_goal_trait 4829 non-null bool
saves_with_feet_trait 4829 non-null bool
selfish_trait 4829 non-null bool
skilled_dribbling_trait 4829 non-null bool
takes_finesse_free_kicks_trait 4829 non-null bool
target_forward_trait 4829 non-null bool
team_player_trait 4829 non-null bool
technical_dribbler_trait 4829 non-null bool
tries_to_beat_defensive_line_trait 4829 non-null bool
poacher_speciality 4829 non-null bool
speedster_speciality 4829 non-null bool
aerial_threat_speciality 4829 non-null bool
preferred_pos 4829 non-null object
dtypes: bool(17), float64(3), int64(30), object(5)
memory usage: 1.5+ MB
# Statisticaly describing the data
train.describe()
| height_cm | weight_kg | eur_wage | skill_moves | weak_foot | crossing | finishing | heading_accuracy | short_passing | volleys | ... | strength | long_shots | aggression | interceptions | positioning | vision | penalties | composure | marking | standing_tackle | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| count | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | ... | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 | 4829.000000 |
| mean | 183.109961 | 76.990267 | 10016.359495 | 2.083868 | 2.819217 | 41.739283 | 39.410644 | 50.172292 | 52.684407 | 37.804307 | ... | 66.328847 | 40.354111 | 52.861876 | 42.873887 | 42.798716 | 47.659350 | 44.654380 | 54.402568 | 40.929178 | 43.800373 |
| std | 6.690962 | 7.292228 | 21720.982928 | 0.726900 | 0.628180 | 19.045680 | 20.721807 | 20.877075 | 16.181135 | 18.377069 | ... | 13.099721 | 19.790496 | 19.088383 | 21.958820 | 21.213381 | 14.071487 | 17.119506 | 14.146100 | 23.301924 | 23.571392 |
| min | 156.000000 | 49.000000 | 0.000000 | 1.000000 | 1.000000 | 6.000000 | 3.000000 | 4.000000 | 10.000000 | 4.000000 | ... | 20.000000 | 3.000000 | 11.000000 | 4.000000 | 2.000000 | 10.000000 | 5.000000 | 5.000000 | 4.000000 | 7.000000 |
| 25% | 178.000000 | 72.000000 | 1000.000000 | 2.000000 | 2.000000 | 27.000000 | 21.000000 | 40.000000 | 42.000000 | 23.000000 | ... | 58.000000 | 22.000000 | 36.000000 | 22.000000 | 25.000000 | 37.000000 | 33.000000 | 46.000000 | 18.000000 | 19.000000 |
| 50% | 184.000000 | 77.000000 | 3000.000000 | 2.000000 | 3.000000 | 43.000000 | 38.000000 | 55.000000 | 57.000000 | 37.000000 | ... | 67.000000 | 42.000000 | 56.000000 | 45.000000 | 47.000000 | 48.000000 | 45.000000 | 56.000000 | 42.000000 | 46.000000 |
| 75% | 188.000000 | 82.000000 | 9000.000000 | 3.000000 | 3.000000 | 58.000000 | 58.000000 | 66.000000 | 65.000000 | 52.000000 | ... | 76.000000 | 57.000000 | 68.000000 | 63.000000 | 60.000000 | 58.000000 | 58.000000 | 65.000000 | 63.000000 | 66.000000 |
| max | 203.000000 | 107.000000 | 510000.000000 | 5.000000 | 5.000000 | 90.000000 | 94.000000 | 94.000000 | 88.000000 | 88.000000 | ... | 96.000000 | 90.000000 | 94.000000 | 90.000000 | 92.000000 | 86.000000 | 90.000000 | 91.000000 | 90.000000 | 92.000000 |
8 rows × 33 columns
# Checking the data size
print(train.shape)
print(test.shape)
(4829, 55)
(4829, 54)
# Finding out how many classes exists
train['preferred_pos'].nunique()
15
Spliting train and test sets
ans['ID'] = test['ID']
X_train = train.drop(['preferred_pos', 'ID'], axis = 1)
y_train = train['preferred_pos']
X_test = test.drop(['ID'], axis = 1)
# Check for nulls
count_nan_train = len(X_train) - X_train.count()
count_nan_test = len(X_test) - X_test.count()
print(count_nan_train.sum())
print(count_nan_test.sum())
0
0
Checking dataset balance
sns.countplot(y_train)
cll.Counter(y_train)
Counter({'prefers_gk': 965,
'prefers_cm': 398,
'prefers_st': 929,
'prefers_cb': 1117,
'prefers_rb': 318,
'prefers_lb': 342,
'prefers_cdm': 213,
'prefers_cam': 157,
'prefers_rw': 53,
'prefers_lm': 124,
'prefers_rm': 158,
'prefers_lw': 37,
'prefers_lwb': 4,
'prefers_cf': 12,
'prefers_rwb': 2})
# Where the posision quantity is too low, change the label to a more commum position (to increase accuracy)
for i in range(len(y_train)):
if y_train[i] == 'prefers_rwb' or y_train[i] == 'prefers_lwb' or y_train[i] =='prefers_lw' or y_train[i] =='prefers_cf' or y_train[i] =='prefers_rw':
y_train[i] = 'prefers_cb'
Bodyshape Analysis
print(X_train['body_type'].unique())
print(X_test['body_type'].unique())
['Stocky' 'Normal' 'Lean' 'Courtois']
['Normal' 'Stocky' 'Lean' 'Akinfenwa' 'Neymar']
- Courtois -> Lean
- Akinfenwa -> Stocky
- Neymar -> Lean
# Changing strange body types
for i in range(len(X_train)):
if X_train['body_type'][i] == 'Courtois':
X_train['body_type'][i] = 'Lean'
for i in range(len(X_test)):
if X_test['body_type'][i] == 'Akinfenwa':
X_test['body_type'][i] = 'Stocky'
if X_test['body_type'][i] == 'Neymar':
X_test['body_type'][i] = 'Lean'
Dummy variables
body_shape = pd.get_dummies(X_train['body_type'])
X_train = pd.concat([X_train, body_shape], axis = 1).drop(['body_type'], axis = 1)
X_train.drop(['nationality'], axis = 1, inplace = True)
preferred_foot = pd.get_dummies(X_train['preferred_foot'])
X_train = pd.concat([X_train, preferred_foot], axis = 1).drop(['preferred_foot'], axis = 1)
Correlation
# Correlation between the attributes
plt.figure(figsize=(12,12))
sns.heatmap(X_train.corr(),cmap='Blues')
Scaling attributes
features = X_train.columns
Scaler = preprocessing.StandardScaler()
X_train = Scaler.fit_transform(X_train)
Models
# Using k_folds to evaluate
k_fold = KFold(n_splits=10, shuffle=True, random_state=42)
scoring = 'accuracy'
Random Forest
%%time
rfc = RandomForestClassifier(n_estimators=100)
score = cross_val_score(rfc, X_train, y_train, cv=k_fold, n_jobs=1, scoring=scoring)
print(round(np.mean(score)*100, 2))
86.35
CPU times: user 6.46 s, sys: 7.39 ms, total: 6.47 s
Wall time: 6.35 s
Naive Bayes
bnb = BernoulliNB()
score = cross_val_score(bnb, X_train, y_train, cv=k_fold, n_jobs=1, scoring=scoring)
print(round(np.mean(score)*100, 2))
78.19
Neural Network
%%time
mlp = MLPClassifier(solver='lbfgs', alpha=1e-5, hidden_layer_sizes=(25, 15, 5), random_state=1,early_stopping=True)
score = cross_val_score(mlp, X_train, y_train, cv=k_fold, n_jobs=1, scoring=scoring)
print(round(np.mean(score)*100, 2))
84.1
CPU times: user 36.7 s, sys: 27.3 ms, total: 36.8 s
Wall time: 9.2 s
XGBoost
n_estimators = [50, 100, 150, 200]
max_depth = [2, 4, 6, 8]
param_grid = dict(max_depth=max_depth, n_estimators=n_estimators)
xgb = XGBClassifier()
grid_search = GridSearchCV(xgb, param_grid, scoring="neg_log_loss", n_jobs=-1, cv=k_fold, verbose=0)
score = cross_val_score(grid_search, X_train, y_train, cv=k_fold, n_jobs=1, scoring=scoring)
print(round(np.mean(score)*100, 2))
87.31
best_xgb = grid_search.fit(X_train, y_train)
best_xgb.best_params_
{'max_depth': 4, 'n_estimators': 150}
Feature importance
# Fit on the Random Forest to calculate the feature importance for each attribute
rfc.fit(X_train, y_train)
importances = rfc.feature_importances_
indices = np.argsort(importances)
# Print the top 15 feature ranking
quant = 15
plt.title('Feature Importances')
plt.barh(range(quant), importances[indices][-quant:], color='b', align='center')
plt.yticks(range(quant), [features[i] for i in indices[-quant:]])
plt.xlabel('Relative Importance')
plt.show()
Making the submission
# Pre-processing the test data
body_shape = pd.get_dummies(X_test['body_type'])
X_test = pd.concat([X_test, body_shape], axis = 1).drop(['body_type'], axis = 1)
X_test.drop(['nationality'], axis = 1, inplace = True)
preferred_foot = pd.get_dummies(X_test['preferred_foot'])
X_test = pd.concat([X_test, preferred_foot], axis = 1).drop(['preferred_foot'], axis = 1)
X_test = Scaler.fit_transform(X_test)
# Making the prediction
prediction = best_xgb.predict(X_test)
ans['preferred_pos'] = prediction
ans.to_csv('answer.csv', index=False)
This submission was submitted to Codenation for review and resulted in 87.43% accuracy, which is not bad at all for this challenge.