目录
1. 背景
在这四年一次的足球世界杯即将来临之际,公司将要举办一个模型预测与人工预测的比赛活动,虽然本人不是足球迷也不是伪球迷,但是绷着对机器学习的热爱,决定凭借自己匮乏的足球知识建个模型参加这个比赛。无论成败与否,享受其中即可。
2. 输入特征选择
特征的选择通俗的讲就是选择什么特征最能影响足球比赛。能我让这个不懂足球的人想到的是下面这些特征:
1. 明星队员:比如梅西,C罗,姆巴佩,内马尔等等。但是这些球员的实力怎么衡量呢?还有就是不断有新的球星出现,是不是这个特征有点多?再有这个数据是不是好得到呢?
2. 主客场:听说篮球这个特征很重要,但是世界杯就只有主办方一家是主场,其它是客场。
3. 裁判:有些比赛裁判可能偏向一个国家队,这就操成该罚球的不罚,但是裁判这个特征随机性太强,每届都在变,而且你也不知道他偏向哪个队。
4. 教练:感觉也是特征,而且比较稳定。但是怎么衡量呢
5.球队状态:影响是有,但感觉无从下手的样子。
6. 观众:历史数据应该还有办法,但是新比赛拿不到。
最后就是球队了,如果球队的整体实力比较高,哪怕是没有明星队员,结果也不差,而且球队的整体实力其实暗含教练,球员和明星球员的这几个特征的。综合这些特征个方面的因素和提取的难易程度,最后选择了球队名字作为输入特征。
3. 数据集-特征获取
由于比赛来的太突然,就几天就要开始比赛当时,所以带着侥幸的心理在网上搜了一下是不是有相关的数据集,果然有很多。最终选择Kaggle里下载已经整理好的数据而且还是免费的。
下载数据集
下载后数据如下:
百度网盘也可以下载
链接: https://pan.baidu.com/s/1ky-73f9YTe2o1YI1TyMt9A 提取码: 1pt6
4. 数据预处理
虽然Kaggle这里提供的数据有很多特征,但是我们只需要三个特征,模型的输入是两个队的名字,模型的输出是结果(2:Home Team队胜,1:平,0:Away Team 队胜)。
数据预处理大致的步骤是
- 从Excel里加载数据;
- 根据每队的得分生成新队列Winning_Team;
- 去掉和本次世界杯无关的国家队;
- 去掉无关的特征;
- winning_team数值化;
- 保持新的Excel;
代码如下:
import pandas as pd
import numpy as np
from sklearn.feature_extraction import DictVectorizer
import joblib
root_path = "models"
#load data
results = pd.read_csv('datasets/WorldCupMatches.csv', encoding='gbk')
#Adding goal difference and establishing who is the winner
winner = []
for i in range (len(results['Home Team Name'])):
if results ['Home Team Goals'][i] > results['Away Team Goals'][i]:
winner.append(results['Home Team Name'][i])
elif results['Home Team Goals'][i] < results ['Away Team Goals'][i]:
winner.append(results['Away Team Name'][i])
else:
winner.append('Draw')
results['winning_team'] = winner
#adding goal difference column
results['goal_difference'] = np.absolute(results['Home Team Goals'] - results['Away Team Goals'])
# narrowing to team patcipating in the world cup, totally there are 32 football teams in 2022
worldcup_teams = ['Qatar','Germany','Denmark', 'Brazil','France','Belgium', 'Serbia',
'Spain','Croatia', 'Switzerland', 'England','Netherlands', 'Argentina',' Iran',
'Korea Republic','Saudi Arabia', 'Japan', 'Uruguay','Ecuador','Canada',
'Senegal', 'Poland', 'Portugal','Tunisia', 'Morocco','Cameroon','USA',
'Mexico','Wales','Australia','Costa Rica', 'Ghana']
df_teams_home = results[results['Home Team Name'].isin(worldcup_teams)]
df_teams_away = results[results['Away Team Name'].isin(worldcup_teams)]
df_teams = pd.concat((df_teams_home, df_teams_away))
df_teams.drop_duplicates()
df_teams.count()
#dropping columns that wll not affect matchoutcomes
df_teams_new =df_teams[[ 'Home Team Name','Away Team Name','winning_team']]
print(df_teams_new.head() )
#Building the model
#the prediction label: The winning_team column will show "2" if the home team has won, "1" if it was a tie, and "0" if the away team has won.
df_teams_new = df_teams_new.reset_index(drop=True)
df_teams_new.loc[df_teams_new.winning_team == df_teams_new['Home Team Name'],'winning_team']=2
df_teams_new.loc[df_teams_new.winning_team == 'Draw', 'winning_team']=1
df_teams_new.loc[df_teams_new.winning_team == df_teams_new['Away Team Name'], 'winning_team']=0
print(df_teams_new.count() )
df_teams_new.to_csv('datasets/raw_train_data.csv', encoding='gbk', index =False)
数据预处理后会在datasets文件下生成raw_train_data.csv,里面的数据如下:
保持数据集大致步骤如下:
- 加载数据集;
- 利用DictVectorizer对两个队伍的名字进行数字化;
- 保持数据集方便后面训练模型用。
代码如下:
df_teams_new = pd.read_csv('datasets/raw_train_data.csv', encoding='gbk')
feature = df_teams_new[['Home Team Name', 'Away Team Name']]
vec = DictVectorizer(sparse=False)
print(feature.to_dict(orient='records'))
X = vec.fit_transform(feature.to_dict(orient='records'))
X = X.astype('int')
print("===")
print(vec.get_feature_names())
print(vec.feature_names_)
y = df_teams_new[['winning_team']]
y = y.astype('int')
print(X.shape)
print(y.shape)
joblib.dump(vec, root_path + "/vec.joblib")
np.savez('datasets/train_data', x=X, y=y)
以上代码会在datasets文件下生成train_data.npz文件
数据预处理的完整代码如下:
import pandas as pd
import numpy as np
from sklearn.feature_extraction import DictVectorizer
import joblib
root_path = "models"
def reprocess_dataset():
#load data
results = pd.read_csv('datasets/WorldCupMatches.csv', encoding='gbk')
#Adding goal difference and establishing who is the winner
winner = []
for i in range (len(results['Home Team Name'])):
if results ['Home Team Goals'][i] > results['Away Team Goals'][i]:
winner.append(results['Home Team Name'][i])
elif results['Home Team Goals'][i] < results ['Away Team Goals'][i]:
winner.append(results['Away Team Name'][i])
else:
winner.append('Draw')
results['winning_team'] = winner
#adding goal difference column
results['goal_difference'] = np.absolute(results['Home Team Goals'] - results['Away Team Goals'])
# narrowing to team patcipating in the world cup, totally there are 32 football teams in 2022
worldcup_teams = ['Qatar','Germany','Denmark', 'Brazil','France','Belgium', 'Serbia',
'Spain','Croatia', 'Switzerland', 'England','Netherlands', 'Argentina',' Iran',
'Korea Republic','Saudi Arabia', 'Japan', 'Uruguay','Ecuador','Canada',
'Senegal', 'Poland', 'Portugal','Tunisia', 'Morocco','Cameroon','USA',
'Mexico','Wales','Australia','Costa Rica', 'Ghana']
df_teams_home = results[results['Home Team Name'].isin(worldcup_teams)]
df_teams_away = results[results['Away Team Name'].isin(worldcup_teams)]
df_teams = pd.concat((df_teams_home, df_teams_away))
df_teams.drop_duplicates()
df_teams.count()
#dropping columns that wll not affect matchoutcomes
df_teams_new =df_teams[[ 'Home Team Name','Away Team Name','winning_team']]
print(df_teams_new.head() )
#Building the model
#the prediction label: The winning_team column will show "2" if the home team has won, "1" if it was a tie, and "0" if the away team has won.
df_teams_new = df_teams_new.reset_index(drop=True)
df_teams_new.loc[df_teams_new.winning_team == df_teams_new['Home Team Name'],'winning_team']=2
df_teams_new.loc[df_teams_new.winning_team == 'Draw', 'winning_team']=1
df_teams_new.loc[df_teams_new.winning_team == df_teams_new['Away Team Name'], 'winning_team']=0
print(df_teams_new.count() )
df_teams_new.to_csv('datasets/raw_train_data.csv', encoding='gbk', index =False)
def save_dataset():
df_teams_new = pd.read_csv('datasets/raw_train_data.csv', encoding='gbk')
feature = df_teams_new[[ 'Home Team Name','Away Team Name']]
vec = DictVectorizer(sparse=False)
print(feature.to_dict(orient='records'))
X =vec.fit_transform(feature.to_dict(orient='records'))
X = X.astype('int')
print("===")
print(vec.get_feature_names())
print(vec.feature_names_)
y = df_teams_new[[ 'winning_team']]
y =y.astype('int')
print(X.shape)
print(y.shape)
joblib.dump(vec, root_path+"/vec.joblib")
np.savez('datasets/train_data', x= X, y = y)
if __name__ == '__main__':
reprocess_dataset()
save_dataset();
reprocess_dataset() 方法是数据进行预处理。
save_dataset() 方法是对预处理过的数据,进行向量化。
由于这个数据集里没有2018年后的数据,所以我自己又手动收集了一下2018后新的数据。可以把这个新的数据集和上面预处理后数据集合在一起。
2018年和2022年的数据可以下面百度网盘下载
链接: https://pan.baidu.com/s/1fe_z6kRXB8T69wx1HBxO8g 提取码: o55d
5. 模型训练与选择
用不同的传统机器学习方法进行训练,训练后的模型比较
Model | Training Accuracy | Test Accuracy |
Logistic Regression(逻辑回归) | 67.40% | 61.60% |
SVM(支持向量机) | 67.30% | 62.70% |
Naive Bayes(朴素贝叶斯) | 65.50% | 63.80% |
Random Forest(随机森林) | 90.80% | 65.50% |
XGB(X (Extreme) GBoosted) | 75.30% | 62.00% |
可以看到随机森林模型在测试集上准确率最高,所以我们可以用它来做预测。
下面是完整训练代码:
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import matplotlib.ticker as ticker
import matplotlib.ticker as plticker
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn import svm
import sklearn as sklearn
from sklearn.feature_extraction import DictVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.ensemble import RandomForestClassifier
import joblib
from sklearn.metrics import classification_report
from xgboost import XGBClassifier
from sklearn.metrics import confusion_matrix
root_path = "models"
def get_dataset():
train_data = np.load('datasets/train_data.npz')
return train_data
def train_by_LogisticRegression(train_data):
X = train_data['x']
y = train_data['y']
# Separate train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, random_state=42)
logreg = LogisticRegression()
logreg.fit(X_train, y_train)
joblib.dump(logreg, root_path+'/LogisticRegression_model.joblib')
score = logreg.score(X_train, y_train)
score2 = logreg.score(X_test, y_test)
print("LogisticRegression Training set accuracy: ", '%.3f'%(score))
print("LogisticRegression Test set accuracy: ", '%.3f'%(score2))
def train_by_svm(train_data):
X = train_data['x']
y = train_data['y']
# Separate train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, random_state=42)
model = svm.SVC(kernel='linear', verbose=True, probability=True)
model.fit(X_train, y_train)
joblib.dump(model, root_path+'/svm_model.joblib')
score = model.score(X_train, y_train)
score2 = model.score(X_test, y_test)
print("SVM Training set accuracy: ", '%.3f' % (score))
print("SVM Test set accuracy: ", '%.3f' % (score2))
def train_by_naive_bayes(train_data):
X = train_data['x']
y = train_data['y']
# Separate train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, random_state=42)
model = MultinomialNB()
model.fit(X_train, y_train)
joblib.dump(model, root_path+'/naive_bayes_model.joblib')
score = model.score(X_train, y_train)
score2 = model.score(X_test, y_test)
print("naive_bayes Training set accuracy: ", '%.3f' % (score))
print("naive_bayes Test set accuracy: ", '%.3f' % (score2))
def train_by_random_forest(train_data):
X = train_data['x']
y = train_data['y']
# Separate train and test sets
X_train = X
y_train = y
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, random_state=42)
model = RandomForestClassifier(criterion='gini', max_features='sqrt')
model.fit(X_train, y_train)
joblib.dump(model, root_path+'/random_forest_model.joblib')
score = model.score(X_train, y_train)
score2 = model.score(X_test, y_test)
print("random forest Training set accuracy: ", '%.3f' % (score))
print("random forest Test set accuracy: ", '%.3f' % (score2))
def train_by_xgb(train_data):
X = train_data['x']
y = train_data['y']
# Separate train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, random_state=42)
model = XGBClassifier(use_label_encoder=False)
model.fit(X_train, y_train)
joblib.dump(model, root_path+'/xgb_model.joblib')
score = model.score(X_train, y_train)
score2 = model.score(X_test, y_test)
print("xgb Training set accuracy: ", '%.3f' % (score))
print("xgb Test set accuracy: ", '%.3f' % (score2))
y_pred = model.predict(X_test)
report = classification_report(y_test, y_pred, output_dict=True)
# show_confusion_matrix(y_test, y_pred)
print(report)
def show_confusion_matrix(y_true, y_pred, pic_name = "confusion_matrix"):
confusion = confusion_matrix(y_true=y_true, y_pred=y_pred)
print(confusion)
sns.heatmap(confusion, annot=True, cmap= 'Blues', xticklabels=['0','1','2'], yticklabels=['0','1','2'], fmt = '.20g')
plt.xlabel('Predicted class')
plt.ylabel('Actual Class')
plt.title(pic_name)
# plt.savefig('pic/' + pic_name)
plt.show()
if __name__ == '__main__':
train_data = get_dataset()
train_by_LogisticRegression(train_data)
train_by_svm(train_data)
train_by_naive_bayes(train_data)
train_by_random_forest(train_data)
train_by_xgb(train_data)
执行上面代码会生成下面文件:
6. 预测
预测逻辑大致如下:
- 输入两个队伍名字
- 对队伍名字进行验证
- 加载模型
- 利用模型进行预测并输出每个的概率
执行下面预测代码,结果是Ecuador胜于Qatar, 英国队胜于伊朗队。
[2]
[[0.05 0.22033333 0.72966667]]
Probability of Ecuador winning: 0.730
Probability of Draw: 0.220
Probability of Qatar winning: 0.050
[2]
[[0.02342857 0.21770455 0.75886688]]
Probability of England winning: 0.759
Probability of Draw: 0.218
Probability of Iran winning: 0.023
预测的完整代码如下:
import joblib
worldcup_teams = ['Qatar','Germany','Denmark', 'Brazil','France','Belgium', 'Serbia',
'Spain','Croatia', 'Switzerland', 'England','Netherlands', 'Argentina',' Iran',
'Korea Republic','Saudi Arabia', 'Japan', 'Uruguay','Ecuador','Canada',
'Senegal', 'Poland', 'Portugal','Tunisia', 'Morocco','Cameroon','USA',
'Mexico','Wales','Australia','Costa Rica', 'Ghana']
root_path = "models"
def verify_team_name(team_name):
for worldcup_team in worldcup_teams:
if team_name==worldcup_team:
return True
return False
def predict(model_dir =root_path+'/LogisticRegression_model.joblib', team_a='France', team_b = 'Mexico'):
if not verify_team_name(team_a):
print(team_a, ' is not correct')
return
if not verify_team_name(team_b) :
print(team_b, ' is not correct')
return
logreg = joblib.load(model_dir)
input_x = [{'Home Team Name': team_a, 'Away Team Name': team_b}]
vec = joblib.load(root_path+"/vec.joblib")
input_x = vec.transform(input_x)
result = logreg.predict(input_x)
print(result)
result1 = logreg.predict_proba(input_x)
print(result1)
print('Probability of ',team_a , ' winning:', '%.3f'%result1[0][2])
print('Probability of Draw:', '%.3f' % result1[0][1])
print('Probability of ', team_b, ' winning:', '%.3f' % result1[0][0])
if __name__ == '__main__':
team_a = 'Ecuador'
team_b = 'Qatar'
predict('models/random_forest_model.joblib', team_a, team_b)
team_a = 'England'
team_b = ' Iran'
predict('models/random_forest_model.joblib', team_a, team_b)
7. 2018后新的数据
2018年和2022年的数据可以下面百度网盘下载
链接: https://pan.baidu.com/s/1fe_z6kRXB8T69wx1HBxO8g 提取码: o55d
8. 个人总结
特征少的可怜,如果可以加一些球员的信息和状态的特征会更好,数据也相对太少,如果可以把欧洲杯,亚洲杯,非洲杯和美洲杯的每届数据加入进来就好了。数据有点旧(1930年后所有数据),这也是没办法,因为没有其它数据了。
如果预测的是比分,是不是也可以用分类做呢?比如最多可以踢入4个球每个队,当然像西班牙队可以踢进去7个球太罕见了,我们可以忽略不计,哈哈。那就是25个类别。效果会不会好呢?有待检验。