Linear Regression

Linear_Regression

Linear Regression

Salary Prediction

te2.jpg

Tennis Ace

Project Goals

  • Create a linear regression model that predicts the earnings of tennis players based on their playing habits. By analyzing and modeling the Association of Tennis Professionals (ATP) data.

Target/Outcomes

  • Winnings: total winnings in USD($) in a year

Data Dictionary:

–Not Available–

Content

  1. Import and Load Data
  2. Preliminary Linear Regression
    • 2.1 Building a Linear Regression Model
    • 2.2 Model Accuracy
  3. Feature Selection
    • 3.1 Wrapper Methods
    • 3.2 RFE
    • 3.3 Select K best
    • 3.4 Further Observation
  4. Final Linear Regression
    • 4.1 Building a Linear Regression Model
    • 4.2 Final Regression Evaluation
    • 4.3 Test Run
  5. Conclusion

1.Import and Load Data

In [1]:
import pandas as pd
import numpy as np
from scipy.stats import pearsonr, skew

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LinearRegression, Lasso
from sklearn.feature_selection import mutual_info_regression, RFE, SelectKBest, f_regression
from functools import partial

from mlxtend.feature_selection import SequentialFeatureSelector as SFS
from mlxtend.plotting import plot_sequential_feature_selection as plot_sfs

import seaborn as sns
import matplotlib.pyplot as plt
import warnings

warnings.filterwarnings('ignore')
pd.set_option('display.max_columns', None)
In [4]:
# load data
df = pd.read_csv('tennis_stats.csv')
print(df.shape)
df.head(2)
(1721, 24)
Out[4]:
Player Year FirstServe FirstServePointsWon FirstServeReturnPointsWon SecondServePointsWon SecondServeReturnPointsWon Aces BreakPointsConverted BreakPointsFaced BreakPointsOpportunities BreakPointsSaved DoubleFaults ReturnGamesPlayed ReturnGamesWon ReturnPointsWon ServiceGamesPlayed ServiceGamesWon TotalPointsWon TotalServicePointsWon Wins Losses Winnings Ranking
0 Pedro Sousa 2016 0.88 0.50 0.38 0.50 0.39 0 0.14 7 7 0.43 2 9 0.11 0.38 8 0.50 0.43 0.50 1 2 39820 119
1 Roman Safiullin 2017 0.84 0.62 0.26 0.33 0.07 7 0.00 7 0 0.57 1 9 0.00 0.20 9 0.67 0.41 0.57 0 1 17334 381

2. Preliminary Linear Regression

We will start by building a Linear Regression with all the features available and compare it later to our final logistic model.

2.1 Building a Linear Regression Model

Split the data into X and y
In [3]:
y = df['Winnings']
X = df.drop(columns=['Winnings', 'Player'])
Train Test split

Spliting our data to training 70% and testing 30%.

In [4]:
x_train, x_test, y_train, y_test = train_test_split(X, y,  test_size = 0.30, random_state = 0)
Normalize the Data

Our features(X) must be in the same scale

In [5]:
scaler = StandardScaler()
x_train = scaler.fit_transform(x_train)
x_test = scaler.transform(x_test)
Linear Regression Model
In [6]:
lr = LinearRegression()
Fit the model

Use the .fit() method on lr to fit the model to X and y.

In [7]:
lr.fit(x_train, y_train)
Out[7]:
LinearRegression()
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

2.2 Model accuracy

Note:

  • Overfitting, It fits the training data well but performs significantly worse on test data.
  • When data is overfit, we do regularization (LAsso, Ridge)
R2 score
In [8]:
print('Training Score: ' + str(lr.score(x_train,y_train)))
print('Testing Score: ' + str(lr.score(x_test,y_test)))
Training Score: 0.8931239790692806
Testing Score: 0.8909760018805483
Mean Squared Error (Training and Testing Dataset)
In [9]:
# high difference is an indicator of overfitting

# prediction from training set
pred_train = lr.predict(x_train)
# prediction from testing set
pred_test = lr.predict(x_test)

# MSE, prediction and actual from training set
MSE_train = np.mean((pred_train - y_train)**2)
print("Training Error: ", MSE_train)

# MSE, prediction and actual from testing set
MSE_test = np.mean((pred_test - y_test)**2)
print("Testing Error: ", MSE_test)

print("Difference %: ", abs((MSE_train - MSE_test) / MSE_train)  )
Training Error:  7143338538.850752
Testing Error:  6216637074.812809
Difference %:  0.12972946179126968

3. Feature Selection

Finding the most suitable features to improve our model score/performance.

3.1 Wrapper Methods

Now that we’ve created a linear regression model and evaluated its performance, we’re ready to do some feature selection.

First lets define a variables of our feature selection methods.

  1. number_of_features the number of features
  2. scoring_method;
    • for classifiers {‘accuracy’, ‘f1’, ‘precision’, ‘recall’, ‘roc_auc’}
    • for regressors {‘r2′, ‘max_error’,’neg_mean_squared_error’, ‘median_absolute_error’}

Documentation; https://scikit-learn.org/stable/modules/model_evaluation.html

Note: We’re fitting the the orignal X and y variables here instead of the training and testing data set.

We will us r2 as our scoring method.

Important formula:

  • y = actual values

  • Å· = predicted value

    • r2 = 1 – (tss / rss)
    • tss = sum (y – mean_of_y)^2
    • rss = sum(y – Å·)^2
In [10]:
# Input variable here:
number_of_features = 8
scoring_method = 'r2'
Sequencial Forward Floating Selection
In [11]:
sffs = SFS(lr,
        k_features=number_of_features,
        forward= True,
        floating= True,
        scoring=scoring_method,
        cv=0)

# Note: our X values here are not standarized or
# does sffs automatically standarized our features?
# should I just fit the x_train, y_train here?
sffs.fit(X, y)
print('SFFS selected features:')

# Saving a copy of sffs selected features in a list variable
X_sffs = list(sffs.subsets_[number_of_features]['feature_names'])

print(X_sffs)    
print('\n' + str(scoring_method) + ': ' +  str(sffs.subsets_[number_of_features]['avg_score']))
SFFS selected features:
['Year', 'BreakPointsFaced', 'BreakPointsOpportunities', 'DoubleFaults', 'ServiceGamesPlayed', 'Wins', 'Losses', 'Ranking']

r2: 0.8923341769921913
Sequencial Backward Floating Selection
In [12]:
sbfs = SFS(lr,
            k_features=1,
            forward= False,
            floating= True,
            scoring=scoring_method,
            cv=0)
sbfs.fit(X, y)
print('SBFS selected features:')

# Saving a copy of sbfs selected features in a list variable
X_sbfs = list(sbfs.subsets_[1]['feature_names'])

print(X_sbfs)
print('\n' + str(scoring_method) + ': ' +  str(sbfs.subsets_[1]['avg_score']))
SBFS selected features:
['Wins']

r2: 0.8341119961391337

Wins has the higest r2 score.

Visualize SFFS and SBFS

In [13]:
# Visualize sffs
# Visualize in DataFrame (Optional, Uncomment to view)
# pd.DataFrame.from_dict(ssfs.get_metric_dict()).T

fig1 = plot_sfs(sffs.get_metric_dict(), kind='std_dev')
plt.title('Sequential Forward Floating Selection (w. StdDev)')
plt.grid()
plt.show()

#Visualize sbfs
fig2 = plot_sfs(sbfs.get_metric_dict(), kind='std_dev')
plt.title('')
plt.grid()
plt.show()

The graph shows that there’s not much variation from using 8 to 21 features though theres is a very very tiny increase using all the features, it’s not enough justification and using all may lead to redundant information. We can simplify by just using 8 features instead.

3.2 RFE

Ranking features by importance, discarding the least important features, and re-fitting the model. This process is repeated until a specified number of features remains.

Before doing applying recursive feature elimination it is necessary to standardize the data.

In [14]:
# Setting a varible 'X_standard' so that our original X features will remain unaffected
# if we rerun this code
X_standard = StandardScaler().fit_transform(X)
    
rfe = RFE(lr, n_features_to_select=number_of_features)
    
rfe.fit(X_standard,y)
    
rfe_features = [f for (f, support) in zip(X, rfe.support_) if support]
print('RFE selected features:')
X_rfe = rfe_features
print(X_rfe)
print('\nrfe score: '+ str(rfe.score(X_standard,y)))
RFE selected features:
['Year', 'BreakPointsFaced', 'BreakPointsOpportunities', 'DoubleFaults', 'ServiceGamesPlayed', 'Wins', 'Losses', 'Ranking']

rfe score: 0.8923341769921913

3.3 SelectKBest

We will be using the f_regression and mutual_info_regression for this test

Select K Best most common test methods;

f_classif ANOVA F-value between label/feature for classification tasks.

mutual_info_classif Mutual information for a discrete target.

chi2 Chi-squared stats of non-negative features for classification tasks.

f_regression F-value between label/feature for regression tasks.

mutual_info_regression Mutual information for a continuous target.

SelectKBest f_regression

In [15]:
# because we want to specify additional arguments (random_state=0) 
# besides the features and targets inputs, we’ll need the help of the partial()
# score_function = partial(test_method, random_state=0)
selection = SelectKBest(score_func = f_regression, k = number_of_features)
 
# fit the fata    
selection.fit_transform(X, y)

# saving a copy of chi2 slected features in a list
X_f_regression = X[X.columns[selection.get_support(indices=True)]]
X_f_regression = list(X_f_regression)
X_f_regression
Out[15]:
['Aces',
 'BreakPointsFaced',
 'BreakPointsOpportunities',
 'DoubleFaults',
 'ReturnGamesPlayed',
 'ServiceGamesPlayed',
 'Wins',
 'Losses']

SelectKBest mutual_info_regression

In [16]:
selection = SelectKBest(score_func = mutual_info_regression, k = number_of_features)
 
# fit the fata    
selection.fit_transform(X, y)

# saving a copy of mutual_info_classifslected features in a list
X_mutual_info_regression = X[X.columns[selection.get_support(indices=True)]]
X_mutual_info_regression = list(X_mutual_info_regression)
X_mutual_info_regression
Out[16]:
['Aces',
 'BreakPointsFaced',
 'BreakPointsOpportunities',
 'DoubleFaults',
 'ReturnGamesPlayed',
 'ServiceGamesPlayed',
 'Wins',
 'Losses']

3.4 Further Observation

Almost half of our features are corellated to our target. However we want to minimize our features to avoid redundant information and multicolinearity as possible. As observed our selection methods accept the losses feature which in contrast with our general knowledge we expect the opposite.

Lets view it in a scatter plot.

In [17]:
plt.scatter(df['Losses'], df['Winnings'])
plt.title('Losses and Winnings')
plt.xlabel('Losses')
plt.ylabel('Winnings')
Out[17]:
Text(0, 0.5, 'Winnings')
In [18]:
corr, p = pearsonr(df['Losses'], df['Winnings'])
corr, p
Out[18]:
(0.8692767614021005, 0.0)

As seen in the graph, Losses is 86.92% correlated to Winnings. I suspect that number of losses and winnings contribute to overall number of ServiceGamesPlayed.

The graph below show’s the correlation between wins, losses and service games played(number of game of a player).

In [19]:
# show plot
plt.figure(figsize=(15, 5)) 

ax1 = plt.subplot(1,2,1)
plt.scatter(df['ServiceGamesPlayed'], df['Losses'])
plt.xlabel('Service Games Played')
plt.ylabel('Losses')

ax1 = plt.subplot(1,2,2)
plt.scatter(df['ServiceGamesPlayed'], df['Wins'] )
plt.xlabel('Service Games Played')
plt.ylabel('Wins')
Out[19]:
Text(0, 0.5, 'Wins')

The number of ServiceGamesplayed produces the number of Wins and Losses affecting our target Winnings value. This is an example of duplicate information, we can drop the one that has lower correlation to our target. We can view that information with the use of correlation heapmap.

In [20]:
# Correlation between features and target
plt.figure(figsize=(8, 8)) 
corr_matrix = df.corr()
 
# Isolate the column corresponding to `Winnings`
corr_target = corr_matrix[['Winnings']].drop(labels=['Winnings'])
corr_target = corr_target.sort_values(by=['Winnings'], ascending=False)

sns.heatmap(corr_target, annot=True, fmt='.3', cmap='RdBu_r')
Out[20]:
<AxesSubplot:>

The correlation table clearly display that Wins has the higher correlation to Winnings than losses. We will drop losses in this case and replace it with ReturnGamesPlayed.

Creating our own feature list:

In [21]:
X_sffs
Out[21]:
['Year',
 'BreakPointsFaced',
 'BreakPointsOpportunities',
 'DoubleFaults',
 'ServiceGamesPlayed',
 'Wins',
 'Losses',
 'Ranking']
In [23]:
X_my_list =['Year',
            'BreakPointsFaced',
            'BreakPointsOpportunities',
            'DoubleFaults',
            'ServiceGamesPlayed',
            'Wins',
#             'Losses',
            'Ranking']

4. Final Linear Regression

Linear Regression with selected features.

4.1 Building Final Logistic Regression

So far we have six different list of selected features, What we can do for now is to try each list and find the one which has the highest scores.

In this case we will using the X_my_list.

  1. X_my_list
  2. X_sffs
  3. X_sbfs
  4. X_rfe
  5. X_f_regression
  6. X_mutual_info_regression
In [29]:
X = df[X_my_list]
y = df['Winnings']
In [30]:
x_train, x_test, y_train, y_test = train_test_split(X, y,  test_size = 0.30, random_state = 0)
In [31]:
scaler = StandardScaler()
x_train = scaler.fit_transform(x_train)
x_test = scaler.transform(x_test)

lr = LinearRegression()

lr.fit(x_train, y_train)
Out[31]:
LinearRegression()
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

We have narrowed down our features from 21 to just 8 without sacrificing too much of our model performance.

4.2 Final Regression Evaluation

Our regresion model score.

In [32]:
# r2 score
# r2 ( 1- (TSS / RSS)

# TSS = how far each observation from mean 
#       sum of (each actual - its mean value)^2

# RSS = sum squared residual (lower the better) 

print(lr.score(x_train,y_train))
print(lr.score(x_test,y_test))
0.8913335512852711
0.8890348563475029
In [33]:
# Mean squared error
# (actual - predicted)^2 / number of observation
# high difference is an indicator of overfitting

#Training Error
pred_train = lr.predict(x_train)
MSE_train = np.mean((pred_train - y_train)**2)
print("Training Error: ", MSE_train)

# 2. Testing Error
pred_test = lr.predict(x_test)
MSE_test = np.mean((pred_test - y_test)**2)
print("Testing Error: ", MSE_test)

print("Difference %: ", abs((MSE_train - MSE_test) / MSE_train)  )
Training Error:  7263006465.099948
Testing Error:  6327322772.4251375
Difference %:  0.12882870160875373

Using just 8 features our final regression model score has no significant difference from our preliminary model with all the features.

Visualization of actual and predicted earnings.

In [34]:
y_predict = lr.predict(x_test)
In [35]:
# visualization of actual vs. predicted value
plt.figure(figsize=(8, 6)) 

# y_test is 20% of our data
plt.scatter(y_test, y_predict, alpha=0.4, color = 'red')
plt.title("Actual Vs. Predicted Winnings")
plt.xlabel("Actual Winnings")
plt.ylabel("Predicted Winnings")
Out[35]:
Text(0, 0.5, 'Predicted Winnings')
In [36]:
# show/hide code
a = y_test.values.reshape(-1,1).flatten()
b = y_predict.flatten()
diff = (b - a)

sim_data={"Actual Winnings":a, 'Predicted Winnings':b, 'Difference':np.round(diff,2)}

sim_data=pd.DataFrame(sim_data)

# Showing first 5 rows
sim_data.head(5)
Out[36]:
Actual Winnings Predicted Winnings Difference
0 492583 486891.437854 -5691.56
1 53210 24305.827453 -28904.17
2 123905 94160.366388 -29744.63
3 93606 89217.295690 -4388.70
4 25260 43458.181254 18198.18

4.3 Model Test Run

Let say we have a very talented tennis player named Alyx. We extracted his gameplay record and save it in a variable alyx_stats.

  • What will be Alyx earning in a year with his statistical record below?

Note: alyx_stats are based on df index number 465. We intentionaly copy this stats so that we can compare it to the real value in pur sim_data.

In [37]:
df[X_my_list].columns
Out[37]:
Index(['Year', 'BreakPointsFaced', 'BreakPointsOpportunities', 'DoubleFaults',
       'ServiceGamesPlayed', 'Wins', 'Ranking'],
      dtype='object')
In [39]:
# showing details for index 570
index_465 = df[df['Winnings'].isin([492583])]
index_465[['Year', 'BreakPointsFaced', 'BreakPointsOpportunities', 'DoubleFaults',
       'ServiceGamesPlayed', 'Wins', 'Ranking']]
Out[39]:
Year BreakPointsFaced BreakPointsOpportunities DoubleFaults ServiceGamesPlayed Wins Ranking
465 2015 284 252 84 450 16 24
In [40]:
alyx_record = [[2015, 284, 252, 84, 450 ,16 ,24]]
In [41]:
# # transform alyx_stats
alyx_record = scaler.transform(alyx_record)
alyx_record
Out[41]:
array([[ 0.55489652,  1.38014095,  1.15494622,  0.56674657,  1.08969239,
         0.75124826, -0.8924122 ]])
In [42]:
alyx_earning = lr.predict(alyx_record)
alyx_earning
Out[42]:
array([486891.4378537])
In [43]:
sim_data.iloc[[0]]
Out[43]:
Actual Winnings Predicted Winnings Difference
0 492583 486891.437854 -5691.56

5. Conclusion

We are comparing continuous features and for this dataset the best method for feature selection is to check for their correlations.

Summary:

  1. We build a preliminary linear regression with all the features included.
  2. Used various feature selection methods combined with our general knowledge to select the best features that most describe our target. Narrowed it down from 24 to just only 8 features.
  3. Re create our final linear regression model and compare it’s performance with our preliminary model.
  4. Finally, we evaluate and test run our final regression model and got an 89% r2 score on both training and testing dataset.

– fin

Leave a Reply