Business Analysis By Machine Learning

case_s=iteration
head(case_s)

The above data is being analysied for prediction.

str(case_s)
case_s$Gender=factor(case_s$Gender)
case_s$Income=factor(case_s$Income)
case_s$Churn=factor(case_s$Churn)

The above shown is descriptive analysis of the given data.

summary(case_s)

The summary describe the central tendency of each variable

Above graph shows the distribution age vs call.

The graph between Calls and Eductaion.

Now we have to see the correlation between variables.

the redish color shows the strong positive correlation and blue indicates the strong negative correlation.

Now for the model building and prediction we have to split the data into train data and test data.

summary(case_glm)

Call:
glm(formula = Churn ~ ., family = binomial(link = "logit"), data = train_case)

Deviance Residuals: 
     Min        1Q    Median        3Q       Max  
-2.33731  -0.68973   0.09699   0.65465   2.54748  

Coefficients:
             Estimate Std. Error z value Pr(>|z|)    
(Intercept) -10.81025    2.42832  -4.452 8.52e-06 ***
Gender1       1.29634    0.49565   2.615 0.008911 ** 
Age          -0.01809    0.01528  -1.184 0.236330    
Income1       1.96473    0.55288   3.554 0.000380 ***
FamilySize    1.31021    0.33967   3.857 0.000115 ***
Education     0.23322    0.10171   2.293 0.021846 *  
Calls         0.06161    0.02244   2.746 0.006035 ** 
Visits        0.45354    0.16956   2.675 0.007478 ** 
---
Signif. codes:  
0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 166.22  on 119  degrees of freedom
Residual deviance: 105.83  on 112  degrees of freedom
AIC: 121.83

Number of Fisher Scoring iterations: 5

As from the above table,we came to know that according to p-value all the independent variables are significant but since in logistics regression not only p-value is considered but also we have to see the residual error and AIC.

summary(model_case2)

Call:
glm(formula = Churn ~ . - Education, family = "binomial", data = train_case)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-2.5593  -0.7613   0.1213   0.7349   2.5640  

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept) -7.33130    1.67112  -4.387 1.15e-05 ***
Gender1      1.36133    0.48545   2.804 0.005044 ** 
Age         -0.02168    0.01526  -1.421 0.155386    
Income1      1.78174    0.52890   3.369 0.000755 ***
FamilySize   1.30276    0.32335   4.029 5.60e-05 ***
Calls        0.06945    0.02055   3.380 0.000726 ***
Visits       0.45257    0.16397   2.760 0.005779 ** 
---
Signif. codes:  
0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 166.22  on 119  degrees of freedom
Residual deviance: 111.46  on 113  degrees of freedom
AIC: 125.46

Number of Fisher Scoring iterations: 5

Even though Age and Education are less significant according to p-value. but as we exclude them the AIC and residual error tends to increase,Hence we are not excluding either of these variables from our model.

predict_case

          1           2           3           4           5 
0.128561514 0.989731137 0.519249066 0.351355752 0.728860084 
          6           7           8           9          10 
0.149581330 0.980514042 0.763882908 0.934801608 0.292808047 
         11          12          13          14          15 
0.984283756 0.049200219 0.205409181 0.204013141 0.122600549 
         16          17          18          19          20 
0.885635058 0.936579754 0.958000352 0.025657305 0.930563161 
         21          22          23          24          25 
0.527723004 0.209611745 0.193657613 0.707139900 0.515624569 
         26          27          28          29          30 
0.317410941 0.486299606 0.452784687 0.988710589 0.002240695 
         31          32          33          34          35 
0.845787059 0.211410029 0.823854969 0.422863887 0.066522226 
         36          37          38          39          40 
0.988104147 0.065648751 0.784132804 0.469726152 0.934879778 

the above table shows the prediction in terms of probablity.

table_case

            predicted_value
actual_value FALSE TRUE
           0    17    3
           1     3   17

from above confusion matrix shows the number of values truely predicted and falsely predicted

A11=TRUE NEGATIVE (truely predicted as “not survived” )

A22=TRUE POSITIVE (truely predicted as “survived”)

A12=FALSE POSITIVE (falsely predicted as “survived”)

A21=FALSE NEGATIVE (falsely predicted as “not survived”)

Now based on the set threshold we have to check its accuracy and find the best threshold value for max accuracy.

print(paste("Accuracy of prediction is",acc_case*100,"percent"))

[1] "Accuracy of prediction is 85 percent"

The graph above used to determine the optimum value of the threshold in order to get the maximum accuracy.

From the above graph we get to know that threshold near 0.50 is having max accuracy.

The rocr curve for showing relation between true positive value and false positive value.

print(plot(table_case1,col=c("yellow","blue")))

NULL

print(paste("Accuracy of prediction is",acc_case*100,"percent"))

[1] "Accuracy of prediction is 85 percent"

As the prediction accuracy is 85%,it concludes that our prediction model is good.