A New Approach for Construction of Geodemographic Segmentation Model and Prediction Analysis

Table 3

Comparison among models for feature selection.

Features	Backward elimination		Forward selection		Bidirectional selection
Features selected	CreditScore; Age; NumOfProducts; IsActiveMember; Female; Germany; Tenure; Log_Balance		CreditScore; Age; NumOfProducts; IsActiveMember; Female; Germany; Log_Balance		CreditScore; Age; NumOfProducts; IsActiveMember; Female; Germany; Tenure; Log_Balance
Adjusted R-squared	0.151006		0.149718		0.152092
Correctly predicted cases	8127 (81.3%)		8122 (81.1%)		8127 (81.3%)
CAP curve (training data)	Accuracy, 81%		Accuracy, 80.5%		Accuracy, 80.5%
CAP curve (test data)	Accuracy, 78%		Accuracy, 77%		Accuracy, 78%
Accuracy drop	3%		3.5%		2.5%
Multicollinearity	CreditScore	1.001	CreditScore	1.001	CreditScore	1.001
	Age	1.012	Age	1.012	Age	1.012
	NumOfProducts	1.151	NumOfProducts	1.151	NumOfProducts	1.151
	IsActiveMember	1.010	IsActiveMember	1.009	IsActiveMember	1.010
	Female	1.003	Female	1.003	Female	1.003
	Germany	1.269	Germany	1.269	Germany	1.269
	Tenure	1.001	Log_Balance	1.420	Tenure	1.001
	Log_Balance	1.420			Log_Balance	1.420