Modelagem e seleção de modelos

Machine Learning for Marketing Analytics in R

Verena Pflieger

Data Scientist at INWT Statistics

logitModelFull <- glm(returnCustomer ~ title + newsletter + 
websiteDesign + ..., family = binomial, churnData)
summary(logitModelFull)

## Coeficientes:
##                          Estimativa  Erro Padrão  z value  Pr(>|z|)
## (Intercept)              -1.49074    0.04930      -30.239  < 2e-16   ***
## titleCompany             -0.21215    0.05286      -4.013   5.99e-05  ***
## titleMrs                  0.03086    0.02953       1.045   0.29586
## newsletter1               0.52373    0.03031      17.280   < 2e-16   ***
## websiteDesign2           -0.45679    0.16267      -2.808   0.00498    **
## websiteDesign3           -0.28800    0.15899      -1.811   0.07007     .
## paymentMethodCredidCard  -0.24192    0.04843      -4.995   5.89e-07  ***
## tvEquipment               -0.51475    1.08141      -0.476   0.63408
...
## Cód. de signif.: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
...
## AIC: 41762

Significância estatística

## Coeficientes:
##               Estimativa  Erro Padrão  z value  Pr(>|z|)
## ...
## newsletter1     0.52373     0.03031    17.280  < 2e-16   ***
## ...

Interpretação dos coeficientes

Equação dos log-odds: $\displaystyle \log \frac{P(\text{returnCustomer}=1)}{P(\text{returnCustomer}=0)} = $ $-1.49 - 0.21 \cdot \text{titleCompany} + 0.52 \cdot \text{newsletter1} + ... $

Transformação para odds:

coefsExp <- coef(logitModelFull) %>% exp() %>% round(2)
coefsExp

## (Intercept)  titleCompany    titleMrs    titleOthers
## 0.23         0.81            1.03        1.77

## newsletter1  websiteDesign2  ...
## 1.69         0.63            ...

Seleção de modelos

library(MASS)
logitModelNew <- stepAIC(logitModelFull, trace = 0)
summary(logitModelNew)

## Coeficientes:
##                          Estimativa  Erro Padrão  z value  Pr(>|z|)
## (Intercept)              -1.49130    0.04928      -30.260  < 2e-16   ***
## titleCompany             -0.21131    0.05285      -3.998   6.38e-05  ***
## titleMrs                  0.03159    0.02951       1.071   0.28432
## newsletter1               0.52332    0.03030      17.269   < 2e-16   *** 
...
## videogameDownload         0.26474    0.05256       5.037   4.74e-07  ***
## prodRemitted              0.89528    0.07619      11.751   < 2e-16   ***

## Cód. de signif.: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
...
## AIC: 41756

Resultados da função step-AIC

Variáveis removidas	Variáveis restantes
tvEquipment	newsletter
prodOthers	paymentMethod
	dvd
	blueray
	...

Vamos aplicar o que eu mostrei!

Machine Learning for Marketing Analytics in R