Feature Engineering, Ta-Feng

變數製作與資料分割流程

Fig-1: Feature Engineering

Fig-2: Feature Engr. & Data Spliting Process

---

Loading & Preparing Data

Sys.setlocale("LC_ALL","C")

[1] "C"

library(dplyr)
library(ggplot2)
library(caTools)
library(Matrix)
library(slam)
library(rpart)
library(rpart.plot)

rm(list=ls(all=TRUE))
load("data/tf2.rdata")
A2 = subset(A, buy)
c(sum(spl), sum(spl2))

[1] 20008  9269

---

Weekday Percentage: W1 ~ W7

X = X %>% mutate(wday = format(date, "%w"))
table(X$wday)

    0     1     2     3     4     5     6 
18011 12615 11288  9898 11245 10651 14587 

mx = xtabs(~ cust + wday, X)
dim(mx)

[1] 28584     7

mx[1:5,]

      wday
cust   0 1 2 3 4 5 6
  1069 1 1 0 0 0 0 0
  1113 2 1 0 0 0 0 1
  1359 0 1 0 0 0 0 0
  1823 0 1 0 1 1 0 0
  2189 0 0 0 1 0 0 1

mx = mx / rowSums(mx)
mx[1:5,]

      wday
cust           0         1         2         3         4         5
  1069 0.5000000 0.5000000 0.0000000 0.0000000 0.0000000 0.0000000
  1113 0.5000000 0.2500000 0.0000000 0.0000000 0.0000000 0.0000000
  1359 0.0000000 1.0000000 0.0000000 0.0000000 0.0000000 0.0000000
  1823 0.0000000 0.3333333 0.0000000 0.3333333 0.3333333 0.0000000
  2189 0.0000000 0.0000000 0.0000000 0.5000000 0.0000000 0.0000000
      wday
cust           6
  1069 0.0000000
  1113 0.2500000
  1359 0.0000000
  1823 0.0000000
  2189 0.5000000

A = data.frame(as.integer(rownames(mx)), as.matrix.data.frame(mx)) %>% 
  setNames(c("cust","W1","W2","W3","W4","W5","W6","W7")) %>% 
  right_join(A, by='cust')
head(A)

---

Classification (Buy) Model

TR = subset(A, spl)
TS = subset(A, !spl)

library(rpart)
library(rpart.plot)
rpart1 = rpart(buy ~ ., TR[,c(2:16,18)], method="class")
pred =  predict(rpart1, TS)[,2]  # predict prob
cm = table(actual = TS$buy, predict = pred > 0.5); cm

       predict
actual  FALSE TRUE
  FALSE  3730  873
  TRUE   1643 2330

acc.ts = cm %>% {sum(diag(.))/sum(.)}; acc.ts   # 0.70662          

[1] 0.7066231

colAUC(pred, TS$buy)                            # 0.6984

                    [,1]
FALSE vs. TRUE 0.6983998

• 利用CART – Classification & Regression Tree建立預測模型
• 使用CART 預測 類別
• 檢視測試資料的準確度
• 檢視AUC

rr rpart.plot(rpart1,cex=0.6)

[1] 0.70417

rr colAUC(pred, TS$buy) # 0.7169

                 [,1]
FALSE vs. TRUE 0.7169

rr rpart.plot(rpart2,cex=0.6)

Regression (Amount) Model

• 由於是預測數量，將資料取Log10可以避免單位不同所造成的數字差異

rr rpart3 = rpart(amount ~ ., TR2[,c(2:17)], cp=0.002) SST = sum((TS2\(amount - mean(TR2\)amount))^ 2) SSE = sum((predict(rpart3, TS2) - TS2$amount)^2) 1 - (SSE/SST)

[1] 0.24606

• 即總變異(SST)=已解釋變異(SSR)+ 未解釋變異(SSE)

---