Diferentes Fontes de Dados no contexto de Modelagem Uplift
Angelica Custodio
Setembro de 2020
Utilizaremos no ajuste dos Modelos Uplift três diferentes bases de dados
- Criteo Uplift Modeling Dataset
Esse dataset é relacionado ao artigo “A Large Scale Benchmark for Uplift Modeling” - Eustache Diemert, Artem Betlei, Christophe Renaudin; (Criteo AI Lab), Massih-Reza Amini (LIG, Grenoble INP). Disponível em: https://ailab.criteo.com/criteo-uplift-prediction-dataset/.
Os dados em questao sao resultantes de varios testes incrementais, onde uma parte aleatoria da populacao é controle ( impedida de ser alvo da acao) e a outra tratamento. Consiste de 25 milhoes de linhas, cada uma representando o usuario com 11 variaveis, um indicador de tratamento e 2 rotulos (visitas e convers?es).
Descricao detalhada dos campos:
f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11: variaveis descaracterizadas que foram observadas tratamento: grupo de tratamento (1 = tratado, 0 = controle) conversao: se ocorreu uma conversao para este usuario visit: se uma visita ocorreu para este usuario exposicao: efeito do tratamento, se o usuario foi efetivamente exposto (binario).
- Dados Sintéticos para Modelagem Uplift
Esses dados foram disponibilizados em Fevereiro de 2020, em: https://zenodo.org/record/3653141#.X1orrWdKhQJ.
- Dados Campanha de Marketing do Kaggle
Desafio no Kaggle com dados de uma Campanha de Marketing, ele foi abordado como um problema no contexto de Modelagem Uplift. Disponível em: https://www.kaggle.com/davinwijaya/customer-retention.
Tabela data3.csv
library(data.table)
library(dplyr)##
## Attaching package: 'dplyr'## The following objects are masked from 'package:data.table':
##
## between, first, last## The following objects are masked from 'package:stats':
##
## filter, lag## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, uniondata=fread('amostra_criteo_uplift')## Warning in fread("amostra_criteo_uplift"): Detected 16 column names but the
## data has 17 columns (i.e. invalid file). Added 1 extra default column name for
## the first column which is guessed to be row names or an index. Use setnames()
## afterwards if this guess is not correct, or fix the file write command that
## created the file to create a valid file.summary(data)## V1 f0 f1 f2
## Min. : 1 Min. :-1.9463 Min. :3.264 Min. :8.272
## 1st Qu.: 25001 1st Qu.:-1.0118 1st Qu.:3.264 1st Qu.:8.272
## Median : 50000 Median : 1.6254 Median :3.264 Median :8.272
## Mean : 50000 Mean : 0.6148 Mean :3.266 Mean :8.505
## 3rd Qu.: 75000 3rd Qu.: 1.9920 3rd Qu.:3.264 3rd Qu.:8.748
## Max. :100000 Max. : 1.9920 Max. :4.503 Max. :9.337
## f3 f4 f5 f6
## Min. :1.682 Min. :3.507 Min. :10.16 Min. :-6.5997
## 1st Qu.:3.736 1st Qu.:3.507 1st Qu.:10.16 1st Qu.: 0.9452
## Median :3.736 Median :3.507 Median :10.16 Median : 2.3861
## Mean :3.656 Mean :3.522 Mean :10.20 Mean : 1.7535
## 3rd Qu.:3.736 3rd Qu.:3.507 3rd Qu.:10.16 3rd Qu.: 2.9817
## Max. :3.736 Max. :6.750 Max. :16.98 Max. : 2.9817
## f7 f8 f9 f10
## Min. :-4.7244 Min. :-31.747 Min. : 9.85 Min. :-13.037
## 1st Qu.:-0.1667 1st Qu.: -2.716 1st Qu.: 9.85 1st Qu.: -1.861
## Median :-0.1667 Median : 1.108 Median : 9.85 Median : -1.861
## Mean :-0.2930 Mean : -2.045 Mean :10.67 Mean : -2.084
## 3rd Qu.:-0.1667 3rd Qu.: 1.108 3rd Qu.: 9.85 3rd Qu.: -1.861
## Max. :-0.1667 Max. : 1.108 Max. :28.10 Max. : -1.861
## f11 treatment conversion visit
## Min. :4.158 Min. :0.0000 Min. :0.0000 Min. :0.00000
## 1st Qu.:4.158 1st Qu.:1.0000 1st Qu.:0.0000 1st Qu.:0.00000
## Median :4.158 Median :1.0000 Median :0.0000 Median :0.00000
## Mean :4.159 Mean :0.8459 Mean :0.0021 Mean :0.04204
## 3rd Qu.:4.158 3rd Qu.:1.0000 3rd Qu.:0.0000 3rd Qu.:0.00000
## Max. :4.795 Max. :1.0000 Max. :1.0000 Max. :1.00000
## exposure
## Min. :0.00000
## 1st Qu.:0.00000
## Median :0.00000
## Mean :0.03487
## 3rd Qu.:0.00000
## Max. :1.00000#retirando a coluna de index da base
datan=select(data,-V1)Analisando as variáveis da base de dados.
library(tidyverse)## ── Attaching packages ─────────────────────────────────────────────────────────────────────── tidyverse 1.3.0 ──## ✓ ggplot2 3.3.0 ✓ purrr 0.3.3
## ✓ tibble 3.0.0 ✓ stringr 1.4.0
## ✓ tidyr 1.0.2 ✓ forcats 0.5.0
## ✓ readr 1.3.1## ── Conflicts ────────────────────────────────────────────────────────────────────────── tidyverse_conflicts() ──
## x dplyr::between() masks data.table::between()
## x dplyr::filter() masks stats::filter()
## x dplyr::first() masks data.table::first()
## x dplyr::lag() masks stats::lag()
## x dplyr::last() masks data.table::last()
## x purrr::transpose() masks data.table::transpose()library(skimr)
#Visao inicial dos dados
glimpse(datan)## Rows: 100,000
## Columns: 16
## $ f0 <dbl> 1.9919806, 1.9919806, 1.9919806, 1.9919806, 1.9919806, 1.9…
## $ f1 <dbl> 3.263641, 3.263641, 3.263641, 3.263641, 3.263641, 3.263641…
## $ f2 <dbl> 8.272483, 8.272483, 9.003290, 8.272483, 8.821857, 8.272483…
## $ f3 <dbl> 3.735871, 3.735871, 3.735871, 3.735871, 3.735871, 3.735871…
## $ f4 <dbl> 3.506733, 3.506733, 3.506733, 3.506733, 3.506733, 3.506733…
## $ f5 <dbl> 10.16128, 10.16128, 10.16128, 10.16128, 10.16128, 10.16128…
## $ f6 <dbl> 2.98172089, 2.98172089, 2.98172089, 2.98172089, 2.98172089…
## $ f7 <dbl> -0.1666894, -0.1666894, -0.1666894, -0.1666894, -0.1666894…
## $ f8 <dbl> 1.107571, 1.107571, -7.012504, 1.107571, -1.784858, 1.1075…
## $ f9 <dbl> 9.850093, 9.850093, 9.850093, 9.850093, 11.700187, 9.85009…
## $ f10 <dbl> -1.8609, -1.8609, -1.8609, -1.8609, -1.8609, -1.8609, -1.8…
## $ f11 <dbl> 4.157648, 4.157648, 4.157648, 4.157648, 4.157648, 4.157648…
## $ treatment <int> 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1…
## $ conversion <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
## $ visit <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
## $ exposure <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…datan %>%
mutate_if(is.character,as.factor) %>%
skimr::skim()| Name | Piped data |
| Number of rows | 100000 |
| Number of columns | 16 |
| _______________________ | |
| Column type frequency: | |
| numeric | 16 |
| ________________________ | |
| Group variables | None |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| f0 | 0 | 1 | 0.61 | 1.54 | -1.95 | -1.01 | 1.63 | 1.99 | 1.99 | ▃▂▁▁▇ |
| f1 | 0 | 1 | 3.27 | 0.02 | 3.26 | 3.26 | 3.26 | 3.26 | 4.50 | ▇▁▁▁▁ |
| f2 | 0 | 1 | 8.50 | 0.36 | 8.27 | 8.27 | 8.27 | 8.75 | 9.34 | ▇▁▁▁▂ |
| f3 | 0 | 1 | 3.66 | 0.24 | 1.68 | 3.74 | 3.74 | 3.74 | 3.74 | ▁▁▁▁▇ |
| f4 | 0 | 1 | 3.52 | 0.11 | 3.51 | 3.51 | 3.51 | 3.51 | 6.75 | ▇▁▁▁▁ |
| f5 | 0 | 1 | 10.20 | 0.23 | 10.16 | 10.16 | 10.16 | 10.16 | 16.98 | ▇▁▁▁▁ |
| f6 | 0 | 1 | 1.75 | 1.67 | -6.60 | 0.95 | 2.39 | 2.98 | 2.98 | ▁▁▁▂▇ |
| f7 | 0 | 1 | -0.29 | 0.66 | -4.72 | -0.17 | -0.17 | -0.17 | -0.17 | ▁▁▁▁▇ |
| f8 | 0 | 1 | -2.05 | 5.47 | -31.75 | -2.72 | 1.11 | 1.11 | 1.11 | ▁▁▁▁▇ |
| f9 | 0 | 1 | 10.67 | 2.29 | 9.85 | 9.85 | 9.85 | 9.85 | 28.10 | ▇▁▁▁▁ |
| f10 | 0 | 1 | -2.08 | 1.36 | -13.04 | -1.86 | -1.86 | -1.86 | -1.86 | ▁▁▁▁▇ |
| f11 | 0 | 1 | 4.16 | 0.02 | 4.16 | 4.16 | 4.16 | 4.16 | 4.80 | ▇▁▁▁▁ |
| treatment | 0 | 1 | 0.85 | 0.36 | 0.00 | 1.00 | 1.00 | 1.00 | 1.00 | ▂▁▁▁▇ |
| conversion | 0 | 1 | 0.00 | 0.05 | 0.00 | 0.00 | 0.00 | 0.00 | 1.00 | ▇▁▁▁▁ |
| visit | 0 | 1 | 0.04 | 0.20 | 0.00 | 0.00 | 0.00 | 0.00 | 1.00 | ▇▁▁▁▁ |
| exposure | 0 | 1 | 0.03 | 0.18 | 0.00 | 0.00 | 0.00 | 0.00 | 1.00 | ▇▁▁▁▁ |
Trabalhando com os dados para depois ajustar um modelo.
#instalando uplift
require(uplift)## Loading required package: uplift## Loading required package: RItools## Loading required package: SparseM##
## Attaching package: 'SparseM'## The following object is masked from 'package:base':
##
## backsolve## Loading required package: MASS##
## Attaching package: 'MASS'## The following object is masked from 'package:dplyr':
##
## select## Loading required package: coin## Loading required package: survival## Loading required package: tables## Loading required package: penalized## Welcome to penalized. For extended examples, see vignette("penalized").#funcao que transforma a variavel resposta - rvtu
require(plyr)## Loading required package: plyr## ------------------------------------------------------------------------------## You have loaded plyr after dplyr - this is likely to cause problems.
## If you need functions from both plyr and dplyr, please load plyr first, then dplyr:
## library(plyr); library(dplyr)## ------------------------------------------------------------------------------##
## Attaching package: 'plyr'## The following object is masked from 'package:purrr':
##
## compact## The following objects are masked from 'package:dplyr':
##
## arrange, count, desc, failwith, id, mutate, rename, summarise,
## summarizedata_rvtu <- rvtu(visit~f1+f2+f3+f4+f5+f6+f7+f8+f9+f10+f11+trt(as.numeric(treatment)),datan,method="none")
names(data_rvtu)## [1] "f1" "f2" "f3" "f4" "f5" "f6" "f7" "f8" "f9" "f10" "f11" "ct"
## [13] "y" "z"explore(y~f1+f2+f3+f4+f5+f6+f7+f8+f9+f10+f11+trt(ct),
data=data_rvtu)## $f1
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0)
## [3.2636406,3.263641] 15291 83728 0.0269
## (3.263641,4.502868] 114 867 0.1316
## Mean Resp.(Treat=1) Uplift
## [3.2636406,3.263641] 0.0425 0.0156
## (3.263641,4.502868] 0.2491 0.1176
##
## $f2
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [8.27,8.75] 11999 63001 0.0307 0.0506
## (8.75,9.34] 3406 21594 0.0173 0.0274
## Uplift
## [8.27,8.75] 0.0199
## (8.75,9.34] 0.0101
##
## $f3
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [1.68,3.74] 15405 84595 0.0277 0.0446
## Uplift
## [1.68,3.74] 0.0169
##
## $f4
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [3.51,6.75] 15405 84595 0.0277 0.0446
## Uplift
## [3.51,6.75] 0.0169
##
## $f5
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1) Uplift
## [10.2,17] 15405 84595 0.0277 0.0446 0.0169
##
## $f6
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [-6.6,0.945] 4200 21637 0.0405 0.0910
## (0.945,2.39] 5430 21569 0.0131 0.0233
## (2.39,2.98] 5775 41389 0.0322 0.0316
## Uplift
## [-6.6,0.945] 0.0505
## (0.945,2.39] 0.0102
## (2.39,2.98] -0.0007
##
## $f7
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0)
## [-4.724423,-0.1666894] 15259 83785 0.0277
## (-0.1666894,-0.16668936] 146 810 0.0342
## Mean Resp.(Treat=1) Uplift
## [-4.724423,-0.1666894] 0.0448 0.0171
## (-0.1666894,-0.16668936] 0.0309 -0.0034
##
## $f8
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [-31.7,-2.72] 3407 24461 0.1071 0.1360
## (-2.72,1.11] 11998 60134 0.0052 0.0075
## Uplift
## [-31.7,-2.72] 0.0288
## (-2.72,1.11] 0.0023
##
## $f9
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1)
## [9.85,28.1] 15405 84595 0.0277 0.0446
## Uplift
## [9.85,28.1] 0.0169
##
## $f10
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0)
## [-13.03661,-1.8609] 15261 83770 0.0278
## (-1.8609,-1.8608999] 144 825 0.0208
## Mean Resp.(Treat=1) Uplift
## [-13.03661,-1.8609] 0.0448 0.0170
## (-1.8609,-1.8608999] 0.0315 0.0107
##
## $f11
## N(Treat=0) N(Treat=1) Mean Resp.(Treat=0) Mean Resp.(Treat=1) Uplift
## [4.16,4.8] 15405 84595 0.0277 0.0446 0.0169# targeted
count(data_rvtu[data_rvtu$ct == 1,], "y")$freq / sum(data_rvtu$ct == 1)## [1] 0.95535197 0.04464803# control
count(data_rvtu[data_rvtu$ct == 0,], "y")$freq / sum(data_rvtu$ct == 0)## [1] 0.97228173 0.02771827Modelagem considerando Regressão Logistica
#ajustando uma logistico simples
logit.formula <- ~f1+f2+f3+f4+f5+f6+f7+f8+f9+f10+f11
#testar intera??es
set.seed(123)
require(glmnet)## Loading required package: glmnet## Loading required package: Matrix##
## Attaching package: 'Matrix'## The following objects are masked from 'package:tidyr':
##
## expand, pack, unpack## Loaded glmnet 3.0-2#library(tidyr)
logit.x.interactions <- model.matrix(logit.formula, data=data_rvtu)
logit.z <- data_rvtu$z
logit.y <- data_rvtu$y
# traditional classifier, y as response
logit.cv.lasso.y.interactions <- cv.glmnet(logit.x.interactions, logit.y, alpha=1, family="binomial")
plot(logit.cv.lasso.y.interactions)
# uplift classifier, z as response
logit.cv.lasso.z.interactions <- cv.glmnet(logit.x.interactions, logit.z, alpha=1, family="binomial")
plot(logit.cv.lasso.z.interactions)
coef(logit.cv.lasso.z.interactions)[which(coef(logit.cv.lasso.z.interactions) != 0),]## (Intercept) f2 f4 f6 f9 f10
## -2.604446556 -0.047059251 0.135638594 -0.062440901 0.106193447 -0.005877698coef(logit.cv.lasso.z.interactions,
s=logit.cv.lasso.z.interactions$lambda.min)[which(coef(logit.cv.lasso.z.interactions,
s=logit.cv.lasso.z.interactions$lambda.min) !=0),]## (Intercept) f1 f2 f6 f8 f9
## -3.60777825 -0.22454262 -0.41308015 -0.08267585 -0.02384136 0.07569372
## f10 f11
## -0.02682158 1.34202229