Tensor Flow con R

Tutorial Edgar Ruíz - México 2019

Etapas de análisis

Bibliotecas para Data Science

Material

• Turorial página web RStudio - Edgar Ruíz
• Github Edgar Ruíz - Tensor Flow con R
• Datos de ejemplo: Predicción de pérdida de clientes

Instalación de herramientas

install.packages("tensorflow")
install.packages("keras")

library(keras)
library(tensorflow)

install_keras()
install_tensorflow()

Lectura de datos

Preprocesamiento de datos

Train y Test

library(rsample)
id_partition <- initial_split(data = datos, prop = 0.3)
df_train <- training(id_partition)
df_test <- testing(id_partition)

“Receta”

library(recipes)
depuration <- df_train %>% 
  recipe(Churn ~ .) %>%                                 # Forma del modelo
  step_rm(customerID) %>%                               # Remueve columnas
  step_naomit(all_outcomes(), all_predictors()) %>%     # Omitir NAs
  step_discretize(tenure, options = list(cuts = 6)) %>% # Discretiza variables 
  step_log(TotalCharges) %>%                            # Conversión logarítmica
  step_mutate(Churn = ifelse(Churn == "Yes", 1, 0)) %>% # Conversión binaria
  step_dummy(all_nominal(), -all_outcomes()) %>%        # Variables dummy
  step_center(all_predictors(), -all_outcomes()) %>%    # Centrar valores
  step_scale(all_predictors(), -all_outcomes()) %>%     # Centrado de variables
  prep()

Aplicación de “Receta”

Train

# Predictores 
x_df_train <- depuration %>% 
  juice(all_predictors(),
        composition = "matrix")

# Respuesta
y_df_train <- depuration %>% 
  juice(all_outcomes()) %>% 
  pull()

Test

# Preprocesado base
base_test <- bake(object = depuration, new_data = df_test)

# Predictores
x_df_test <- base_test %>% 
  select(-Churn) %>% 
  as.matrix()

# Respuesta
y_df_test <- base_test %>% 
  select(Churn) %>% 
  pull()

Entrenamiento de modelos

Preparación con keras

• Guías para complementar la preparación del modelo con keras:
  • Modelo secuencial keras
  • Machine Learning con caret y keras - Edimer

modelo_keras <- keras_model_sequential() %>% 
  layer_dense(units = 16,
              kernel_initializer = "uniform",
              activation = "relu",
              input_shape = ncol(x_df_train)) %>% 
  layer_dropout(rate = 0.1) %>% 
  layer_dense(units = 16,
              kernel_initializer = "uniform",
              activation = "relu") %>% 
  layer_dropout(rate = 0.1) %>% 
  layer_dense(units = 1,
              kernel_initializer = "uniform",
              activation = "sigmoid") %>% 
  compile(optimizer = "adam",
          loss = "binary_crossentropy",
          metrics = c("accuracy"))

Ajuste del modelo

resultado <- fit(
  object = modelo_keras,
  x = x_df_train,
  y = y_df_train,
  batch_size = 50,
  epochs = 35,
  validation_split = 0.30,
  verbose = 0
)

plot(resultado) +
  theme_bw()

Predicciones

# Predicción de clases 
rta_class <- modelo_keras %>% 
  predict_classes(x_df_test) %>% 
  as.factor() %>% 
  fct_recode(yes = "1", no = "0")

# Predicción de probabilidades
rta_probs <- modelo_keras %>% 
  predict_proba(x_df_test) %>% 
  as.vector()

# Valores reales en test
real_test <- y_df_test %>% 
  as.factor() %>% 
  fct_recode(yes = "1", no = "0")

# Juntando predichos y reales
df_real_predict <- tibble(
  real = real_test,
  predicho = rta_class,
  predicho_prob = rta_probs
)
df_real_predict

Validación del modelo

Métricas de desempeño

library(yardstick)

• Matriz de confusión:

conf_mat(data = df_real_predict, truth = real, estimate = predicho)

          Truth
Prediction   no  yes
       no  3230  612
       yes  398  684

• Accuracy y Kappa:

metrics(data = df_real_predict, truth = real, estimate = predicho)

• Área bajo la curva (ROC):

roc_auc(data = df_real_predict, real, predicho_prob)

• Precisión:

precision(data = df_real_predict, truth = real, estimate = predicho)

• Sensitividad:

recall(data = df_real_predict, truth = real, estimate = predicho)

• Medidas F:

f_meas(data = df_real_predict, truth = real, estimate = predicho, beta = 1)

Entendiendo el modelo black box

library(lime)
suppressWarnings(explicar <- x_df_train %>% 
  as_tibble() %>%  
  lime(modelo_keras, bin_continuous = FALSE))

explanation <- x_df_test %>% 
  as_tibble() %>%
  lime::explain(
    explainer = explicar, n_labels = 1,
    n_features = 2, kernel_width = 0.5
  )

Recursos

• Aprendizaje automático con Tensorflow y R
• Deep Learning with Keras & Lime in R