README.md

In AlexAfanasev/NeuralNetworkVisualization: Create Partial Dependence Plots for Neural Networks, Visualizes the marginal effects of Neural Networks

NeuralNetworkVisualization

Copyright 2019 Jacqueline Dorothea Seufert, Alex Afanasev. Licensed under the MIT license.

NeuralNetworkVisualization is a R Package for creating beautiful partial dependence plots for Neural Networks using ggplot2, plotly and shiny.

This software was developed by Jacqueline Dorothea Seufert and Alex Afanasev and is available freely.

Installation

NeuralNetworkVisualization is available through GitHub:

To install the latest development version from GitHub:

install.packages("devtools")
devtools::install_github("AlexAfanasev/NeuralNetworkVisualization")

Examples

Step by step procedure: 1. Fit a neural network using the NeuralNetwork class (uses the famous neuralnet package for fitting the model) 2. Create visualizations of the partial dependencies using the plot_partial_dependencies function (creates visualizations in ggplot2 and if wanted in plotly, adds a confidence interval)

Confidence intervals for the partial dependencies can be created directly using the plot_partial_dependencies function or by providing the options parameter while creating the NeuralNetwork model.

The following example models can be obtained using the example_nn_model function!

Numerical response variable:

library(NeuralNetworkVisualization)

# Example for Plotting with numerical dependent variable
library(MASS)
data <- Boston; data$chas <- as.factor(data$chas)
train <- data

set.seed(1)
system.time(
    model <- NeuralNetwork(medv ~ ., data = train, layers = c(5, 3),
                           scale = TRUE, linear.output = TRUE, threshold = 0.5,
                           options = list(store = TRUE, parallel = TRUE,
                                          probs = c(0.05, 0.95),
                                          nrepetitions = 1000)))

plot_partial_dependencies(model, use_stored_data = TRUE, type = "ggplotly")

plot_partial_dependencies(model, predictors = "crim", use_stored_data = TRUE)

plot_partial_dependencies(model, predictors = "crim")

plot_partial_dependencies(model, predictors = c("crim", "age"),
                          type = "ggplotly", probs = c(0.05, 0.95),
                          nrepetitions = 100, parallel = TRUE)

Categorical response variable:

library(NeuralNetworkVisualization)

# Example for Plotting with categorical dependent variable
library(datasets)
data("iris")
train_model <- iris

set.seed(1)
system.time(
    model <- NeuralNetwork(
        Species ~ ., data = train_model, layers = c(5, 5), rep = 5,
        linear.output = FALSE, scale = TRUE, err.fct = "ce", stepmax = 1000000,
        threshold = 0.5, options = list(
            store = TRUE, parallel = TRUE, nrepetitions = 1000,
            probs = c(0.05, 0.95))))

plot_partial_dependencies(model, type = "ggplotly", use_stored_data = TRUE)

plot_partial_dependencies(model, predictors = "Sepal.Length",
                          use_stored_data = TRUE)

plot_partial_dependencies(model, predictors = "Sepal.Length")

plot_partial_dependencies(model, predictors = c("Sepal.Length", "Petal.Length"),
                          type = "ggplotly", probs = c(0.1, 0.9),
                          nrepetitions = 100, parallel = TRUE)

Binary response variable:

library(NeuralNetworkVisualization)

# Example for Plotting with binary dependent variable
library(faraway)
pima$glucose[pima$glucose == 0] <- NA
pima$diastolic[pima$diastolic == 0] <- NA
pima$triceps[pima$triceps == 0] <- NA
pima$insulin[pima$insulin == 0] <- NA
pima$bmi[pima$bmi == 0] <- NA
pima <- pima[complete.cases(pima), ]
pima$test <- as.factor(pima$test)
levels(pima$test) <- c("Negative", "Positive")
train <- pima

set.seed(1)
system.time(
    model <- NeuralNetwork(
        test ~ pregnant + glucose + diastolic + triceps + insulin + bmi +
            diabetes + age, data = train, layers = 2, err.fct = "ce",
        linear.output = FALSE, threshold = 1.0, stepmax = 1e6, scale = TRUE,
        options = list(store = TRUE, parallel = TRUE, probs = c(0.05, 0.95),
                       nrepetitions = 1000)))

plot_partial_dependencies(model, use_stored_data = TRUE, type = "ggplotly")

plot_partial_dependencies(model, predictors = "glucose", use_stored_data = TRUE)

plot_partial_dependencies(model, predictors = "glucose")

plot_partial_dependencies(model, predictors = c("pregnant", "diastolic"),
                          type = "ggplotly", probs = c(0.05, 0.95),
                          nrepetitions = 100, parallel = TRUE)

Shiny App

Please run the shiny app within a modern browser. Dont be stupid trying to use Microsoft Internet Explorer or Edge.

Within the shiny app NeuralNetwork models can be uploaded as a .rds file, selected from the global environment or the available example models can be used.

Run the shiny app:

library(NeuralNetworkVisualization)
run_shiny_app()

Sources

• https://towardsdatascience.com/first-neural-network-for-beginners-explained-with-code-4cfd37e06eaf
• https://www.datacamp.com/community/tutorials/neural-network-models-r
• https://towardsdatascience.com/introducing-pdpbox-2aa820afd312
• https://christophm.github.io/interpretable-ml-book/pdp.html
• https://bgreenwell.github.io/pdp/articles/pdp.html

AlexAfanasev/NeuralNetworkVisualization documentation built on Sept. 23, 2019, 2:29 a.m.