In gloverd2/admr: Contains standard functions for Admiral Advance Analytics

Introduction

knitr::opts_chunk$set(echo = TRUE, include =TRUE)

.libPaths('C:/Program Files/R/R-3.6.3/library')
library(ADMR)
library(tidyverse)
library(checkmate) 
library(magrittr)
library(data.table)
library(testthat)

sessionInfo()

What functions are included in the package?

Currently the master branch contains the following functions however we have many others as pull request

ls("package:ADMR")

Using the Package

Import 'Predicting Household Sale Prices' dataset

Importing the packages inbuilt testing datasets

data(hh_train)
data(hh_test) 

This data is the Californian house prices dataset and has a target of r "SalePrice"

target <- 'SalePrice' 
hh_train %<>% select(all_of(c(intersect(names(hh_train), names(hh_test)), target))) 
hh_test %<>% select(intersect(names(hh_train), names(hh_test)))
names(hh_train)

Data Exploration

expl_na

For each variable, find the number or percentage of NA's in a dataset

na_vars <- ADMR::expl_na(df = hh_train, na.strings = NULL, ignore.case = FALSE) 

We can now see that we have some factors which are useless as they are mostly unknown

head(dplyr::arrange(na_vars, -na))

expl_categorical

For each categorical variable, find the number of / percentage of each level

ADMR::expl_categorical(df = hh_train, char.level = 3, num.level =2)

This format can be easily plotted to show factor distributions for each factor (either in a loop or using ggplot2 facet functions)

encode_freq

Apply frequency encoding with the option to group the rarest levels

# Add some unknowns to show what it looks like
hh_train$Neighborhood[5:10] <- NA

From the below output we can see that the factor Neighborhood has more levels than can usefully be modelled

table(hh_train$Neighborhood, exclude = NULL)

We decide to frequency encode the factor

nhood_freq <- ADMR::encode_freq(data = hh_train$Neighborhood, 
                                n_levels = 5, 
                                min_level_count = NULL, 
                                unknown_levels = NULL, 
                                unknown_treatment_method = 1)

which levels us with the transformed new factor

hh_train$nhood_freq <- factor(x = nhood_freq$data, labels = nhood_freq$levels)
table(hh_train$nhood_freq)

Build a Model

We build a model using standard R functions

hh_train %<>% mutate(GarageYrBlt = coalesce(GarageYrBlt, YearBuilt))
model_vars <- c('GarageYrBlt', 'nhood_freq', 'OverallCond', 'ExterCond', target)
hh_train %<>% select(all_of(model_vars))

model <- stats::glm(formula = 'SalePrice ~ .',
                   family = 'gaussian',
                   data = hh_train)
summary(model)
hh_train$pred <- predict(object = model, newdata = hh_train)

Model Evaluation

plot_PDP

We can now plot PDP plots for any factor

ADMR::plot_PDP(data = hh_train,
               model = model,
               explain_col = 'GarageYrBlt',
               n_bins = 20)

plot_PDP for 2 factors

We can also plot pairs of factors

ADMR::plot_PDP(data = hh_train,
               model = model,
               explain_col = c('GarageYrBlt', 'OverallCond'),
               n_bins = 10)

Note: OverallCond: Rates the overall condition of the house 10 - Very Excellent 1 - Very Poor

plot_ALE

We can now plot ALE plots for any factor

ADMR::plot_ALE(data = hh_train,
               model = model,
               explain_col = 'GarageYrBlt',
               n_bins = 20)

plot_lift_curve

We can also plot useful global model outputs such as lift curves

ADMR::plot_lift_curve(actual = hh_train$SalePrice,
                      predicted = hh_train$pred)

plotly vs ggplot2

Some users may require plotly and others ggplot2 so all plots work with both engines

ADMR::plot_lift_curve(actual = hh_train$SalePrice,
                      predicted = hh_train$pred,
                      use_plotly = FALSE)

Metrics

We can use codebase to compute metrics

library(Metrics)

ADMR::metric_mae(actual = hh_train$SalePrice, predicted = hh_train$pred, na.rm = TRUE)
Metrics::mae(actual = hh_train$SalePrice, predicted = hh_train$pred)

we have many metrics all of which have the same construction of arguments meaning they can be used by other functions or in loops

ADMR::metric_rmse(actual = hh_train$SalePrice, predicted = hh_train$pred)
ADMR::metric_mae(actual = hh_train$SalePrice, predicted = hh_train$pred)

ADMR::metric_deviance(actual = hh_train$SalePrice, predicted = hh_train$pred, family = "gaussian")
ADMR::metric_nloglik(actual = hh_train$SalePrice, predicted = hh_train$pred, family = "gaussian")

gloverd2/admr documentation built on Dec. 2, 2020, 11:16 p.m.