build_pipeline.Rmd

---
title: "Build a pipeline"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{build_pipeline}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  rmarkdown.html_vignette.check_title = FALSE
)
```

## Load libraries 

```{r eval = FALSE}
library(autotextclassifier) # Auto text classifier 
library(parallel) # Parallel processing 
library(doParallel) # Parallel processing 
library(here) # Creating reproducible file paths 
library(patchwork) # Putting ggplots together
library(recipes) # Preprocessing 
library(zeallot) # Multiple assignments 
library(yardstick) # Metrics  
```

## Import data 

```{r eval = FALSE}
load(file = here("inst/extdata/sample_data.rda"))
```

## Data munging 

Don't forget to make sure that the type of the outcome variable should be factor.

```{r eval = FALSE}
names(sample_data) <- c("category", "org_name", "ein", "text")
sample_data$category <- as.factor(sample_data$category)
```

## Apply basic recipe 

The `rec` object provides the following information. The function also checks whether the `text` column has missing values or includes extremely short documents (less than five words).

There are two broad basic options for text preprocessing. 

1. Without word embedding 

* Tokenization for text [trained]
* Stop word removal for text [trained]
* Text filtering for text [trained]
* Term frequency-inverse document frequency with text [trained]

2. With word embedding 

* Tokenization for text [trained]
* Stop word removal for text [trained]
* Text filtering for text [trained]
* Word embeddings aggregated from text [trained]

```{r eval = FALSE}
# Without word embedding
rec <- apply_basic_recipe(sample_data, category ~ text, text)

# With word embedding 
rec_alt <- apply_basic_recipe(sample_data, category ~ text, text, add_embedding = TRUE)
```

## Build a pipeline 

The `build_pipeline` function reduces the steps one needs to take a classifier pipeline. The pipeline involves data splitting, creating tuning parameters, search spaces, workflows, 10-fold cross-validation samples, finding the best model from each algorithm and fitting the best model from each algorithm to the data. 

```{r eval = FALSE}
# Using parallel processing to speed up
all_cores <- parallel::detectCores(logical = FALSE)

cl <- makeCluster(all_cores[1] - 1)

registerDoParallel(cl)
```

```{r eval = FALSE}
set.seed(1234)

c(lasso_fit, rand_fit, xg_fit) %<-% build_pipeline(rec, category, rec, prop_ratio = 0.8, metric_choice = "roc_auc")
```

## Evaluate the model using visualization

```{r eval = FALSE}
# Based on the class-based metrics 

viz_class_fit(lasso_fit, "Lasso", test_x_class, test_y_class, "class") + 

viz_class_fit(rand_fit, "Random forest", test_x_class, test_y_class, "class") +

viz_class_fit(xg_fit, "XGBoost", test_x_class, test_y_class, "class")

# Based on the probability-based metrics 

viz_class_fit(lasso_fit, "Lasso", test_x_class, test_y_class, "probability") +

viz_class_fit(rand_fit, "Random forest", test_x_class, test_y_class, "probability") +

viz_class_fit(xg_fit, "XGBoost", test_x_class, test_y_class, "probability")
```