7-isolation-forests.Rmd

---
title: "Isolation forest"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```

## Load data

```{r load_data}
# From 1-clean-data.Rmd
data = rio::import("data/clean-data-imputed.RData")

str(data)
```

## Basic isolation forest

```{r basic_isolation}
library(h2o)

# Start up h2o.
h2o.init()

h2o_df = as.h2o(data)

# Split dataset giving the training dataset 75% of the data
h2o_split <- h2o.splitFrame(data = h2o_df, ratios=0.75)

# Create a training set from the 1st dataset in the split
train <- h2o_split[[1]]

# Create a testing set from the 2nd dataset in the split
test <- h2o_split[[2]]

# Build an Isolation forest model
model <- h2o.isolationForest(training_frame=train,
                             sample_rate = 0.1,
                             max_depth = 20,
                             ntrees = 50)

# Calculate score
score <- h2o.predict(model, test)
result_pred <- as.data.frame(score$predict)

summary(result_pred)

library(ggplot2)
qplot(result_pred$predict) + theme_minimal()

# Predict the leaf node assignment
ln_pred <- h2o.predict_leaf_node_assignment(model, test)
```

## Resources

* [h2o blog post on isolation forests](https://www.h2o.ai/blog/anomaly-detection-with-isolation-forests-using-h2o/) - Nov. 2018
* [h2o documentation for isolation forests](http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/if.html)

## References