---

title: "Regression Example: Boston Housing"

author: "Jo-fai (Joe) Chow - [email protected]"

date: "H2O + LIME Workshop at eRum 2018 (Updated for MilanoR Workshop)"

output:

html_document:

df_print: kable

fig_height: 10

fig_width: 14

highlight: tango

number_sections: yes

theme: spacelab

toc: yes

toc_depth: 2

---

<br>

```{r setup, include=FALSE}

knitr::opts_chunk$set(echo = TRUE)

```

# Get Ready

```{r, message=FALSE}

# Libraries

library(h2o) # for H2O Machine Learning

library(lime) # for Machine Learning Interpretation

library(mlbench) # for Datasets

```

```{r}

# Your lucky seed here ...

n_seed = 12345

```

<br>

# Data Prep - Boston Housing

```{r}

data("BostonHousing")

dim(BostonHousing)

head(BostonHousing)

```

<br>

## Define Target and Features

```{r}

target = "medv" # Median House Value

features = setdiff(colnames(BostonHousing), target)

print(features)

```

<br>

## Convert R dataframe into H2O dataframe (JVM)

```{r}

# Start a local H2O cluster (JVM)

h2o.init()

h2o.no_progress() # disable progress bar for RMarkdown

h2o.removeAll() # Optional: remove anything from previous session

```

```{r}

# H2O dataframe

h_boston = as.h2o(BostonHousing)

head(BostonHousing)

```

```{r}

# Split Train/Test

h_split = h2o.splitFrame(h_boston, ratios = 0.75, seed = n_seed)

h_train = h_split[[1]] # 75% for modelling

h_test = h_split[[2]] # 25% for evaluation

```

<br>

# Build H2O Models

<br>

## Single Model - Default H2O GBM

```{r}

# Train a Default H2O GBM model

model_gbm = h2o.gbm(x = features,

y = target,

training_frame = h_train,

model_id = "gbm_default_reg",

seed = n_seed)

print(model_gbm)

```

<br>

## Evaluate Single Model

```{r}

# Evaluate performance on test

h2o.performance(model_gbm, newdata = h_test)

```

<br>

## H2O AutoML: Multiple H2O Models + Stacked Ensemble

```{r}

# Train multiple H2O models with H2O AutoML

# Stacked Ensembles will be created from those H2O models

# You tell H2O ...

# 1) how much time you have and/or

# 2) how many models do you want

# Note: H2O deep learning algo on multi-core is stochastic

model_automl = h2o.automl(x = features,

y = target,

training_frame = h_train,

nfolds = 5, # Cross-Validation

max_runtime_secs = 120, # Max time

max_models = 100, # Max no. of models

stopping_metric = "RMSE", # Metric to optimize

project_name = "automl_reg",

exclude_algos = NULL, # If you want to exclude any algo

seed = n_seed)

```

<br>

## AutoML Leaderboard

```{r}

model_automl@leaderboard

```

<br>

## Best Model (Single / Stacked Ensemble)

```{r}

# H2O: Model Leader

# Best Model (either an individual model or a stacked ensemble)

model_automl@leader

```

<br>

## Evaluate Performance

```{r}

# Default GBM Model

h2o.performance(model_gbm, newdata = h_test)

```

```{r}

# Best model from AutoML

h2o.performance(model_automl@leader, newdata = h_test) # lower RMSE = better

```

<br>

## Make Predictions (Optional)

```{r}

yhat_test = h2o.predict(model_automl@leader, h_test)

head(yhat_test)

```

<br>

## Export Models (Optional)

- Use `h2o.saveModel()` to save model to disk

- Use `h2o.loadModel()` to re-load model

- Also see `h2o.download_mojo()` and `h2o.download_pojo()`

```{r, eval=FALSE}

# Save model to disk

h2o.saveModel(object = model_automl@leader,

path = "./models/",

force = TRUE)

```

<br>

# Explain the Model

<br>

## Step 1: Create an `explainer`

```{r}

explainer = lime::lime(x = as.data.frame(h_train[, features]),

model = model_automl@leader)

```

<br>

## Step 2: Turn `explainer` into `explanations`

```{r}

# Extract one sample (change `1` to any row you want)

d_samp = as.data.frame(h_test[1, features])

```

```{r}

# Assign a specifc row name (for better visualization)

row.names(d_samp) = "Sample 1"

```

```{r}

# Create explanations

explanations = lime::explain(x = d_samp,

explainer = explainer,

n_permutations = 5000,

feature_select = "auto",

n_features = 13) # Look top x features

```

<br>

## Look at Explanations (Bar Chart)

```{r}

lime::plot_features(explanations, ncol = 1)

```

<br>

## Look at Explanations (Full Table)

```{r}

# Sort explanations by feature weight

explanations =

explanations[order(explanations$feature_weight, decreasing = TRUE),]

```

```{r}

# Print Table

print(explanations)

```

<br>

# Try it Yourself

Replace `BostonHousing` with your own data. Good luck!

<br>

# Session Info

```{r}

sessionInfo()

```

<br>