Data Science

How to use R, H2O, and Domino for a Kaggle competition

Nick Elprin2014-09-20 | 17 min read

Return to blog home

This is a guest post by Jo-Fai Chow

If you're in a hurry, feel free to skip to:

  • Tutorial 1: Using Domino
  • Tutorial 2: Using H2O to Predict Soil Properties
  • Tutorial 3: Scaling up your analysis

Introduction

This blog post is the sequel to TTTAR1 a.k.a. An Introduction to H2O Deep Learning. If the previous blog post was a brief intro, this post is a proper machine learning case study based on a recent Kaggle competition: I am leveraging R, H2O and Domino to compete (and do pretty well) in a real-world data mining contest.

kaggle_topspot

The competition got tougher since then but I am still very proud to say I stayed top for a few days!

Note that the public score is only based on a small subset (13%) of test data. My position on the final leaderboard might change dramatically. In my opinion, climbing up the public leaderboard is just the first step, making sure the models are generalized enough to stay up is the real challenge. Anyway, that is a topic for a future blog post.

Accelerate Your Data Analysis with Domino

I think it is fair to say that Kaggle contests are very challenging. Even more so when you put the time constraint into the mix. Domino's data science tools let me easily scale up my analysis, take advantage of parallelized code and carry out many tests within a very short space of time. Competing in a data mining contest is all about understanding the problem better and identifying the best strategies sooner than your opponents. Having the ability to ask the data many questions at the same time is crucial.

As a result, Domino cut down days or even weeks of analysis to just a few hours. I could quickly evaluate many different machine learning strategies for the Kaggle contest. On top of that, I didn't have to waste any time on infrastructure setup and hardware maintenance, so I could focus on understanding the machine learning problem better and improving the model performance steadily over time.

Go Deeper with H2O

To say that I am very impressed with the predictive performance of the H2O deep learning module is an understatement. All data geeks know about the "No Free Lunch" theorem. Yet, against all odds and my past experience, I have managed to climb up the ladder using just the h2o.deeplearning algorithm and some simple model averaging strategies.

Performance is not the only thing, H2O's highly efficient memory management and parallelized algorithms are just as amazing. It got even better when I realized that I can use R, H2O and Domino together. At that moment, I knew I had found a very promising machine learning solution and hence the motivation of this blog post.

This post is not only about Kaggle competitions, it is more about finding a practical and scalable data analysis solution. A generic and robust solution that I can rely on and apply to many of my machine learning problems in life (like this one). My experience with this R + H2O + Domino combination in the last couple weeks has been very promising. So I would like to use this opportunity to share my experience. I hope this post encourages more data geeks out there to give R, H2O and Domino a try.

Using H2O alone is a huge topic. In order to keep things simple, I will just walk you through the starter code here. I promise you that there will be a more in-depth discussion on various machine learning algorithms and strategies after the contest. I also hope that these blog posts can eventually evolve into proper tutorials (as my tiny contribution to the data science community).

Learning-by-Doing

I always believe that learning-by-doing is the best way to learn. So I have prepared two short tutorials with R starter code below to get you started with Domino and H2O. Hopefully, by going through the tutorials, you will be able to master the basics and to try H2O and Domino for your own data analysis very soon. Moreover, if you follow all the steps, you will get a CSV file that is ready for the Kaggle competition. Finally, when you are ready to up your game, check out the third tutorial about scaling up your analysis in just a few easy steps.

Wait ... H2O, Domino, Kaggle? Who Are They?

Just in case you haven't heard of them yet ...

  • H2O is an open-source math and in-memory prediction engine for big data science developed by 0xdata.
  • Domino is an enterprise-grade platform for analysts and data scientists that handles all the “plumbing” required to run, scale, share, and deploy analytical models.
  • Kaggle is a platform for predictive modeling and analytics competitions and consulting.

Ready?

OK, that's quite enough for an intro. Let's dive into it!!!

Tutorial One: Using Domino

In the first tutorial, I will only focus on the basic Domino operations via their web UI and R interface. Although the R script is all about H2O machine learning, it is hard to learn two new things at the same time. So I will keep this simple and leave the H2O discussion for the next tutorial.

Step 1.1 - Install R Interface

Domino has its own R package on CRAN. After installing the package, you can log into your Domino account and manage your projects directly from R. The Domino team has done a great job streamlining the whole process. Personally, I think this is yet another excellent example of using R as an interface.

install.packages("domino")
library(domino)
domino.login("your_username", "your_password")

Step 1.2 - Download Data and Code

This tutorial is based on the Kaggle Africa Soil Property Prediction Challenge. In order to carry out the data analysis, you will need to download the original datasets from Kaggle first. Then you can run a simple analysis using my sample R script, Kaggle_AfSIS_with_H2O.R.

kaggle_data

Step 1.3 - Upload Data and Code

You will need to upload the files to a specific project folder on Domino. To keep things simple, I will use the quick-start folder here because most of you should have this folder somewhere on your local drive after going through Domino's getting started guide.

On your machine, copy the Kaggle datasets and R script to the quick-start folder. The file structure should look like this:

Before

 

After

  • results /
  • main.m
  • main.py
  • main.r
  •  

  • data / train.zip (note: create a new 'data' folder)
  • data / test.zip
  • data / sample_submission.csv
  • results /
  • Kaggle_AfSIS_with_H2O.R
  • main.m
  • main.py
  • main.r
  • After that, you can upload the files to your quick-start folder on Domino.

    ## Set Working Directory to the 'quick-start' folder
    setwd("quick_start_dir_on_your_local_drive")## Upload new files to the 'quick-start' folder on Domino clouddomino.upload()

    Step 1.4 - Check the Files

    Before running the analysis, check and make sure the files are on the cloud. Log into your Domino account and browse your files in your quick-start project.

    kaggle_files

    Step 1.5 - Run your Code

    Got all the files there? Great, you now are ready to run the analysis on the cloud!

    Use the web UI to start a run (Runs -> Run -> Enter of the File Name -> Start Run).

    kaggle_startrun

    Alternatively, you can do this directly from R.

    domino.run("Kaggle_AfSIS_with_H2O.R")

    Step 1.6 - Monitor Progress

    Domino's web UI lets you monitor your runs online. You can get access to all your files and project settings on the left panel. When you click on any of your runs, it will display the native console (R in this case) on the right with a summary below showing CPU load, memory usage, and other information about the run.

    kaggle_monitor

    Step 1.7 - Download the Results

    The analysis takes about half an hour. After that, you can download the results my_Kaggle_submission.csv either from the web ...

    kaggle_monitor

    or via R.

    ## Download the results to your local drive
    domino.download()

    By default, you will receive an email notification once the run has finished. There are more ways to customize notifications if you are interested.

    kaggle_monitor

    Step 1.8 - Submit it to Kaggle (Optional)

    You can now go to the submission page and upload the CSV file. Good luck :)

    Try Your Own Code

    I have shown you the basic steps to get my Kaggle starter code up and running on the Domino cloud. Now it is time to try your own code. Remember Domino also supports Python, Matlab and Julia!

    Share Your Projects

    It is very easy to share your Domino projects with others. You can enable public access to your project folder in the project settings so everyone can view your files and runs.

    kaggle_monitor

    Version Control

    As an aside: the version control on Domino is a well-thought process specially designed for data scientists. Instead of versioning individual files one at a time, Domino creates snapshots of your data, code and results altogether. That means it is very easy to go back in time and make reference to a particular run with the exact data and code used.

    Tutorial Two: Using H2O to Predict Soil Properties

    In this section, I will explain how I use R and H2O to train predictive models for the Kaggle competition. In short, we have to develop regression models based on 1158 rows of training data and then use the models to predict five soil properties Ca, P, pH, SOC and Sand for each of the 728 test records. For more information, please go to the description page on Kaggle.

    Step 2.1 - Install R Interface for H2O

    Again, R acts as an interface. Having R as the common interface,
    H2O and Domino integrate seamlessly.

    Like Domino, H2O also has an excellent R package. To begin with, you can install the CRAN Version of H2O:

    install.packages("h2o")

    If you want to take advantage of the new H2O features, you can install the bleeding edge version.

    ## Specify H2O version here
    h2o_ver <- "1511"## Install H2Olocal({r <- getOption("repos"); r["CRAN"] <- "https://cran.us.r-project.org"; options(repos = r)})txt_repo <- (c(paste0(paste0("https://s3.amazonaws.com/h2o-release/h2o/master/", h2o_ver),"/R"), getOption("repos")))install.packages("h2o", repos = txt_repo, quiet = TRUE)

    In order to find out the latest H2O version number, go to this page and look for the last four digits as shown in the screenshot below.

    h2o_version

    (Optional) I use H2O a lot these days so I wrote a wrapper function to make my life easier. With two lines of code, you can install or upgrade H2O to the latest bleeding-edge version:

    ## Use the wrapper function in my package 'deepr'
    devtools::install_github("woobe/deepr")
    deepr::install_h2o()

    Step 2.2 - Initiate and Connect to a Local H2O Cluster

    Using the argument max_mem_size in function h2o.init(...), we can manually limit the physical memory usage. In this tutorial, we set a limit of 1GB of memory as it is the maximum allowance of the Domino free tier.

    library(h2o)
    localH2O <- h2o.init(max_mem_size = '1g')
    R connected to H2O Cluster

    ### Step 2.3 - Import Data

    Importing data is one of the things I like most about H2O!!! You can load compressed files (.zip, .gz) directly into the H2O cluster. This may not be significantly important for the small Kaggle dataset in this tutorial but it is certainly useful for bigger datasets.

    ## Import Data to H2O Cluster
    train_hex <- h2o.importFile(localH2O, "./data/train.zip")
    test_hex <- h2o.importFile(localH2O, "./data/test.zip")

    Step 2.4 - Train Deep Neural Networks Models and Make Predictions

    Without further ado, here is the starter code to train one deep neural networks model for each target variable and then use the models to make predictions. There are quite a few variables to tweak in the h2o.deeplearning(...) function. To keep things simple, I only provide the barebone code to get you started. For more information, I strongly recommend going through the vignette and demo code.

    ## Split the dataset into 80:20 for training and validation
    train_hex_split &amp;lt;- h2o.splitFrame(train_hex, ratios = 0.8, shuffle = TRUE)
    ## Split the dataset into 80:20 for training and validationtrain_hex_split <- h2o.splitFrame(train_hex, ratios = 0.8, shuffle = TRUE)## One Variable at at Timels_label <- c("Ca", "P", "pH", "SOC", "Sand")for (n_label in 1:5) {## Displaycat("nnNow training a DNN model for", ls_label[n_label], "...n")## Train a 50-node, three-hidden-layer Deep Neural Networks for 100 epochsmodel <- h2o.deeplearning(x = 2:3595,y = (3595 + n_label),data = train_hex_split[[1]],validation = train_hex_split[[2]],activation = "Rectifier",hidden = c(50, 50, 50),epochs = 100,classification = FALSE,balance_classes = FALSE)## Print the Model Summaryprint(model)## Use the model for prediction and store the results in submission templateraw_sub[, (n_label + 1)] <- as.matrix(h2o.predict(model, test_hex))## One Variable at at Time
    ls_label <- c("Ca", "P", "pH", "SOC", "Sand")for (n_label in 1:5) {## Display
    cat("nnNow training a DNN model for", ls_label[n_label], "...n")## Train a 50-node, three-hidden-layer Deep Neural Networks for 100 epochs
    model <- h2o.deeplearning(x = 2:3595,
    y = (3595 + n_label),
    data = train_hex_split[[1]],
    validation = train_hex_split[[2]],
    activation = "Rectifier",
    hidden = c(50, 50, 50),
    epochs = 100,
    classification = FALSE,
    balance_classes = FALSE)## Print the Model Summary
    print(model)## Use the model for prediction and store the results in submission template
    raw_sub[, (n_label + 1)] <- as.matrix(h2o.predict(model, test_hex))}

    Step 2.5 - Save Results as a CSV

    The final step is to save all predictions as a CSV file. This is the CSV file in the correct Kaggle submission format.

    write.csv(raw_sub, file = "./results/my_Kaggle_submission.csv", row.names = FALSE)

    Try Other Datasets

    That's it! You have just trained some Deep Neural Networks models and used them for prediction! Why don't you try the same thing on other datasets? For example, the Breast Cancer and MNIST datasets I used in the previous TTTAR blog post.

    Try Other H2O Machine Learning Algorithms

    The h2o.deeplearning algorithm is just one of the algorithms available in H2O. You can also try h2o.gbm, h2o.glm, h2o.naiveBayes, h2o.randomForest and more. All these functions can be called directly from R like the code example above. They have similar syntax so it should be very straightforward.

    Tutorial Three: Scaling Up Your Analysis!

    I have walked you through a simple H2O machine learning exercise for Kaggle on the Domino cloud. I hope the starter code and tutorials can help you better understand the R + H2O + Domino mechanism. However, the simplicity of the model means you will not be able to achieve a very good score on the Kaggle leaderboard.

    When you are ready to gear up your effort (e.g. building more complex models) or analyze much bigger datasets, you can easily scale up your data analysis on Domino in a just few steps. Ready? Read on!

    Choose a Higher Hardware Tier

    You can find the hardware tier dropdown list in the project settings. All you need to do is to choose a tier that suits your needs. Domino will sort out the rest for you.

    Choosing hardware tier for H20 & R

    Modify Your Code

    If you use the R + H2O + Domino combination, you can easily improve H2O's performance by increasing the physical memory allowance. H2O will automatically parallelize the machine learning algorithms and utilize all CPU cores. For general code parallelization, read this Domino blog post about easy parallel loops.

    ## Increase the H2O Cluster Memory Allowance (e.g. 25GB)
    localH2O <- h2o.init(max_mem_size = '25g')

    Use Multiple Instances

    To further increase productivity, you can start multiple runs at the same time. Again, no complex set up is required. Just start your runs as usual. Domino will take care of the rest. In the screenshot below, you can see I would run multiple experiments simultaneously, each of which could take many hours and use 30GB of memory and 30 cores of CPU.

    Multiple instances of H20 & R

    Conclusions

    If I have to summarize this experience, I would say that the discovery of this R + H2O + Domino combination has been very exciting and promising! I can confidently say I have a practical and scalable solution to deal with not just Kaggle but many other machine learning problems. I also know this solution can get even better! I will continue to explore and experiment with other tools.

    All the Links

    This was quite a lengthy post so let me summarize all the useful resources here:

    Domino

    H2O

    It's the people, not just the tools!

    I have told you how great these tools are. But I haven't mentioned the awesome people behind the scenes.

    First of all, I would like to thank Nick Elprin for his tremendous support since the early days of Domino. Not only has Nick given me unlimited cloud resources in the last couple weeks to experiment with this R + H2O + Domino combination (it all began with a quick email "Hi Nick, can I use H2O on Domino?"), he also continuously helps me out with many technical issues and turns my requests into new features.

    Then, of course, is the Oxdata team (Sri Ambati, Arno Candel, Viraj Parmar, Cliff Click, Kevin Normoyle, Raymond Peck, Spencer Aiello and many others) behind the H2O engine. Since the early days when I began taking my baby steps to understand H2O, they have been emailing me tips and information about H2O as well as encouraging me to push the boundaries. Especially Arno Candel who gave me some very specific advice on fine-tuning h2o.deeplearning and other functions.

    I hope you find this post useful. If you decide to give this R + H2O + Domino combination a try but have questions regarding the setup, feel free to email / tweet me as well as the Domino and Oxdata team. I know your questions and feedback will help them evolve and shape even better products for data science. Unfortunately, I need to follow Kaggle competition rules and I can't tell you exactly what I do to climb up the leaderboard. Yet, you can pretty much guess it from the name of my blog "Blend It Like a Bayesian!".

    I afraid I can't cover everything in one post but I will continue to blog about other cool Domino and H2O features in the coming months.

    Nick Elprin is the CEO and co-founder of Domino Data Lab, provider of the open data science platform that powers model-driven enterprises such as Allstate, Bristol Myers Squibb, Dell and Lockheed Martin. Before starting Domino, Nick built tools for quantitative researchers at Bridgewater, one of the world's largest hedge funds. He has over a decade of experience working with data scientists at advanced enterprises. He holds a BA and MS in computer science from Harvard.

    Subscribe to the Domino Newsletter

    Receive data science tips and tutorials from leading Data Science leaders, right to your inbox.

    *

    By submitting this form you agree to receive communications from Domino related to products and services in accordance with Domino's privacy policy and may opt-out at anytime.