Using Machine Learning to Improve UX

PredictionIO and Drupal

@aczietlow

@ccrampton

Mindgrub

Technical Agency, that also happens to specialize in Drupal

Now Hiring

How'd we get here?

1997 - IBM Deep Blue defeats Garry Kasparov at Chess

2011 - IBM's Watson crushes it on Jeopardy

2012 – Google’s X Lab builds ML algorithm to identify cat videos

2014 - Sergey Bin announces Google is working on self driving cars

Machine Learning

Machine Learning

What it isn't

• Magic
• AI
• A replacement for anyone's job
• Skynet ... yet

Machine Learning

What it is

Machine Learning

Learns

Supervised

All data is labeled, and algorithms learn to predict outcomes.

Unsupervised

Data is not labeled and the algorithm learns to structure the data.

semi-supervised

Some of the data is labeled, but most is not.

Supervised

UnSupervised

Machine Learning

PredictionIO

• Open source (apache-2.0)
• Has library of algorithms (engines) available
• Easy to wire to web apps
• Has an awesome animal mascot

PredictionIO

• Communicates with your web apps via Rest
• Event Server
• Data Storage
• Machine learning Engines

PredictionIO: Event Server

• Apache HBase, MySQL, PostgreSQL
• Passes data on to ML Engines

PredictionIO: Engine

• The 'brain' of the stack
• Contains 1 (or more) algorithms to build training models
• Can be queried to make predictions
• More data === better

Demo #1

Product Recommendations

Initial setup

• Have PredictionIO & dependencies installed
• We have a self contained Docker image
• Works beautifully with Docksal
• Install desired engines
• PredictionIO documentation is your friend

Build Pio

pio build --verbose

	[INFO] [Engine$] Using command '/PredictionIO-0.12.0-incubating/sbt/sbt' at /engines/similar-text to build.
	[INFO] [Engine$] If the path above is incorrect, this process will fail.
	[INFO] [Engine$] Uber JAR disabled. Making sure lib/pio-assembly-0.12.0-incubating.jar is absent.
	[INFO] [Engine$] Going to run: /PredictionIO-0.12.0-incubating/sbt/sbt  package assemblyPackageDependency in /engines/similar-text
	[INFO] [Engine$] [info] Loading project definition from /engines/similar-text/project
	[INFO] [Engine$] [info] Set current project to pio-template-text-similarity (in build file:/engines/similar-text/)
	[INFO] [Engine$] [success] Total time: 2 s, completed Aug 18, 2018 3:30:58 PM
	[INFO] [Engine$] [info] Including from cache: scala-library-2.11.8.jar
	[INFO] [Engine$] [info] Checking every *.class/*.jar file's SHA-1.
	[INFO] [Engine$] [info] Merging files...
	[INFO] [Engine$] [warn] Merging 'META-INF/MANIFEST.MF' with strategy 'discard'
	[INFO] [Engine$] [warn] Strategy 'discard' was applied to a file
	[INFO] [Engine$] [info] Assembly up to date: /engines/similar-text/target/scala-2.11/pio-template-text-similarity-assembly-0.1-SNAPSHOT-deps.jar
	[INFO] [Engine$] [success] Total time: 12 s, completed Aug 18, 2018 3:31:09 PM
	[INFO] [Engine$] Compilation finished successfully.
	[INFO] [Engine$] Looking for an engine...
	[INFO] [Engine$] Found pio-template-text-similarity_2.11-0.1-SNAPSHOT.jar
	[INFO] [Engine$] Found pio-template-text-similarity-assembly-0.1-SNAPSHOT-deps.jar
	[INFO] [Engine$] Build finished successfully.
	[INFO] [Pio$] Your engine is ready for training.
						

Pio train

pio train

	[INFO] [CoreWorkflow$] Inserting persistent model
	[INFO] [CoreWorkflow$] Updating engine instance
	[INFO] [CoreWorkflow$] Training completed successfully.
						

Demo

For when the live demo doesn't work

	{
	  "itemScores":[
		{"item":"i0","score":0.7071067811865475},
		{"item":"i1","score":0.7071067811865475},
		{"item":"i2","score":0.5773502691896258},
		{"item":"i3","score":0.5773502691896258}
	  ]
	}
						

Prediction Engine

You can build your own

  def sumArray (m: Array[Double], n: Array[Double]): Array[Double] = {
    for (i <- 0 until m.length) {m(i) += n(i)}
    return m
  }

  def divArray (m: Array[Double], divisor: Double) : Array[Double] = {
    for (i <- 0 until m.length) {m(i) /= divisor}
    return m
  }

  def wordToVector (w:String, m: Word2VecModel, s: Int): Vector = {
    try {
      return m.transform(w)
    } catch {
      case e: Exception => return Vectors.zeros(s)
    }
  }

  def normalizet(line: String) = java.text.Normalizer.normalize(line,java.text.Normalizer.Form.NFKD).replaceAll("\\p{InCombiningDiacriticalMarks}+","").toLowerCase

					

Prediction Engine

Proudly invented else where

To the library!

• Recommendations
• The Universal Recommender - Cross-Occurrence (CCO) algorithm
• Recommendation - Collaborative Filtering
• E-Commerce Recommendation - MLLib ALS E-commerce algorithm fork
• E-Commerce Recommendation (Java) - Like before but now with more Java!
• Similar Product - recommends based on user's actions
• Product Ranking
• Music Recommendations
• Complementary Purchase
• Viewed This Bought That
• Frequent Pattern Mining - FP Growth algorithm
• Similar Product with Rating

• Classification
• Classification - Naive Bayes algorithm
• Lead Scoring - likelihood user will convert
• Text Classification - uses OpenNLP library for text vectorization
• Churn Prediction - Attrition rate users across cell phone carriers
• Classification Deeplearning4j - Uses Deeplearning4j library
• Document Classification
• Circuit End Use Classification - predicts circuit energy consumption
• GBRT_Classification
• classifier-kafka-streaming-template
• sentiment analysis
• Classification template for Iris

• Regression
• Survival Regression - Profiling customers who has a higher survival rate
• Deep Learning Energy Forecasting
• Electric Load Forecasting
• Linear Regression BFGS
• Regression template for Boston House Prices

• Natural Language Processing
• Text similarity engine
• Topic Labelling with Wikipedia - auto tagging
• Text Classification
• OpenNLP Sentiment Analysis Template
• Recursive Neural Networks (Sentiment Analysis)

• Clustering
• Topc Model (LDA)
• KMeans-Clustering-Template - KMeans Algorithm

• Similarity
• Content Based SVD Item Similarity Engine
• Similar Product with Rating

Stay tuned

Releasing Official PredictionIO Docker Images

Drupal Module(s) (?) maybe

Resources

• Slides
• PredictionIO Dockerfile
• PredictionIO Image