1.
James Ward
Platform Evangelist
jamesward.com
@_JamesWard
Introduction to
Machine Learning

2.
Machine Learning • Deep Learning • Predictive Analytics • NLP • Smart Data Discovery
AI in the Salesforce Platform
Salesforce Einstein
Introducing
World’s smartest
CRM
Empowering Sales,
Service, Marketing & IT
Everyone can build
AI-powered apps fast

3.
How Humans Learn
Observe
world
Compare to
expectations
Analyze
differences
Refine
model

4.
How Machines Learn
Data to
Model
Evaluate
model
Feedback
signals
Refine
model

5.
What is Machine Learning?
Decision
if (a == “foo”) {
if (b == “bar”) {
if (c == “baz”) {
Data
Ask Model
Prediction Categorization
Apply Algorithm
Create Model
Data
Pattern Recognition via Zeros & Ones

6.
Machine Learning

7.
Machine Learning Spectrum
Use Cases
• Search Relevance
• Collaborative Filtering
• Decision Support
• Time Series Forecasting
Core Methods
• Compression
• Classification
• Regression
• Reinforcement
Frameworks
• Neural Networks
• Support Vector Machines
• Decision Trees
• Bayesian Models
• K-Nearest Neighbor
• Logistic Regression

8.
Use Cases
o Rate leads with how likely they are to close
o Categorize this text as offensive or not
o Recommend products based on what I’ve liked
o Detect anomalies in credit card purchase behavior

9.
ML Core Methods
• Compression – Determine the pattern
• Classification – Determine if “thing” is an x or y
• Regression – Determine the correct output for an input
• Reinforcement – Determine what action yields an award

10.
Product Recommendation
User Likes Prediction
Collaborative Filtering

11.
Demo!
https://dreamhouseapp.io/pio

12.
Prediction IO = Open Source Machine Learning Server
Heroku = Cloud platform to deliver, scale, and monitor apps built in any technology
Heroku Postgres = Relational database in the cloud, managed by Heroku
Heroku Connect = Data sync between Salesforce and Heroku Postgres
Apache Spark = Big Data Framework
Spark ML = Machine Learning libraries built on Apache Spark
Pieces of the Puzzle

13.
Product Recommendation
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
Human Intelligence First
★
★
★ ★
★

 

14.
Product Recommendation
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
Features
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔

15.
Product Recommendation
Deriving Features from Favorites
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
×
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔
Property Features User FeaturesFavorites
=
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
×
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔
Property FeaturesUser FeaturesFavorites
=
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔

16.
Product Recommendation
Deriving Recommendations from Features
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
×
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔
Property Features User Features Recommendations
=
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
★
Property 2 ✔ ✔
★ ★
Property 3 ✔ ✔ ✔
Property 4
★ ✔ ✔
Property 5 ✔
Favorites
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
?
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
&
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔
Property Features User Features

17.
Alternating Least Squares
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔
User 4 ✔ ✔
User 5 ✔
1. Generate Random User Feature Data
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔
Property 4 ✔
Property 5 ✔
2. Using Favorites, Compute Optimal Property Features
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
× =
3. Using Property Features and Favorites, Compute Optimal User Features
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔
Property 4 ✔
Property 5 ✔
×
4 Bedroom Near Park Good School Views New Build
User 1 ✔
User 2 ✔ ✔ ✔
User 3 ✔
User 4 ✔ ✔
User 5 ✔ ✔
=
4. Repeat
The User Features & Property Features converge to a model that fits the training data

18.
Latent Features
Feature 1 Feature 2 Feature 3 Feature 4 Feature N
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
Feature 1 Feature 2 Feature 3 Feature 4 Feature N
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔

19.
Data -> Model -> Prediction
Recommendations
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
★
Property 2 ✔ ✔
★ ★
Property 3 ✔ ✔ ✔
Property 4
★ ✔ ✔
Property 5 ✔
Favorites
User 1 User 2 User 3 User 4 User 5
Property 1 ✔
Property 2 ✔ ✔
Property 3 ✔ ✔ ✔
Property 4 ✔ ✔
Property 5 ✔
4 Bedroom Near Park Good School Views New Build
Property 1 ✔ ✔
Property 2 ✔ ✔
Property 3 ✔ ✔
Property 4 ✔
Property 5 ✔ ✔
4 Bedroom Near Park Good School Views New Build
User 1 ✔ ✔
User 2 ✔ ✔
User 3 ✔ ✔
User 4 ✔ ✔
User 5 ✔ ✔
Property Features
User Features

20.
Dot Product
Calculating Predictions
Feature 1 Feature 2 Feature 3 Feature 4 Feature N
Property 1 ✔ ✔
Feature 1 Feature 2 Feature 3 Feature 4 Feature N
User 1 ✔ ✔
0 × 0 ++ ++ 1 × 1 1 × 0 0 × 1 0 × 0

21.
Recommendations
Matrix Factorization Model
User 1 User 2 User 3 User 4 User 5
Property 1
9 7 5 5 0
Property 2
9 9 7 7 0
Property 3
7 9 9 9 0
Property 4
5 9 9 9 0
Property 5
0 0 0 0 9

22.
ALS Storage Efficiency
User 1 User 2 User 3 User 4 User 5 User6 User7 User8 User9 User10
Propert
y 1 9 7 5 5 0 9 7 5 5 0
Propert
y 2 9 9 7 7 0 9 9 7 7 0
Propert
y 3 7 9 9 9 0 7 9 9 9 0
Propert
y 4 5 9 9 9 0 5 9 9 9 0
Propert
y 5 0 0 0 0 9 0 0 0 0 9
Propert
y6 9 7 5 5 0 9 7 5 5 0
Propert
y7 9 9 7 7 0 9 9 7 7 0
Propert
y8 7 9 9 9 0 7 9 9 9 0
Propert
y9 5 9 9 9 0 5 9 9 9 0
Propert
y10 0 0 0 0 9 0 0 0 0 9
User
1
User
2
User
3
User
4
User
5
User
6
User
7
User
8
User
9
User
10
Featu
re1
9 7 5 5 0 9 7 5 5 0
Featu
re2
9 9 7 7 0 9 9 7 7 0
Featu
re3
7 9 9 9 0 7 9 9 9 0
Users x Properties
10 x 10
Feature1 Feature2 Feature3
Property 1 9 7 5
Property 2 9 9 7
Property 3 7 9 9
Property 4 5 9 9
Property 5 0 0 0
Property 6 9 7 5
Property 7 9 9 7
Property 8 7 9 9
Property 9 5 9 9
Property 10 0 0 0
Features x Properties
3 x 10
+
Users x Features
10 x 3

23.
Alternating Least Squares
val favorites: Seq[Favorite] = ...
val ratings = favorites.map(fav => Rating(fav.user, fav.prop, 1))
val (userFeatures, propertyFeatures) = ALS.train(ratings = ratings, rank = 5)
val predictions = userFeatures * propertyFeatures
val predictionsForUser = predictions.filter(_.userId == userId)
Spark ML Makes it Easy

24.
Code!
https://github.com/jamesward/dreamhouse-sparkml

25.
Deployment
• Deploy it on Heroku
• Run on a Spark Cluster
• When to train?
• Where to put the model?

26.
PredictionIO
• Template Gallery
• Event Server
• Model Persistence
• Engine Tuning & Versioning
• REST Endpoint for Predictions
Open Source Machine Learning Server

27.
Introducing: Predictive Vision & Sentiment Services `
Leverage existing and train your own deep learning
models to recognize and classify images
Access deep learning models to classify the
sentiment of text
Predictive Vision Service Predictive Sentiment Service
Predictive Sentiment Service
Get started today!
metamind.io
Introducing: Predictive Vision & Sentiment Services

28.
Predictive Services

29.
Thank Y u