Recommendation service using Prediction IO framework. Facebook Open Academy October, 2014

1. Yuki Furuta NaotoYamamoto Tran Hoan
Facebook Open Academy
International

2. What is ?
An open source machine
learning server
For software developers
to create predictive
features in their web
and mobile app.
Currently powering
thousands of developers
and hundred of
applications1
1 http://github.com/PredictionIO/
1

3. Architecture
Horizontally scalability
Spark
Data
Preparator
Model 1
Model N
HBase
Query
Prediction
ResultData
Source
Import Data
(EventServer)
Algorithm
1
Algorithm
N
ServingHDFS
Spark
.
.
.
http://docs.prediction.io/resources/systems/
Web App
Mobile App
Productivity
Data In Data Out
2

4. What can do?
Content-based recommendation
Trend detection Sentiment Analysis
Restaurant recommendation User similarity
Data analysis
Engine
(recommendation, rank,…)
YELPIO-NAVI
3
MovieLens

5. YELPIO-NAVI
NaotoYamamoto Tran Hoan
Recommendation App for Restaurants
UsingYelp! Dataset
Inhwan Eric Lee
（JP）（JP）（USA）

6. What is YELPIO-NAVI
Yelp:
食べログ in America
Information of
restaurants’ address, stars
users’ stars
Recommendation of Restaurants
Using These Information
５

7. YELPIO-NAVI Demo Setup
Batch import data
through RubySDK
Store & Retrieve
Business Data
Retrieve & Store
Business Data
through REST API
Retrieve Prediction
Results through REST API
https://github.com/OminiaVincit/predictionio_rails
http://yelpio.hongo.wide.ad.jp/
https://github.com/OminiaVincit/YELPIO_demo2
(1) Neighbourhood model
(2) Collaborative Filtering
http://www.yelp.com/
６

8. YELPIO-NAVI Demo
http://yelpio.hongo.wide.ad.jp/
７ http://zorovn.hongo.wide.ad.jp/

9. MovieLens
Content-based Movie Recommendation
Yuki FurutaNhu-Quynh BethYue ShiShaocong Mo
（JP）（USA） （USA） （USA）

10. x MovieLens
- Content-Based Movie Recommendation Engine -
A B
A. Collaborative Filtering
９

11. x MovieLens
MovieLens Datasets
• 100,000 ratings (1-5)
from 943 users on 1682
movies
• Simple demographic
info for the users (age,
gender, occupation, zip)
• Information about the
movies (title, release
date, genre)
- Content-Based Movie Recommendation Engine -
B. Content-Based
A (age: 20, male, RUS) B (age: 21, male, KZH)
20-year-old man likes:
• Action 60%
• Comedy 10%
• English 10%
• etc.
10

12. x MovieLens
- Content-Based Movie Recommendation Engine -
Dataset
val DataSourceAttributeNames
= AttributeNames(
user = "pio_user",
item = "pio_item",
u2iActions = Set("rate"),
itypes = "pio_itypes",
starttime = "pio_starttime",
endtime = "pio_endtime",
inactive = "pio_inactive",
rating = "pio_rating")
Feature Based
User Based
Algorithms
PreparationReading Data
Query
Serve
MovieLens
- User (ID, Age, Gender,
Occupation, Zip)
- Movie (ID, Title, Year,
Genre, Actors,…)
Prepare Train
11

13. x MovieLens
- Content-Based Movie Recommendation Engine -
Stanlay Kubricks
America
Comedy
Black
SF
Rowan Atkinson
United Kingdom
Comedy
SF
Action
Feature Based Algorithm
Michael
12

14. x MovieLens
- Content-Based Movie Recommendation Engine -
Stanlay Kubricks
USA
Comedy
Black
Scientific
Fantasy
Rowan Atkinson
United Kingdom
Comedy
SF
Action
Fantasy
Comedy
Fantasy
Action
USA
Mark Wahlberg
USA
Comedy
Fantasy
Action
Recommend!
Feature Based Algorithm
Michael
13

15. x MovieLens
- Content-Based Movie Recommendation Engine -
Feature Based Algorithm
UserID: 1, Age: 24, Gender: M, Occupation: technician, Zip: 85711
UserID: 2, Age: 53, Gender: F, Occupation: other, Zip: 94043
UserID: 3, Age: 23, Gender: M, Occupation: writer, Zip: 32067
UserID: 4, Age: 24, Gender: M, Occupation: technician, Zip: 43537
UserID: 5, Age: 33, Gender: F, Occupation: other, Zip: 15213
User: 196 rates Movie: 242 (3.0 / 5)
User: 186 rates Movie: 302 (3.0 / 5)
User: 22 rates Movie: 377 (1.0 / 5)
User: 244 rates Movie: 51 (2.0 / 5)
User: 166 rates Movie: 346 (1.0 / 5)
Threshold (e.g. 2.0)
BUY
BUY
-
-
-
Train
Query
e.g.
Recommend 5 movies for UserID: 2
Recommend 5 movies which are “Comedy” for UserID:2
Recommend 2 movies which are “Action” by Rowan Atkinson for UserID: 2
1. MovieID: 297 Score: -8.53295620539528
2. MovieID: 251 Score: -13.326537513274323
3. MovieID: 292 Score: -15.276804370241758
4. MovieID: 290 Score: -32.944167483781335
5. MovieID: 314 Score: -37.45527366828404
Predict
14

16. …to be continued
Scale for Big Data
Multi-engines & Multi-algorithms
Predict with more features
…
15
Evaluation

17. Thank you for listening
Japanese team
Yuki Furuta NaotoYamamoto Tran Hoan