talk bedcon berlin 2012

7/31/2019 Talk BedCon Berlin 2012

1/42

How to build a recommender system based on Mahout and

Java EE

Berlin Expert Days 29. 30. March 2012 Manuel Blechschmidt CTO Apaxo GmbH


2/42

All the web content will be personalized in threeto five years.

Sheryl Sandberg COO Facebook 09.2010


3/42

Agenda

- What is personalization?

- What algorithms can be used?- Architecture of a recommender system

- How to bundle Mahout into an Java EE

application- Interface with the recommender

- Conclusion


4/42

What is personalization?

Personalization involves using technology toaccommodate the differences betweenindividuals. Once confined mainly to the Web, itis increasingly becoming a factor in education,health care (i.e. personalized medicine),television, and in both "business to business"and "business to consumer" settings.

Source:https://en.wikipedia.org/wiki/Personalization


5/42

Amazon.com


6/42

TripAdvisor.com


7/42

eBay


8/42

criteo.com - Retargeting


9/42

Zalando


10/42

Plista


11/42

YouTube


12/42

Naturideen.de (coming soon)


13/42

Recommender

This talk will concentrate on recommender technology based on collaborative filtering (cf)to personalize a web site

- a lot of research is going on

- cf has shown great success in movie and musicindustry

- recommenders can collect data silently and useit without manual maintenance


14/42

What is a recommender?Let U be a set of users of the recommendation system and I be the

set of items from which the users can choose. A recommender r is a function which produces for a user u

ia set of recommended

items R kwith k entries and a binary, transitive, antisymmetric and

total relation prefers_over ui

which can be used for sorting the

recommendations for the user. The recommender r is oftencalled a top-k recommender.


15/42

WARNING!

The next slides contain math which takes normally3 years to understand so don't be disappointedif you don't get it.


16/42

What should wolf and sheep eat?Ms. Baumgartner your sun is ignoringagain the food chain.

SON!

Source: http://static.nichtlustig.de/comics/full/081009.jpg
http://static.nichtlustig.de/comics/full/081009.jpghttp://static.nichtlustig.de/comics/full/081009.jpg


17/42

Demo Data

Carrots Grass Pork Beef Corn FishRabbit 10 7 1 2 ? 1

Cow 7 10 ? ? ? ?

Dog ? 1 10 10 ? ?

Pig 5 6 4 ? 7 3Chicken 7 6 2 ? 10 ?

Pinguin 2 2 ? 2 2 10

Bear 2 ? 8 8 2 7

Lion ? ? 9 10 2 ?Tiger ? ? 8 ? ? 5

Antilope 6 10 1 1 ? ?

Wolf 1 ? ? 8 ? 3

Sheep ? 8 ? ? ? 2


18/42

Characteristics of Demo Data

Ratings from 1 10

Users: 12

Items: 6

Ratings: 43 (unusual normally 100,000 100,000,000)

Matrix filled: ~60% (unusual normally sparse around 0.5-2%)

Average Number of Ratings per User: ~3.58 Average Number of Ratings per Item: ~7.17

Average Rating: ~5.579

https://github.com/ManuelB/facebook-recommender-demo/tree/master/docs/BedConExamples.R
https://github.com/ManuelB/facebook-recommender-demo/tree/master/docs/BedConExamples.Rhttps://github.com/ManuelB/facebook-recommender-demo/tree/master/docs/BedConExamples.R


19/42


20/42


21/42

YouTube Rating Distribution

http://techcrunch.com/2009/09/22/youtube-comes-to-a-5-star-realization-its-ratings-are-useless/
http://techcrunch.com/2009/09/22/youtube-comes-to-a-5-star-realization-its-ratings-are-useless/http://techcrunch.com/2009/09/22/youtube-comes-to-a-5-star-realization-its-ratings-are-useless/


22/42

Model and Memory Approaches

- Item- or User- Based Collaborative Filtering

- Idea suggest what similair users did

- Matrix Factorization e.g Singular Value Decomposition- Try to create a profile based of factors for users and items

Main difference:

A model base approach tries to extract the underlying logic fromthe data.


23/42

User Based Approach

- Find similar or opposite animals like wolf

- Checkout what these other animals like- Recommend this to wolf


24/42

Find animals which voted for beef,fish and carrots too

Carrots Grass Pork Beef Corn Fish

Wolf 1 ? ? 8 ? 3

Pinguin 2 2 ? 2 2 10

Bear 2 ? 8 8 2 7

Rabbit 10 7 1 2 ? 1

Cow 7 10 ? ? ? ?

Dog ? 1 10 10 ? ?

Pig 5 6 4 ? 7 3

Chicken 7 6 2 ? 10 ?Lion ? ? 9 10 2 ?

Tiger ? ? 8 ? ? 5

Antilope 6 10 1 1 ? ?

Sheep ? 8 ? ? ? ?


25/42

Pearson Correlation

- 1 = very similar

- (-1) = complete opposite votings

- similarity between wolf and pinguin: -0.2401922

- cor(c(8,3,1),c(2,10,2))

- similarity between wolf and bear: 0.8196562

- cor(c(8,3,1),c(8,7,2))- similarity between wolf and rabbit: -0.6465846

- cor(c(8,3,1),c(2,1,10))


26/42

Weighted sum of the ratings

Pork (10):

(0.8196 * 8 + (-0.6466) * 1) / (0.8196 + (-0.6465)

= 34 ~ 10

Grass (5.65):

(2*(-0.2401)+7*(-0.6465)) / ((-0.2401) + (-0.6465))= 5,65

Corn (2,0):

2*(-0.2401)+2*(0.8196)) / (-0.2401+0.8196) = 2,0


27/42

Predicted ratings

- Wolf should eat: Pork Rating: 10.0

- Wolf should eat: Grass Rating: 5.645701

- Wolf should eat: Corn Rating: 2.0


28/42

SVD

http://public.lanl.gov/mewall/kluwer2002.html
http://public.lanl.gov/mewall/kluwer2002.htmlhttp://public.lanl.gov/mewall/kluwer2002.html


29/42

Factorized Matrixes


30/42

Predicted Matrix (k = 2)


31/42

Predicted Ratings

- Sheep should eat: Carrots Rating: 6.01

- Sheep should eat: Corn Rating: 5.83


32/42

What other algorithms can be used?

Similarity Measures for Item or User based:

- LogLikelihood Similarity

- Cosine Similarity

- Pearson Similarity

- etc.

Estimating algorithms for SVD:- ALSWRFactorizer

- ExpectationMaximizationSVDFactorizer


33/42

Let's built it


34/42

Mahout - Taste

Taste is a flexible, fast collaborative filteringengine for Java.

It was primarly written by Sean Owen and is partof the Apache Mahout library.

Another important author is: Sebastian Schelter

It offers all the algorithms introduced here and isOpen Source

http://mahout.apache.org/taste.html
http://mahout.apache.org/taste.htmlhttp://mahout.apache.org/taste.html


35/42

Architecture of the recommender


36/42

Packaging


37/42

Eclipse Project


38/42

Maven pom.xml


39/42


40/42

Demo and live Debugging


41/42

Conclusion

Recommendation is a lot of math

You shouldn't implement the algorithms again

There are a lot of unsanswered questions

- Scalibility, Performance, Usability

You can gain a lot from good personalization


42/42

More sources

http://www.apaxo.de /

http://mahout.apache.org /

http://research.yahoo.com /

http://www.grouplens.org/

http://recsys.acm.org/

https://github.com/ManuelB/facebook-recommender-demo/

http://docs.oracle.com/javaee/6/tutorial/doc/
http://www.apaxo.de/http://mahout.apache.org/http://research.yahoo.com/http://www.grouplens.org/http://recsys.acm.org/https://github.com/ManuelB/facebook-recommender-demo/http://docs.oracle.com/javaee/6/tutorial/doc/http://docs.oracle.com/javaee/6/tutorial/doc/https://github.com/ManuelB/facebook-recommender-demo/http://recsys.acm.org/http://www.grouplens.org/http://research.yahoo.com/http://mahout.apache.org/http://www.apaxo.de/

talk bedcon berlin 2012

Documents