Recommender Systems and SPARQL: More than a Shotgun …jperez/amw2014/Arras... · Recommender Systems and SPARQL: More than a Shotgun Wedding? 5 1. Motivation Motivation Recommender

Recommender Systems and SPARQL: More than a Shotgun

Wedding?

8th Alberto Mendelzon International Workshop

Cartagena, Colombia

(AMW 2014)

Victor Anthony Arrascue Ayala Martin Przyjaciel-Zablocki Thomas Hornung Alexander Schätzle Georg Lausen University of Freiburg Databases & Information Systems

June 5 2014

1. Motivation

2. RecSPARQL

3. Experiments

4. Summary

Recommender Systems and SPARQL: More than a Shotgun Wedding?

Overview

2 1. Motivation

RDF’s flexibility ◦ Recommendation domain

Users, Items, Ratings

Consumption relationship

Recommender Systems and SPARQL: More than a Shotgun Wedding? 3 1. Motivation

Motivation

How to get recommendations from RDF-graphs: SPARQL? ◦ Retrieval of explicit data from RFD-graphs

◦ Graph pattern matching

◦ Flexible

◦ No possible to express fuzzy queries

Similarity


Motivation

SPARQL

Source: http://courses.ischool.berkeley.edu/i253/f11/


Motivation Recommender

Systems

How to get recommendations from RDF-graphs: Recommender Systems ◦ Predicts a degree of preference for a user towards a set of

non-consumed items

◦ Based on Information Retrieval techniques

Source: http://courses.ischool.berkeley.edu/i253/f11/

Recommender Systems and SPARQL: More than a Shotgun Wedding? 6

Motivation

A straightforward approach

1. Motivation

Classic recommender ◦ Input: users, items, ratings

◦ Output: users, recommended items, predicted rating

Recommendations

7

Motivation

Example: lack of flexibility ◦ Collaborative filtering approach

Recommendations from similar users

1. Motivation

1 3 5 0 2

1 0 4 1 1

3 2 0 0 3

0 3 4 1 5

1 3 5 0 1

2 1 2 3 2

Recommender Systems and SPARQL: More than a Shotgun Wedding?

(A)Extraction /

Pre-processingSPARQL

(B)Recommendation

System (CF)


Motivation

Example: lack of flexibility ◦ Collaborative filtering approach


1. Motivation

U1

U5

U17

U43

U55

1 3 5 0 2

1 0 4 1 1

3 2 0 0 3

0 3 4 1 5

1 3 5 0 1

2 1 2 3 2

(B)Recommendation

System (CF)


Motivation

Example: lack of flexibility ◦ Collaborative approach


1. Motivation

U1

U5

U17

U43

U55

◦ Customization

Neighbors geographically close

Age of neighbors differ by 𝛿

Speak same languages

etc….


Motivation



1. Motivation

U1

U5

U17

U43

U55

◦ Customization



Speak same languages

etc….

Where is the problem? ◦ Fixed recommender model

1. Motivation

2. RecSPARQL

3. Experiments

4. Summary

Recommender Systems and SPARQL: More than a Shotgun Wedding? 2. Approach

Overview

11


Our Approach

RecSesame

◦ Platform for the evaluation of RecSPARQL queries

Based on Sesame’s framework

Cache System

2. Approach

RecSPARQL: Recommendations + SPARQL

◦ Extension of SPARQL 1.1

Consistent mechanism to select parts of the graph

Flexibility


RecSPARQL in a nutshell

2. Approach

RECOMMEND [Projected Variables]

USING [Recommendation Algorithm]

WHERE { [Basic Graph Pattern] }

BASED ON { [RecSPARQL Type Pattern]

[RecSPARQL Model Building Pattern] }



2. Approach






◦ Must contain recommendation entities



2. Approach






◦ Algorithm used

Algorihtms ◦ Content-based (CB)

◦ Collaborative Filtering (CF)

◦ Hybrid (H)



2. Approach






◦ Similarity

criteria

Content-based ◦ Genre

◦ Cast

◦ Director

Collaborative filtering ◦ Watched movies and ratings

◦ Age

◦ Geographical location




2. Approach






◦ Similar to projection

◦ Makes it possible to project recommendations

◦ ?movie.REC

Filters ◦ Recommend only action movies


RecSPARQL’s flexibility

2. Approach

FILTER ( ?genre.REC = “action”)

?genre.REC

?genre.REC

Filters ◦ Recommend only from users whose age differ by at most 5

years



2. Approach

FILTER ( abs(xsd:integer(?age) - xsd:integer(?age.REC)) <= 5)

?age.REC

?age

Much more… ◦ Recommend movies watched under a certain context

◦ Recommend movies whose directors have the same citizenship of the user for which we want the recommendations

◦ Recommend ….



2. Approach

FILTER ( ?watchTime.REC = “weekend” && ?company.REC = “partner”) .

1. Motivation

2. RecSPARQL

3. Experiments

4. Summary

Recommender Systems and SPARQL: More than a Shotgun Wedding? 4. Experiments

Experiments

21


Experiments



1. Motivation

U1

U5

U17

U43

U55

Gradual restriction of Neighborhood: ◦ Neighbors geographically close

◦ Age of neighbors differ by 𝛿


Experiments



1. Motivation

U1

U5

U17

U43

U55

FILTER ( abs(xsd:integer(?age) - xsd:integer(?age.REC)) <= %K%) . FILTER ( abs(xsd:integer(?zip) - xsd:integer(?zip.REC)) <= %L% )


Experiments

Benefitial ◦ Backed by our experiments

1. Motivation

U1

U5

U17

U43

U55

◦ Similarity among users in the neighborhood

Increases

◦ Customization




Experiments

Benefitial ◦ Backed by our experiments

1. Motivation

U1

U5

U17

U43

U55

◦ Similarity among users in the neighborhood

Increases

◦ Customization



1 3 5 0 2

1 0 4 1 1 3 2 0 0 3 0 3 4 1 5

1 3 5 0 1

2 1 2 3 2

Tight integration of recommender systems with SPARQL

Customizable recommendations on arbitrary RDF graphs

Future Work ◦ Enhance the integration of both paradigms

◦ Support more recommendations techniques

◦ Increase the expressiveness of RecSPARQL

Binding variables

Sub-queries

Property paths

Recommender Systems and SPARQL: More than a Shotgun Wedding? 26 5. Summary

Summary

V. A. Arrascue Ayala, M. Przyjaciel-Zablocki, T. Hornung, A. Schätzle, G. Lausen, Extending SPARQL for Recommendations. In SWIM (ACM SIGMOD), pages 1-8, 2014.

Recommender Systems and SPARQL: More than a Shotgun Wedding? Thanks

Thanks for your attention!

27

Documents

Recommender Systems and SPARQL: More than a Shotgun …jperez/amw2014/Arras... · Recommender Systems and SPARQL: More than a Shotgun Wedding? 5 1. Motivation Motivation Recommender