recommender systems and linked open data

Recommender Systemsand

Linked Open Data

Tommaso Di Noia

Polytechnic University of BariITALY

11th Reasoning Web Summer School – Berlin August 1, 2015

[email protected]@TommasoDiNoia

Agenda

• A quick introduction to Linked Open Data

• Recommender systems

• Evaluation

• Recommender Systems and Linked Open Data

LINKED OPEN DATAA quick introduction to

Linked Open Datathe Giant Global Graph

Linked (Open) Data

Some definitions:

– A method of publishing data on the Web

– (An instance of) the Web of Data

– A huge database distributed in the Web

– Linked Data is the Semantic Web done right

Web vs Linked Data

Web Linked Data

Analogy File System Database

Designed for Men Machines

(Software Agents)

Main elements Documents Things

Links between Documents Things

Semantics Implicit Explicit

Courtesy of Prof. Enrico Motta, The Open University, Milton Keynes – Uk – Semantic Web: Technologies and Applications.

LOD is the Web

Which technologies?

Which technologies?

Data Language

Query Language

SchemaLanguages

URI

• Every resource/entity/thing/relation isidentified by a (unique) URI

– URI: <http://dbpedia.org/resource/Berlin>

– CURIE: dbpedia:Berlin

– URI: <http://purl.org/dc/terms/subject>

– CURIE: dcterms:subject

Which vocabularies/ontologies?

• Most popular on http://prefix.cc (July 25, 2015)

– YAGO: http://yago-knowledge.org/resource/

– FOAF: http://xmlns.com/foaf/0.1/

– DBpedia Ontology: http://dbpedia.org/ontology/

– DBpedia Properties: http://dbpedia.org/property/

Which vocabularies/ontologies?

• Most popular on http://lov.okfn.org (July 25, 2015)

– VANN: http://purl.org/vocab/vann/

– SKOS: http://www.w3.org/2004/02/skos/core

– FOAF

– DCTERMS

– DCE: http://purl.org/dc/elements/1.1/

RDF – Resource Description Framework

• Basic element: triple

[subject] [predicate] [object]

URI URI

URI | Literal

"string"@lang | "string"^^datatype


dbpedia:Berlin dbo:country dbpedia:Germany .

dbpedia:Berlin rdfs:label "Berlin"@en .

dbpedia:Berlin rdfs:label "Berlino"@it .

dbpedia:Berlin dbo:populationTotal "3517424"^^xsd:integer .

dbpedia:Berlin dcterms:subject category:Capitals_in_Europe .

dbpedia:Berlin rdf:type yago:UrbanArea108675967 .

dbpedia:Germany dbo:language dbpedia:German_Language .

dbpedia:Germany dbo:firstDriverCountry dbpedia:2014_German_Grand_Prix .


Germany Berlin

2014_German_Grand_Prix

German_Language

Capitals_in_Europe

UrbanArea108675967"Berlin"@en

"Berlin"@it

"3517424"^^xsd:integer

country

language

firstDriverCountry

type

subject

label

populationTotal

RDFS and OWL in two statements

dbo:country rdfs:range dbo:Country .

dbpedia:Berlin owl:sameAs freebase:Berlin .

SPARQL

PREFIX dbo: <http://dbpedia.org/ontology/>PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>PREFIX dcterms: <http://purl.org/dc/terms/>PREFIX category: <http://dbpedia.org/resource/Category:>

SELECT DISTINCT ?name ?city WHERE {?city dcterms:subject category:Capitals_in_Europe .?city rdfs:label ?name .?city dbo:populationTotal ?population .FILTER (?population < 30000) .}

SPARQL

curl -g -H 'Accept: application/json' 'http://dbpedia.org/sparql?query=SELECT+DISTINCT+?name+?city+WHERE+{?city+dcterms:subject+category:Capitals_in_Europe+.+?city+rdfs:label+?name+.+?city+dbpedia-owl:populationTotal+?population+.+FILTER+(?population+<+30000)+.}'

RECOMMENDER SYSTEMSThe information overload problem

60 seconds in the Web

Personalized Information Access

• Help the user in finding the information theymight be interested in

• Consider their preferences/past behaviour

• Filter irrelevant information

Recommender Systems

• Help users in dealing with Information/Choice Overload• Help to match users with items

Some definitions

– In its most common formulation, the recommendation problem is reduced to the problem of estimating ratings for the items that have not been seen by a user.

[G. Adomavicius and A. Tuzhilin. Toward the Next Generation of Recommender Systems: A survey of the State-of-the-Art and Possible Extension. TKDE, 2005.]

– Recommender Systems (RSs) are software tools and techniques providing suggestions for items to be of use to a user.

[F. Ricci, L. Rokach, B. Shapira, and P. B. Kantor, editors. Recommender Systems Handbook. Springer, 2011.]

The problem

• Estimate a utility function to automatically predict how much a user will like an item which is unknown to them.

InputSet of users

Set of items

Utility function

𝑈 = {𝑢1 , … , 𝑢𝑀}

𝑋 = {𝑥1 , … , 𝑥𝑁}

𝑓: 𝑈 × 𝑋 → 𝑅

∀ 𝑢 ∈ 𝑈, 𝑥𝑢′ = arg 𝑚𝑎𝑥𝑥∈𝑋 𝑓(𝑢, 𝑥)

Output

The rating matrix

5 1 2 4 3 ??

2 4 5 3 5 2

4 3 2 4 1 3

3 5 1 5 2 4

4 4 5 3 5 2

The

Mat

rix

Tita

nic

I lo

ve s

ho

pp

ing

Arg

o

Love

Act

ual

ly

The

han

gove

r

Tommaso

Vito

Phuong

Jessica

Paolo

The rating matrix(in the real world)

5 ? ? 4 3 ?

2 4 5 ? 5 ?

? 3 ? 4 ? 3

3 5 ? 5 2 ?

4 4 5 ? 5 2

The

Mat

rix

Tita

nic

I lo

ve s

ho

pp

ing

Arg

o

Love

Act

ual

ly

The

han

gove

r

Tommaso

Vito

Phuong

Jessica

Paolo

How sparse is a rating matrix?

Ratings

Explicit

Implicit

Rating Prediction vs Ranking

Best Worst

Recommendation techniques

• Content-based

• Collaborative filtering

• Demographic

• Knowledge-based

• Community-based

• Hybrid recommender systems

Collaborative Recommender Systems

Collaborative RSs recommend items to a user by identifying other users with a similar profile

Recommender System

User profile

Users

Item7

Item15Item11…

Top-N Recommendations

Item1, 5Item2, 1Item5, 4Item10, 5….

….




Content-based Recommender Systems

Recommender System

User profile

Item7

Item15Item11…



Items

Item1Item2

Item100Item’s

descriptions

….

CB-RSs recommend items to a user based on their description and on the profile of the user’s interests

Knowledge-based Recommender Systems

Recommender System

Item7

Item15Item11…


Items

Item1Item2

Item100Item’s descriptions

….

KB-RSs recommend items to a user based on their description and domain knowledge encoded in a knowledge base

Knowledge-base

Collaborative Filtering

• Memory-based

– Mainly based on k-NN

– Does not require any preliminary model building phase

• Model-based

– Learn a predictive model before computingrecommendations

k-Nearest Neighbors

k = 5N

User-based Collaborative Recommendation

5 1 2 4 3 ??

2 4 5 3 5 2

4 3 2 4 1 3

3 5 1 5 2 4

4 4 5 3 5 2

The

Mat

rix

Tita

nic

I lo

ve s

ho

pp

ing

Arg

o

Love

Act

ual

ly

The

han

gove

r

Tommaso

Vito

Phuong

Jessica

Paolo

Pearson’s correlation coefficient

Rate prediction

= 𝑋

Item-based Collaborative Recommendation

5 1 2 4 3 ??

2 4 5 3 5 2

4 3 2 4 1 3

3 5 1 5 2 4

4 4 5 3 5 2

The

Mat

rix

Tita

nic

I lo

ve s

ho

pp

ing

Arg

o

Love

Act

ual

ly

The

han

gove

r

𝑠𝑖𝑚 𝑥𝑖 , 𝑥𝑗 = 𝑥𝑖 ⋅ 𝑥𝑗

|𝑥𝑖| ∗ |𝑥𝑗 |=

σ 𝑟𝑢,𝑥𝑖∗ 𝑟𝑢,𝑥𝑗

𝑢

σ 𝑟𝑢,𝑥𝑖2

𝑢

∗ σ 𝑟𝑢,𝑥2

𝑢

Cosine Similarity

Rate prediction

𝑟ǁ 𝑢𝑖 , 𝑥′ = σ 𝑠𝑖𝑚 𝑥Ԧ, 𝑥Ԧ′ ∗ 𝑟𝑥,𝑢𝑖

𝑥∈𝑋𝑢𝑖

σ 𝑠𝑖𝑚 𝑥Ԧ, 𝑥Ԧ′ 𝑥∈𝑋𝑢𝑖

Adjusted Cosine Similarity

= 𝑋𝑢𝑖

Tommaso

Vito

Phuong

Jessica

Paolo

CF drawbacks

• Sparsity / Cold-start

– New user

– New item

• Grey sheep problem

Content-Based Recommender Systems

• Items are described in terms of attributes/features

• A finite set of values is associated to eachfeature

• Item representation is a (Boolean) vector

Content-based

CB-RSs try to recommend items similar* to those a given user has liked in the past

[P. Lops, M. de Gemmis, G. Semeraro. Content-based Recommender Systems: State of the Art and Trends. Recommender Systems Handbook. 2011]

• Heuristic-based

– Usually adopt techniques borrowed from IR

• Model-based

– Often we have a model for each user

(*) similar from a content-based perspective

CB drawbacks

• Content overspecialization

• Portfolio effect

• Sparsity / Cold-start

– New user

Knowledge-basedRecommender Systems

• Conversational approaches

• Reasoning techniques

– Case-based reasoning

– Constraint reasoning

Hybrid recommender systems

• Weighted

• Switching

• Mixed

• Feature combination

• Cascade

• Feature augmentation

• Meta-level

Robin D. Burke. Hybrid recommender systems: Survey and experiments. User Model. User-Adapt. Interact., 12(4):331–370, 2002.

EVALUATION

Dataset split

20%80%

…

hold-out

k-fold cross-validation

Training Set

Test Set (TS)

Protocols

• Rated test-items

• All unrated items: compute a score for everyitem not rated by the user (also items notappearing in the user test set)

Accuracy metrics for rating prediction

𝑀𝑒𝑎𝑛 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝐸𝑟𝑟𝑜𝑟

𝑅𝑜𝑜𝑡 𝑀𝑒𝑎𝑛 𝑆𝑞𝑢𝑎𝑟𝑒𝑑 𝐸𝑟𝑟𝑜𝑟

MAE and RMSE drawback

• Not very suitable for top-N recommendation

– Errors in the highest part of the recommendationlist are considered in the same way as the ones in the lowest part

Accuracy metrics for top-N recommendation

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 @ 𝑁

𝑃𝑢@𝑁 = |𝐿𝑢 𝑁 ∩ 𝑇𝑆𝑢

+|

𝑁

𝑅𝑒𝑐𝑎𝑙𝑙 @ 𝑁

𝑅𝑢@𝑁 = |𝐿𝑢 𝑁 ∩ 𝑇𝑆𝑢

+|

|𝑇𝑆𝑢+|

𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡 𝐶𝑢𝑚𝑢𝑙𝑎𝑡𝑖𝑣𝑒 𝐺𝑎𝑖𝑛 @ 𝑁

𝐿𝑢 𝑁 is the recommendation list up to the N-th element

𝑇𝑆𝑢+ is the set of relevant test

items for 𝑢

𝐼𝐷𝐶𝐺@𝑁 indicates the score Obtained by an ideal ranking of 𝐿𝑢 𝑁

Is all about precision?

Is all about precision?

• Diversity

– Avoid to recommend only items in a small subset of the catalog

– Suggest diverse items in the recommendation list

• Novelty

– Recommend items in the long tail

• Serendipity

– Suggest unexpected but interesting items

Novelty

𝐸𝑛𝑡𝑟𝑜𝑝𝑦 − 𝐵𝑎𝑠𝑒𝑑 𝑁𝑜𝑣𝑒𝑙𝑡𝑦

RECOMMENDER SYSTEMS AND LINKED OPEN DATA

Content-Based Recommender Systems

P. Lops, M. de Gemmis, G. Semeraro. Content-based recommender Systems: State of the Art and Trends. In: P. Kantor, F. Ricci, L. Rokach, B. Shapira,

editors, Recommender Systems Hankbook: A complete Guide for Research Scientists & Practitioners

Need of domain knowledge!We need rich descriptions of the items!

No suggestion is available if the analyzed content does not contain enough information to discriminate items the user might like from items the user might not like.*

(*) P. Lops, M. de Gemmis, G. Semeraro. Content-based Recommender Systems: State of the Art and Trends. In: P. Kantor, F. Ricci, L. Rokach and B. Shapira, editors, Recommender Systems Handbook: A Complete Guide for Research Scientists & Practitioners

The quality of CB recommendations are correlated with the quality of the features that are explicitly associated with the items.

Limited Content Analysis

Traditional Content-based RecSys

• Base on keyword/attribute -based item representations

• Rely on the quality of the content-analyzer to extract expressive item features

• Lack of knowledge about the items

Semantic-aware approaches

Traditional Ontological/Semantic

Recommender Systems

make use of limited

domain

ontologies;

What about Linked Data?

Use Linked Data to mitigate the limited content analysis issue

• Plenty of structured data available

• No Content Analyzer required

Linking Open Data cloud diagram 2014, by Max Schmachtenberg, Christian Bizer, Anja Jentzsch and Richard Cyganiak. http://lod-cloud.net/

Why RS + LOD

• Multi-Domain knowledge

Why RS + LOD

• Standardized access to dataPREFIX dbpedia: <http://dbpedia.org/resource/>

PREFIX dbo: <http://dbpedia.org/ontology/>SELECT ?actor WHERE {

dbpedia:Pulp_Fiction dbo:starring ?actor .

}

PREFIX yago: <http://yago-knowledge.org/resource/>

PREFIX owl: <http://www.w3.org/2002/07/owl#>PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>PREFIX dbpedia-owl: <http://dbpedia.org/ontology/>

CONSTRUCT{?book ?p ?o .?book yago:linksTo ?yagolink .

}WHERE{

SERVICE <http://live.dbpedia.org/sparql> {

?book rdf:type dbpedia-owl:Book .?book ?p ?o .?book owl:sameAs ?yago .

FILTER(regex(str(?yago),"http://yago-knowledge.org/resource/")) .}

SERVICE <http://lod2.openlinksw.com/sparql> {?yago yago:linksTo ?yagolink .

}

}

Why RS + LOD

• Semantic Analysis

A high level architecture

Item Linker

• Direct Item Linking

• Item Description Linking

Direct item Linking

dbpedia:I_Am_Legend_(film)

Direct item Linking


dbpedia:Troy_(film)

dbpedia:Troy

Direct item Linking


dbpedia:Troy_(film)

dbpedia:Scarface_(1983_film)

dbpedia:Scarface:_The_World_Is_Yours

Direct Item Linking

dbpedia:The_Da_Vinci_Code

dbpedia:Divine_Comedy

Direct Item Linking


Direct Item Linking


dbpedia:Divine_Comedy???

Direct Item Linking

• The easy way

SELECT DISTINCT ?uri, ?title WHERE {?uri rdf:type dbpedia-owl:Film.?uri rdfs:label ?title.FILTER langMatches(lang(?title), "EN") .FILTER regex(?title, "matrix", "i")

}

Direct item Linking

• Other approaches

– DBpedia Lookup

https://github.com/dbpedia/lookup

– Silk Framework

http://silk-framework.com/

Item Description Linking

Item Graph Analyzer

• Build your own knowledge graph

– Select relevant properties. Possible solutions:

• Ontological properties

• Categorical properties

• Frequent properties

– Explore the graph up to a limited depth

Which LOD RSs?

• Content-based

– Heuristic-based

– Model based

• Hybrid

• Knowledge-based

Common features

Linked Data as a structuredinformation source for item descriptions

Rich item descriptions

Different item featuresrepresentations

• Direct properties

• Property paths

• Node paths

• Neighborhoods

• …

Datasets

Subset of Movielens mapped to DBpedia

Subset of Last.fm mapped to DBpedia

Subset of The Library Thing mapped to DBpedia

Mappings

http://sisinflab.poliba.it/semanticweb/lod/recsys/datasets/

Direct properties

Jaccard similarity

Content-based prediction

Vector Space Model for LOD

Righteous Kill

starringdirector

subject/broadergenre

Heat

Ro

ber

tD

e N

iro

Joh

n A

vne

t

Seri

al k

ille

r fi

lms

Dra

ma

Al P

acin

oB

rian

Den

neh

y

He

ist

film

sC

rim

efi

lms

starring

Ro

be

rtD

e N

iro

Al P

acin

o

Bri

an D

en

ne

hy

Righteous KillHeat

… …


Righteous Kill

STARRINGAl Pacino

(v1)

Robert De Niro

(v2)

BrianDennehy

(v3)

Righteous Kill (m1)

X X X

Heat (m2) X X

Heat

Righteous Kill (x1) wv1,x1 wv2,x1 wv3,x1

Heat (x2) wv1,x2 wv2,x2 0

𝑤𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜,𝐻𝑒𝑎𝑡 = 𝑡𝑓𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜,𝐻𝑒𝑎𝑡 ∗ 𝑖𝑑𝑓𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜


Righteous Kill

STARRINGAl Pacino

(v1)

Robert De Niro

(v2)

BrianDennehy

(v3)

Righteous Kill (m1)

X X X

Heat (m2) X X

Heat

Righteous Kill (x1) wv1,x1 wv2,x1 wv3,x1

Heat (x2) wv1,x2 wv2,x2 0

𝑤𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜,𝐻𝑒𝑎𝑡 = 𝑡𝑓𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜,𝐻𝑒𝑎𝑡 ∗ 𝑖𝑑𝑓𝐴𝑙𝑃𝑎𝑐𝑖𝑛𝑜

𝑡𝑓 ∈ {0,1}


+

+

+

… =

𝒔𝒊𝒎𝒔𝒕𝒂𝒓𝒓𝒊𝒏𝒈(𝒙𝒊 , 𝒙𝒋) = 𝒘𝒗𝟏,𝒙𝒊

∗ 𝒘𝒗𝟏,𝒙𝒋+ 𝒘𝒗𝟐,𝒙𝒊

∗ 𝒘𝒗𝟐,𝒙𝒋+ 𝒘𝒗𝟑,𝒙𝒊

∗ 𝒘𝒗𝟑,𝒙𝒋

𝒘𝒗𝟏,𝒙𝒊𝟐 + 𝒘𝒗𝟐,𝒙𝒊

𝟐 + 𝒘𝒗𝟑,𝒙𝒊𝟐

∗ 𝒘𝒗𝟏,𝒙𝒋

𝟐 + 𝒘𝒗𝟐,𝒙𝒋𝟐 + 𝒘𝒗𝟑,𝒙𝒋

𝟐

𝜶𝒔𝒕𝒂𝒓𝒓𝒊𝒏𝒈 ∗ 𝒔𝒊𝒎𝒔𝒕𝒂𝒓𝒓𝒊𝒏𝒈(𝒙𝒊, 𝒙𝒋)

𝜶𝒅𝒊𝒓𝒆𝒄𝒕𝒐𝒓 ∗ 𝒔𝒊𝒎𝒅𝒊𝒓𝒆𝒄𝒕𝒐𝒓(𝒙𝒊, 𝒙𝒋)

𝜶𝒔𝒖𝒃𝒋𝒆𝒄𝒕 ∗ 𝒔𝒊𝒎𝒔𝒖𝒃𝒋𝒆𝒄𝒕(𝒙𝒊, 𝒙𝒋)

𝒔𝒊𝒎 (𝒙𝒊, 𝒙𝒋)

VSM Content-based Recommender

We predict the rating using a Nearest Neighbor Classifier wherein the similarity measure is a linear combination of local property similarities

If this similarity is greater or equal to 0, we suggest the movie m to the user u.

Tommaso Di Noia, Roberto Mirizzi, Vito Claudio Ostuni, Davide Romito, Markus Zanker. Linked Open Data to support Content-based Recommender Systems. 8th International Conference on Semantic Systems (I-SEMANTICS) - 2012




Selected properties




heuristic-based → model-based

Property subset evaluation

The subject+broadersolution is better than only subject or subject+morebroaders.

The best solution is achieved with subject+broader+genres.

Too many broadersintroduce noise.

Rated test items protocol

Evaluation against othercontent-based approaches


Evaluation against other approaches


Property paths

Path-based features

Analysis of complex relations between the user preferences and the target item

Vito Claudio Ostuni, Tommaso Di Noia, Eugenio Di Sciascio, Roberto Mirizzi. Top-N Recommendations from Implicit Feedback leveraging Linked Open Data.

7th Conference on Recommender Systems (RecSys ) – 2013

Data model

I1 i2 i3 i4

u1 1 1 0 0

u2 1 0 1 0

u3 0 1 1 0

u4 0 1 0 1

Implicit Feedback Matrix Knowledge Graph

^

S

Data modelImplicit Feedback Matrix Knowledge Graph

^

S

I1 i2 i3 i4

u1 1 1 0 0

u2 1 0 1 0

u3 0 1 1 0

u4 0 1 0 1

Path-based features

Path: acyclic sequence of relations ( s , .. rl , .. rL )

Frequency of j-th path in the sub-graph related to u and x

• The more the paths, the more the relevance of the item.• Different paths have different meaning.• Not all types of paths are relevant.

u3 s i2 p2 e1 p1 i1 (s, p2 , p1)

Problem formulation

Feature vector

Set of irrelevant items for u

Set of relevant items for u

Training Set

Sample of irrelevant items for u

𝑋𝑢+ = 𝑥 ∈ 𝑋 𝑠Ƹ𝑢𝑥 = 1}

𝑋𝑢− = 𝑥 ∈ 𝑋 𝑠Ƹ𝑢𝑥 = 0}

𝑋𝑢−∗ ⊆ 𝑋𝑢

−

𝑤𝑢𝑥 ∈ ℝ𝐷

TR = ڂ < 𝑤𝑢𝑥 , 𝑠Ƹ𝑢𝑥 > 𝑥 ∈ (𝑋𝑢+ ∪ 𝑋𝑢

−∗)} 𝑢

u1

x1

u2

u3

x2

x3

e1

e3

e4

e2

e5

u4

x4

Path-based features

wu3x1 ?

u1

u2

u3

e1

e3

e4

e2

e5

u4

Path-based features

path(1) (s, s, s) : 1x1

x2

x3

x4

u1

u2

u3

e1

e3

e4

e2

e5

u4

Path-based features

path(1) (s, s, s) : 2x1

x2

x3

x4

u1

u2

u3

e1

e3

e4

e2

e5

u4

Path-based features

path(1) (s, s, s) : 2path(2) (s, p2, p1) : 1

x1

x2

x3

x4

u1

u2

u3

e1

e3

e4

e2

e5

u4

Path-based features

path(1) (s, s, s) : 2path(2) (s, p2, p1) : 2

x1

x2

x3

x4

u1

u2

u3

e1

e3

e4

e2

e5

u4

Path-based features

path(1) (s, s, s) : 2path(2) (s, p2, p1) : 2path(3) (s, p2, p3, p1) : 1

x1

x2

x3

x4

Path-based features

path(1) (s, s, s) : 2path(2) (s, p2, p1) : 2path(3) (s, p2, p3, p1) : 1

u1

u2

u3

e1

e3

e4

e2

e5

u4

x1

x2

x3

x4

Evaluation of different ranking functions

0

0,1

0,2

0,3

0,4

0,5

0,6

given 5 given 10 given 20 given 30 given 50 given All

reca

ll@5

user profile size

Movielens

BagBoo

GBRT

Sum

Evaluation of different ranking functions

0

0,1

0,2

0,3

0,4

0,5

0,6

given 5 given 10 given 20 given All

reca

ll@5

user profile size

Last.fm

BagBoo

GBRT

Sum

Comparative approaches

• BPRMF, Bayesian Personalized Ranking for Matrix Factorization

• BPRLin, Linear Model optimized for BPR (Hybrid alg.)

• SLIM, Sparse Linear Methods for Top-N Recommender Systems

• SMRMF, Soft Margin Ranking Matrix Factorization

MyMediaLite

Comparison with other approaches

0

0,1

0,2

0,3

0,4

0,5

0,6

given 5 given 10 given 20 given 30 given 50 given All

user profile size

Movielens

SPrank

BPRMF

SLIM

BPRLin

SMRMF

pre

cisi

on

@5

Comparison with other approaches

0

0,1

0,2

0,3

0,4

0,5

0,6

given 5 given 10 given 20 given All

user profile size

Last.fm

SPrank

BPRMF

SLIM

BPRLin

SMRMF

pre

cisi

on

@5

Neighborhoods

Graph-based Item Representation

The Godfather

Mafia_films

Gangster_films

American Gangster

Films_about_organized_crime_in_the_United_States

Best_Picture_Academy_Award_winners

Best_Thriller_Empire_Award_winners

Films_shot_in_New_York_City

subject

subjectsubject

subject

subject

subject

subject

Vito Claudio Ostuni, Tommaso Di Noia, Roberto Mirizzi, Eugenio Di Sciascio. A Linked Data Recommender System using a Neighborhood-based Graph Kernel. The 15th International Conference on Electronic Commerce and Web Technologies – 2014


The Godfather

Mafia_films Films_about_organized_crime

Gangster_films

American Gangster


Films_about_organized_crime_by_country



Awards_for_best_film


subject

subjectsubject

broader

broader

broader

broader

broader

subject

subject

subject

subject


The Godfather


Gangster_films

American Gangster







subject

subjectsubject

broader

broader

broader

broader

broader

broader

subject

subject

subject

subject


The Godfather


Gangster_films

American Gangster







subject

subjectsubject

broader

broader

broader

broader

broader

broader

subject

subject

subject

subject

Exploit entities descriptions

h-hop Item Neighborhood Graph

The Godfather


Gangster_films

Best_Picture_Academy_Award_winners Awards_for_best_film


subject

subjectsubject

broader

broader

broader

Kernel Methods

Work by embedding data in a vector space and looking for linear patterns in such space

𝑥 → 𝜙(𝑥)

[Kernel Methods for General Pattern Analysis. Nello Cristianini . http://www.kernel-methods.net/tutorials/KMtalk.pdf]

𝜙(𝑥)

𝜙𝑥Input space Feature space

We can work in the new space F by specifying an inner product function between points in it

𝑘 𝑥𝑖, 𝑥𝑗 = < 𝜙(𝑥𝑖), 𝜙(𝑥𝑗)>

h-hop Item Entity-based Neighborhood Graph Kernel

Explicit computation of the feature map

entity importance in the item neighborhood graph

𝑘𝐺ℎ 𝑥𝑖, 𝑥𝑗 = 𝜙𝐺ℎ 𝑥𝑖 , 𝜙𝐺ℎ 𝑥𝑗

𝜙𝐺ℎ 𝑥𝑖 = (𝑤𝑥𝑖 ,𝑒1, 𝑤𝑥𝑖 ,𝑒2

, …, 𝑤𝑥𝑖 ,𝑒𝑚, … , 𝑤𝑥𝑖 ,𝑒𝑡

)


# edges involving 𝑒𝑚 at l hops from 𝑥𝑖

a.k.a. frequency of the entity in the

item neighborhood graph

factor taking into account at which hop the entity appears

h-hop Item Entity-based Neighborhood Graph Kernel


𝜙𝐺ℎ 𝑥𝑖 = (𝑤𝑥𝑖 ,𝑒1, 𝑤𝑥𝑖 ,𝑒2

, …, 𝑤𝑥𝑖 ,𝑒𝑚, … , 𝑤𝑥𝑖 ,𝑒𝑡

)

Weights computation example

i

e1e2

p3

p2

e4

e5

p3p3

h=2

𝑐𝑃1 𝑥𝑖 ,𝑒1= 2

𝑐𝑃1 𝑥𝑖 ,𝑒2= 1

𝑐𝑃2 𝑥𝑖 ,𝑒4= 1

𝑐𝑃2 𝑥𝑖 ,𝑒5= 2

Weights computation example

i

e1e2

p3

p2

e4

e5

p3p3

h=2

𝑐𝑃1 𝑥𝑖 ,𝑒1= 2

𝑐𝑃1 𝑥𝑖 ,𝑒2= 1

𝑐𝑃2 𝑥𝑖 ,𝑒4= 1

𝑐𝑃2 𝑥𝑖 ,𝑒5= 2

Informative entity about the item even if not directly related to it

Experimental Settings

• Trained a SVM Regression model for each user

• Accuracy Evaluation: Precision, Recall

• Novelty Evaluation: Entropy-based Novelty (All Items protocol) [the lower the better]

Kernel calibrationimpact of alpha params

Comparative approaches

•NB: 1-hop item neigh. + Naive Bayes classifier

•VSM: 1-hop item neigh. Vector Space Model (tf-idf) + SVM regr

•WK: 2-hop item neigh. Walk-based kernel + SVM regr

Comparison with other approaches (i)

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

Prec@10 [20/80] Prec@10 [40/60] Prec@10 [80/20]

NK-bestPrec

NK-bestEntr

NB

VSM

WK


Comparison with other approaches (ii)

0

0,2

0,4

0,6

0,8

1

1,2

1,4

1,6

1,8

EBN@10 [20/80] EBN@10 [40/60] EBN@10 [80/20]

NK-bestPrec

NK-bestEntr

NB

VSM

WK

Node paths

The FreeSound case study

Vito Claudio Ostuni, Sergio Oramas, Tommaso Di Noia, Xavier Serra, Eugenio Di Sciascio. A Semantic Hybrid Approach for Sound Recommendation. 24th

World Wide Web Conference - 2015

FreeSound Knowledge Graph

Item textual descriptions enrichment: Entity Linking tools can be usedto enrich item textual descriptions with LOD


# sequences and subsequences of nodes

from 𝑥𝑖 to em

Normalization factor

h-hop Item Node-Based Neighborhood Graph Kernel

𝜙𝐺ℎ 𝑥𝑖 = (𝑤𝑥𝑖 ,𝑝∗1, …, 𝑤𝑥𝑖 ,𝑝∗𝑚

, … , 𝑤𝑥𝑖 ,𝑝∗𝑡)


Hybrid Recommendation via Feature Combination

The hybridizations is based on the combination of different data sources

Final approach: collaborative + LOD + textual description + tags

Users who rated the item

u1 u2 u3 …. entity1 entity2 …. keyw1 keyw2 … tag1 …

entities from the knowledgegraph (explicit feature mapping)

Keywords extracted from the textual description

tags associated to the item

Item Feature Vector

Accuracy

All items protocol

Long Tail

Aggregate Diversity

• Feature combination hybrid approach

• adding collaborative features to item content feature vectors can improveconsiderably recommendation accuracy

• Semantic Enrichment

• semantics can help in improving different performances beyond accuracysuch as novelty and catalog coverage

Hybrid approaches: some lessons learnt

Dataset selection

Select the domain(s) of your RS

SELECT count(?i) AS ?num ?c

WHERE {

?i a ?c .

FILTER(regex(?c, "^http://dbpedia.org/ontology")) .

}

ORDER BY DESC(?num)

Does the LOD dataset selectionmatter?

A comparison betweenDBpedia and Freebase

Accuracy Coverage Diversity Novelty

Freebase + + - -

DBpedia - - + +

Phuong Nguyen, Paolo Tomeo, Tommaso Di Noia, Eugenio Di Sciascio. Content-based recommendations via DBpedia and Freebase: a case study in the music domain. The 14th International Semantic Web Conference - ISWC 2015

A comparison betweenDBpedia and Freebase

Accuracy Coverage Diversity Novelty

1-hop - - - +

2-hop + + + -

Phuong Nguyen, Paolo Tomeo, Tommaso Di Noia, Eugenio Di Sciascio. Content-based recommendations via DBpedia and Freebase: a case study in the music domain. The 14th International Semantic Web Conference - ISWC 2015

Conclusions

• Linked Open Data to enrich the content descriptions of item

• Exploit different characteristcs of the semantic network to represent/learn features

• Improved accuracy• Improved novelty• Improved Aggregate Diversity• Entity linking for a better expoitation of text-based

data• Select the right approach, dataset, set of properties to

build your RS

Open issues

• Generalize to graph pattern extraction to represent features

• Automatically select the triples related to the domain of interest

• Automatically select meaningful properties to represent items

• Analysis with respect to «knowledgecoverage» of the dataset– What is the best approach?

Not covered here

• User profile

• Preferences

• Context-aware

• Knowledge-based approaches

• …

Many thanks to the RecSys crew @ SisInf Lab

Roberto Mirizzi

now at Yahoo! CA

Vito Claudio Ostuni

now at

Jessica Rosati

Phd Fellowship Awardee @

Paolo Tomeo

Jindřich Mynarz

Phuong Nguyen

Sergio Oramas

Aleksandra Karpus

Visiting Students and PostDoc

Recommender Systemsand

Linked Open Data

Tommaso Di Noia

Polytechnic University of BariITALY

11th Reasoning Web Summer School – Berlin August 1, 2015

[email protected]@TommasoDiNoia

recommender systems and linked open data

Presentations & Public Speaking