Computing the Semantic Similarity of Resources

in DBpedia for Recommendation Purposes

Guangyuan Piao, Safina showkat Ara, John G. Breslin Insight Centre for Data Analytics @NUI Galway, Ireland

Unit for Social Software

The 5th Joint International Semantic Technology Conference Yichang, China, 12/11/2015

Contents

•  Introduction

•  Related Work

•  Resim (Resource similarity) Measure

•  Evaluation Setup and Results

•  Study of Linked Data Sparsity Problem

•  Conclusions

2

•  Linked Data (especially DBpedia) has been used for various applications including recommendations:

•  LOD-enabled Recommender Systems Challenge (ESWC’14, 15)

•  User Modeling for Personalization in Online Social Networks

•  Use entities/resources in a Knowledge Graph (e.g., DBpedia, Freebase) to represent user interests

•  measuring the semantic similarity between resources is important

3

Introduction

•  Linked Data for Recommendation Purposes (single domain)

4

Introduction

dbpedia:Cheryl_Cole

•  measure the semantic similarity in the context of DBpedia

•  recommend similar items based on what you like in a single domain (e.g., music, movie)

Who is the most similar artist to Cheryl Cole?

•  Linked Data for Recommendation Purposes (social domain)

5

Introduction

dbpedia:Cheryl_Cole

•  user interests can be any topical resources in DBpedia

•  can we reuse the similarity measures that were designed for recommendations in single domain?

dbpedia:SIOC

dbpedia:Linked_data

wi1:preference

What news the user will be interested in?

1.  http://smiy.sourceforge.net/wi/spec/weightedinterests.html

6

Related Work

•  LDSD (Linked Data Semantic Distance) – Passant, 2010 •  evaluated on music artist recommendations •  widely used and has comparative performance with supervised learning

approaches

•  Shakti – Leah, 2012 •  similarity was measured based on proximity: two entities are more

similar if they have more number of paths (penalty for longer paths)

•  some problems need to be addressed: •  not suitable for measuring the similarity between general resources •  fundamental axioms are violated •  performance over each other is unproven

•  supervised learning approaches (Di Noia etc.)

sim(ra, ra) = sim(rb, rb), for all resources ra and rb

7

Fundamental Axioms

equal self-similarity

sim(ra, rb) = sim(rb, ra), for all resources ra and rb

symmetry

sim(ra, ra) > sim(ra, rb), for all resources ra ≠ rb

minimality

•  http://www.scholarpedia.org/article/Similarity_measures

8

The goal of the paper

propose a semantic similarity measure - Resim on top of a revised LDSD to

satisfy fundamental axioms

be able to measure the semantic similarity between general resources

provide a comparative study

study Linked Data sparsity problem

9

Linked Data Semantic Distance (LDSD) List_of_The_Tonight_Show_with_Jay_Leno

_episodes_(2013–14)

Category:21st-century_American_singers

Ariana_Grande

Selena_Gomez

musicalguests musicalguests

subject

subject

associatedMusicArtist

influences

Cd(influences, Ariana_Grande, Selena_Gomez) = 1

Cii(musicalguests, Ariana_Grande, Selena Gomez) = 1

10

Resim (Resource similarity) Measure - 1

sim(ra, ra) = sim(rb, rb), for all resources ra and rb

equal self-similarity

sim(ra, ra) > sim(ra, rb), for all resources ra ≠ rb

minimality

✔

✔

•  to satisfy “equal self-similarity” and “minimality” axioms

11

Resim (Resource similarity) Measure - 2

sim(ra, rb) = sim(rb, ra), for all resources ra and rb

symmetry

✔

•  to satisfy “symmetry” axiom

•  incorporating property similarity

•  from the definition of an ontology, the properties of each concept describe various features and attributes of the concept.

•  Thus, property similarity is important when there is no similarity can be indicated using LDSD’

•  property similarity measure

•  based on the number of shared incoming/outgoing properties

•  Csip: shared incoming properties, Cip: # of incoming properties

12

Resim (Resource similarity) Measure - 3

•  w1 = 1 and w2 = 2 for the experiment

13

Resim (Resource similarity) Measure

final equation for Resim

14

Evaluation Setup and Results

1.  similarity measures evaluated on axioms

2.  evaluation on calculating similarities for general resources

Axiom LDSDsim Shakti Resim

equal self-similarity ✔

symmetry ✔ ✔

minimality ✔ ✔

(1) extract word pairs from WordSim353 dataset

sim(Wa, Wb) > sim(Wa, Wc)

the difference is higher than 2

(2) retrieve the corresponding DBpedia resources

construct a test pair as sim(ra, rb) > sim(ra, rc)

15

Evaluation Setup and Results

•  Resim performs best compared to other approaches •  satisfy 23 out of 28 test pairs of general resources

Test pairs of resources LDSDsim Shakti Resim

sim(dbpedia:Money, dbpedia:Currency) > sim(dbpedia:Money,

dbpedia:Business_operations) ✔ ✔

sim(dbpedia:Money, dbpedia:Cash) > sim(dbpedia:Money,

dbpedia:Demand_deposit) ✔ ✔

… > … …

sim(dbpedia:Planet, dbpedia:Moon) > sim(dbpedia:Planet,

dbpedia:People) ✔ ✔

sim(dbpedia:Coast, dbpedia:Shore) > sim(dbpedia:Coast,

dbpedia:Hill) ✔ ✔

… … …

Total: 13 18 23

Evaluation Setup and Results

•  10 similar music artists from Last.fm for given artist

golden truth

•  200 randomly selected music artists from 75,682 resources in DBpedia of type dbpedia-owl:MusicArtist or dbpedia-owl:Band

candidate list

•  Recall and Mean Reciprocal Rank (MRR)

evaluation methods

3.  evaluation on LOD recommender system (music domain)

Evaluation Setup and Results

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Shakti3 Shakti5 LDSDsim Resim

MRR

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Shakti3 Shakti5 LDSDsim Resim

R@5

R@10

R@20

Recall@5, 10, 20

MRR

3.  evaluation on LOD recommender system (music domain)

18

Study of Linked Data Sparsity Problem

•  Linked Data Sparsity Problem: •  the performance of the recommender system based on similarity

measures of resources decreases when resources lack information (i.e., when they have a lesser number of incoming/outgoing relationships to other resources).

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

R@5 R@10 R@20

Random

Popular

The average performance of recommendations

on popular music artists

The average performance of recommendations

on random music artists

19

Study of Linked Data Sparsity Problem

1XPEHU�ORJ��RI�LQFRPLQJ�RXWJRLQJ�OLQNV��������������������

5#���RI�UHFRP

PHQGHU�V\VWHP

����

���

���

���

���

���

3DJH��

H0 : The number(log) of incoming/outgoing links for resources has no relationship to the performance of a recommender system. •  in other words, the performance of the recommender system

decreases for the resources with sparsity.

Pearson’s correlation of 0.798 thus, we reject H0

20

Conclusions

•  Results show that our proposed similarity measure:

•  satisfy the fundamental axioms

•  outperforms baselines for measuring the semantic similarity between general resources

•  outperforms Sharkti on single-domain recommendations

•  Linked Data sparsity problem for LOD recommender system

•  on one hand, utilizing Linked Data to build a recommender system can mitigate the traditional sparsity problem of collaborative recommender systems, but on the other hand, the system can also have a Linked Data sparsity problem for resources in the Linked Data set that the recommender system has adopted

•  extend the current similarity measure (with longer paths)

•  investigate different normalization strategies

•  apply it to social recommendations (e.g., news recommendations in Twitter)

21

Future Work