Untitled Lesson

Authors

AHMED MOUNAF MAHDI Prof. Dr. ASAAD SABAH HADI

College of Information Technology

University of Babylon

Linked Data-based Recommender System Using

Hybrid Semantic Similarity Measure

1

Word Cloud

ما هي البرامج التي تستخدمها والتي تحتوي على نظام توصية

Introduction

 LOD is a community project aiming to provide free and

public data. LOD datasets are available on the web
under an open license in machine readable format
(RDF triples) and linked together with unique URIs

 RDF(Resource Description Framework) offer a graph-

based

data

model

for

describing

resources

and

identifying relationships with other resources

 Vast amount of RDF data have been published in

freely accessible datasets connected with each other
to form the so called LOD cloud.

2

Introduction (Cont.)

 The knowledge encoded in LOD(Linked Open Data)

cloud can assist to tackle with problem when few or no
features can describe the items to be recommended

 The main idea of the proposed system is designing a

Linked data-based recommendation system that use
semantic features extracted from two relevant datasets in
LOD.

3

1301 datasets

with 16283 links May 2021

lod-cloud.net
4

Problem Statement (Cont.)

5

 Most previous research depend on pure link-based similarity measures

which will degrade the accuracy of recommendation if there are
insufficient links between the resources even though they share many
features.

 Furthermore, the values of the features in LOD cloud are stored as

literals and non-literal and some approaches consider the lexical content
such as the literals of the items to be as features of the items. However,
literals may not be properly formatted or have invalid values that affect
the quality of the selected features For example; the object value could
be a specific date like 01-01-1550^^xsd:date or string value like “16th
century”[5].

 Some features such as; actor, genre, director or writer are missing in

one dataset and available in other.

 However, sometimes the links between the same resources in two

datasets are missing and the generating of such links is necessary to
have all the features from the linked datasets.

Objectives

6


Design a Linked-data-based recommendation system that depends on the
Linked data and utilizes the transversal relations exist in LinkedMDB and
DBpedia datasets as well as the hierarchical structure exist in DBpedia in
order to find the candidate resources for the recommendations from movie
domain and other related domains (Book).


Propose an Extended LinkedMDB dataset that includes extra same as links
for fixing the missing links between the similar movies in LinkedMDB and
DBpedia using the Silk framework and also extra links from the movies to the
related books in BNB Library dataset will be included.


Employs Word Embedding to represent the item's non-literal features and
generate a vector for each resource, as well as FOAF vocabulary to represent
the user's preferences by analyzing the user's input and fetching the features
required from LinkedMDB and DBpedia.


Conduct a several experiments to evaluate the proposed system using offline
evaluation method with MovieLens 1M dataset and calculate the effectiveness
using Hit rate measure.

Methodology

7

The proposed system is composed of six major phases

1.

Data Preprocessing

2.

Access Endpoints, Features
Extraction, and Vectorization

3.

Building a User Model

4.

Generation of Candidates List

5.

Calculation of Semantic Distance for
Ranking

6.

present Top-N items and related
books

Data Preprocessing - Datasets

8

 LinkedMDB

 DBpedia

 The movies in these two datasets are interlinked with

SameAs relations

 some cases when some movies are without interlinking

relations.

 BNB Library dataset is used, which consists of

metadata about books and different kinds of
publications to recommend a related book from this
dataset.

Data Preprocessing - Interlinking

9

 We use the Silk framework (Isele et al., 2010) to enrich the

LinkedMDB with more SameAs links to DBpedia.

 If the movie's labels in the two datasets are exactly matched,

the URI of the movie in LinkedMDB will be linked to the URI of
the movie in DBpedia with SameAs relation.

 Number of sameAs links to DBpedia is 30354

 We generate 40142 links

10

 The first step in this stage is extracting the features of the input

invoked by the user and create a file of interests for the user. The
second step is building a user profile based on FOAF vocabularies
and using the features extracted in the first step. Therefore, we will get a
profile that can be used in RS task, and queried in efficient way using
semantic query language (SPARQL).

 Examples of The Terminator movie features

Building a User Model

Feature Name

URI

Dataset

Actor /Starring

https://triplydb.com/Triply/linkedmdb/id/actor/13309

LinkedMDB

http://dbpedia.org/resource/Linda_Hamilton

DBpedia

Director

https://triplydb.com/Triply/linkedmdb/id/director/5617

LinkedMDB

http://dbpedia.org/resource/James_Cameron

DBpedia

Genre /Subject

https://triplydb.com/Triply/linkedmdb/id/film_genre/23

LinkedMDB

http://dbpedia.org/resource/Category:American_action_films

DBpedia

Writer

https://triplydb.com/Triply/linkedmdb/id/writer/8820

LinkedMDB

http://dbpedia.org/resource/James_Cameron

DBpedia

11

 interest property, which is

considered as a social web
term in FOAF vocabularies,
in order to populate it with
the features extracted in the
previous

step

that

are

interested by the user.

 Therefore, the output of this

step will be an RDF file that
can

be

used

by

the

recommendation

system

and query it using the same
known

query

language

(SPARQL).

Building a User Model

Data Preprocessing - Access Endpoints and Features Extraction

12

 LinkedMDB has a dump file that we uploaded to

local Virtuoso Server in addition to, the resulted file
from interlinking process while we used remote
endpoint for DBpedia

 SPARQL query on Extended LinkedMDB to get the

URIs of the movie from both datasets
 extract all the features from both datasets
 each movie will be represented by multi lines
 the features include actors, genre, writer and director

URIs

 The

features

of

the

resources

must

be

represented numerically to calculate the similarity.

 To vectorize the features we utilized Word2Vec

which

uses

neural

network

to

learn

word

association.

13

Data Preprocessing - Vectorization

Word2Vec Algorithms Mikolov, Tomas; et al. (2013)

14

 Word2vec is a particularly computationally-efficient two-

layer neural net model for learning word embeddings from
raw text.

 There are two different algorithms, the Continuous Bag of-

Words model (CBOW) and the Skip-gram model.

 In the CBOW model, the distributed representations of

context (or surrounding words) are combined to predict the
word in the middle.

 While in the Skip-gram model, the distributed

representation of the input word is used to predict the
context.

 CBOW is faster while skip-gram does a better job for

infrequent words.

Word2Vec Algorithms

15

 Once the training is finished, all words are projected into a

lower-dimensional feature space, and semantically similar
words (or entities) are positioned close to each other.

Word2Vec Algorithms – hyper parameters

16

 “hyper-parameter selection is a task specific decision”

(Mikilov et al., 2013)

 size: The number of dimensions of the embeddings

 the default is 100.

 We set this parameter to 50

 window: The maximum distance between a target

word and words around the target word. The default
window is 5.

 We set this parameter to 2

Recommendation Layer

The aim of this layer is to discover the related

resources to a given resource by analyzing the
semantic relationships.

 Recommendation of movies that are linked

directly

 Recommendation of movies that have common

direct or indirect categories.

 Recommendation of items that are linked to the

same movie features.
17

Methodology- Ranking

 This stage mainly ranks candidate resources

obtained in the previous layer, based on semantic
similarity measures. The candidate resources
are sorted according to values of a semantic
similarity measure, which measures the similarity
between the initial resource and each one of
these candidate resources.

18

Hybrid Semantic Similarity Measure

 Linked Data Semantic Distance (LDSD)

Passant,2010

 Cosine distance of vectors generated by

Word2vec algorithm is implemented

Simi,j = p * LDSDi,j+(1-p) * Cosi,j

A trial and error experiments are done by varying
the value of (p), which is the amount of impact of

each distance measure from 0 to 1

19

Evaluation

 We use offline evaluation

 leave one out protocol is used and the MovieLens

dataset is divided into a training and testing set

 Hit Rate measure, the proportion of items in the test

set that appeared in the top-N recommended items
returned for each user

Hit rate=

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 ℎ𝑖𝑡𝑠

𝑛

20

Results

21

p=0

p=0.25

p=0.50

p=0.75

p=1

Hit
rate
0.20

0.70

0.179

0.177

0.13

Table. 1 Hit rate with different P values

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

P=0.25

P=0.50

P=0.75

Hit rate comparisons of the proposed

measure with different P values

0

0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8

Proposed
Measure

LDSD

Cos

Hit rate comparisons of LDSD,

cosine and the proposed measure

Conclusion & Future Works

22


This work contributes to overcome the common issues in recommendation system and
enhance the accuracy of recommendation results.


Firstly, An extended version of LinkedMDB dataset is proposed that includes extra SameAs
links to DBpedia. This will help other researchers in different kinds of applications who need
to utilize the features of the movies and to find the corresponding DBpedia URI from the
LinkedMDB URI.


Secondly, the proposed system utilized LOD and exploited two datasets, DBpedia and
LinkedMDB,


Thirdly, the system recommend items from other domain (books) that are related to the
recommended top-N movies.


Finally, the proposed approach combines LDSD with cosine measure to recommend top-N
movies to users and the presented results were better than using the LDSD or cosine
separately.


In addition, we have found that giving a higher probability to cosine measure over LDSD
can enhance the accuracy of the hit rate measure


Various experiments will be conducted with other evaluation metrics such as Diversity and
Novelty. Also, other datasets from other domains such as books or music will be used as a
future work.

References

23

1.A. Passant, "Measuring Semantic Distance on Linking Data and Using it for Resources Recommendations," in AAAI spring
symposium: linked data meets artificial intelligence, 2010, vol. 77, p. 123.

2.C. C. Aggarwal, "Knowledge-based recommender systems," in Recommender systems: Springer, 2016, pp. 167-197.

3.C. Figueroa et al., "Executing, Comparing, and Reusing Linked-Data-Based Recommendation Algorithms with the Allied
Framework," in Semantic Web Science and Real-World Applications: IGI Global, 2019, pp. 18-47.

4.C. Musto, P. Lops, M. de Gemmis, and G. Semeraro, "Semantics-aware Recommender Systems exploiting Linked Open Data and
graph-based features," Knowledge-Based Systems, vol. 136, pp. 1-14, 2017.

5.F. Komeiha, N. Cheniki, Y. Sam, A. Jaber, N. Messai, and T. Devogele, "LDS: Java Library for Linked Open Data Based Similarity
Measures," in 2020 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT),
2020: IEEE, pp. 476-481.

6.G. O. M. D. Silva, P. R. D. Souza, and F. A. Durão, "HSLD: a hybrid similarity measure for linked data resources," International
Journal of Metadata, Semantics and Ontologies, vol. 14, no. 1, pp. 16-25, 2020.

7.I. Vagliano, "Content Recommendation Through Linked Data," PhD, Computer and Control Engineering, Politecnico di Torino, 2017.

8.Isele, R., Jentzsch, A., & Bizer, C. (2010). Silk Server-Adding missing Links while consuming Linked Data. COLD, 665.

9.J. Rosati, P. Ristoski, T. Di Noia, R. d. Leone, and H. Paulheim, "RDF graph embeddings for content-based recommender systems,"
in CEUR workshop proceedings, 2016, vol. 1673: RWTH, pp. 23-30.

10. R. Meymandpour and J. G. Davis, "A semantic similarity measure for linked data: An information content-based approach,"

Knowledge-Based Systems, vol. 109, pp. 276-293, 2016.

11. S. Alfarhood, K. Labille, and S. Gauch, "PLDSD: propagated linked data semantic distance," in 2017 IEEE 26th International

Conference on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE), 2017: IEEE, pp. 278-283.

12. S. Alfarhood, S. Gauch, and K. Labille, "Semantic Distance Spreading Across Entities in Linked Open Data," Information, vol. 10,

no. 1, p. 15, 2019.

13. T. Mikolov, K. Chen, G. Corrado, and J. Dean, "Efficient estimation of word representations in vector space," arXiv preprint

arXiv:1301.3781, 2013.

Thank You

24