A Method to Manage the Difference of Precision between Profiles and
Items for Recommender System
Applied Upon a News Recommender System using SVM Approach
David Werner and Christophe Cruz
LE2I Laboratory, UMR CNRS 6306, Université de Bourgogne, BP 47870, 21078 Dijon Cedex, France
Keywords: Recommender System, News, Domain Ontology, Ontologies, Knowledge Base, Indexing,
Recommendation, Vector Space Model.
Abstract: Contractors, commercial and business decision-makers need economical information to drive their
decisions. The production and distribution of a press review about French regional economic actors
represents a prospecting tool on partners and competitors for the businessman. Our goal is to propose a
customized review for each user, thus reducing the overload of useless information. Some systems for
recommending news items already exist. The usefulness of external knowledge to improve the process has
already been explained in information retrieval. The system’s knowledge base includes the domain
knowledge used during the recommendation process. Our recommender system architecture is standard, but
during the indexing task, the representations of content of each article and interests of users’ profiles created
are based on this domain knowledge. Articles and Profiles are semantically defined in the Knowledge base
via concepts, instances and relations. This paper deals with the similarity measure, a critical subtask in
recommendation systems. The Vector Space Model is a well-known model used for relevance ranking. The
problematic exposed here is the utilization of the standard VSM method with our indexing method.
1 INTRODUCTION
The decision-making process in the economic field
requires the centralization and intake of a large
amount of information. The aim is to keep abreast
with current market trends. Thus, contractors,
business men and sales persons need to continuously
be aware of the market conditions. This means to be
up-to-date regarding ongoing information and
projects undergoing development. With the help of
economic monitoring, prospects can be easily
identified, so as to establish new contracts. Our tool
is specialized in the production and distribution of
press reviews about French regional economic
actors.
The reviews sent are the same for each user, but
personalized according to a geographic area. All
articles sent do not necessarily correspond to a
person’s needs, and can be a waste of time. To
reduce the overload of useless information, we are
moving towards a customized review for each user.
Therefore, an opinion survey on magazine readers
that covers a broad array of subjects, including news
services, was undertaken. Criteria for a relevant
customization of the review were identified as a
result of this survey as well as expert domain
knowledge. These criteria are economic themes (i.e.
main economic events), economic sectors, major
transverse projects, temporal and localization data
about each element underlined. Therefore, the
complete production process of the review was
redesigned to produce and to automatically
distribute a customized review for each user.
Another drawback in the existing process is the
produced information storage. News articles are
stored as PDF file reviews (i.e. the same format that
is sent to customers, natural language), but this
unstructured format is hard to handle and reuse.
The aim of the architecture is to manage all news
produced, and provide the most relevant article for
each customer, using our domain knowledge. The
overload of news information is a particular case of
information overload, which is a well-known
problem, studied by Information Retrieval and
Recommender Systems research fields. News
recommender systems already exist (Middleton et al.
2004), (Getahun et al. 2009), (Liu et al. 2010),
465
Werner D. and Cruz C..
A Method to Manage the Difference of Precision between Profiles and Items for Recommender System - Applied Upon a News Recommender System
using SVM Approach.
DOI: 10.5220/0004373404650470
In Proceedings of the 9th International Conference on Web Information Systems and Technologies (WEBIST-2013), pages 465-470
ISBN: 978-989-8565-54-9
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
(Tanev et al. 2008) SCENE (Li et al. 2011)
NewsJunkie (Gabrilovich et al. 2004) Athena
(Ijntema et al. 2010) GroupLens (Resnick et al.
1994) News Dude (Billsus, Pazzani. 1999) et
YourNews (Brusilosky et al. 2007). Some of these
systems use domain knowledge to improve the
recommendation task (Ijntema et al. 2010),
(Middleton et al. 2004).
To achieve this goal, a content-based
recommender system is being developed. A
recommender system is necessary for the item
ranking. And content-base is required to analyze the
content of each article to structure and preserve
information content. The results of the analysis
enable to link the domain knowledge to the articles.
Because the domain knowledge can be reused to
improve the recommendation task (Ijntema et al.
2010), (Middleton et al. 2004).
This paper is organized as follows: Section 2
presents a brief review of related work. In section 3,
we outline the proposed solution of a recommender
system. Section 4 presents the similarity measure
task, and our implementation of VMS. Section 5 is
the conclusion and future work.
2 RELATED WORK
Our work is related to several works in News
recommender systems, SCENE (Li et al. 2011)
NewsJunkie (Gabrilovich et al. 2004) Athena
(Ijntema et al. 2010) News Dude (Billsus, Pazzani.
1999) et YourNews (Brusilosky et al. 2007). The
survey of K. Nagewara Rao (Nageswara, Talwar,
2008) proposed a general comparison of the main
advantages and drawbacks of each kind of
Recommender System (e.g. content based or
collaborative filtering). The advantages of content-
based recommender systems for news
recommendation are also explained in (Liu et al.
2010) to improve the Google news platform. The
main drawback of collaborative filtering systems is
the recommendation of new items. Novelty is very
important in the particular case of news; each new
article must be quickly recommended. Waiting for
enough users to have read it before recommending is
a big waste of time. Furthermore, we need to be able
to recommend for very particular user profiles, as
some customer needs are unique.
There are many systems that work without
Knowledge (Liu et al. 2010), (Billsus, Pazzani.
1999), (Resnick et al. 1994). The advantages of
using exterior knowledge for enhancing the
recommendation were exposed by Ijntema (Ijntema
et al. 2010). He uses the name semantic-based
recommender system to distinguish standard
content-based systems from the systems using
external knowledge (e.g. domain ontologies or
lexical knowledge as WordNet (Fellbaum 1998)).
Lexical knowledge is used by (Getahun et al. 2009)
and domain knowledge by (Middleton et al. 2004).
Athena uses both (Ijntema et al. 2010). Ontologies
used by these systems already exist or are created by
hand, and maintained. Unlike previous systems, the
Knowledge base containing domain Knowledge is
used as an index for articles and profiles, as it is
explained in section 3. To compare profiles and
articles, classic VSM is not directly usable, so we
have adapted it, as presented in section 4.
3 IMPLEMENTATION
Our system is an ontology drive content based
recommender system. An ontology schema is
created and populated with the help of company
experts, in order to model the domain knowledge in
a knowledge base. In a classic content-based
recommender system, we distinguish two main
tasks. The first is indexing. The task is to create a
representation of the users’ needs, and item content.
The Knowledge base will be populated during this
task. The quality of content analysis is important for
the knowledge base population and for indexing, so
our system is semi-supervised by an expert. The
second task is comparison. This task is the
comparison with item representation so as to
measure the degree of relevance for each profile.
Items are ranked with the help of the similarity
measure, after being provided to the user. These
subjects are developed as follows.
3.1 The Knowledge Base
The knowledge base used for this system is
composed of several ontologies (Werner et al. 2012).
An Upper level ontology is used to manage
information shared by all application areas (in the
case of an extension of the system to new fields of
application). High level concepts like location,
geospatial information, temporality, events, agents,
etc. Domain ontology is used to manage domain-
specific knowledge. Concepts of this ontology are
mainly specialization of concepts from the upper
level. Other ontologies are used to manage articles,
profiles, and lexical resources used by a gazetteer.
The lexical resource ontology is based on the
ontology PROTON used on the KIM platform
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
466
(Popov et al. 2003). In this paper the knowledge
base model defined by Ehrig et al. (Ehrig et al.
2004) and based on the Karlsruhe Ontology
(SOURCE) Model is used.
Definition 1. Ontology:
Ο,,
,
, ,
,
,
,
Wherein,,,are disjoint sets of concepts,
data types, relations and
attributes,
,
,
,
are hierarchy of classes,
data types, relations and attributes, and
,
are
function that provide the signature for each
: →
attribute and
: → relation.
Definition 2. Knowledge Base:
,
,
,
,,,
,
,
,
Wherein
,
,
,
,, are disjoint sets
of concepts, data types, relation attributes, instances
and data values.
is the classes instantiation
function
:
→2
.
is the data type
instantiation function
:
→2
.
is the relation
instantiation function
:
→2
.
is the
attribute instantiation function
:
→2
.
3.2 Indexing
To archive the recommendation of articles to
customers, the system needs are a representation of
the content of each article, and representation of the
needs of each customer. The index used in our
system is the same for articles and profiles. The
knowledge base used has an index. Articles and
profiles are represented by instances in our
knowledge base. Some relations in the system
ontologies are used to model of article content, and
users’ interests.
3.2.1 Article Indexing
The ambition is to create a machine understandable
representation of the content of each article, so as to
compare with profiles. The unstructured information
contained in articles is analyzed. Two kinds of
information can be distinguished: explicit pieces of
information (e.g. places, persons, organizations and
so on) and implicit pieces of information (e.g. the
theme of each article). In the system, the theme is
one criterion, corresponding to the main event
related by the article. Company experts have
predefined a hierarchical list of themes wherein each
article must be classified. The tasks of information
extraction, annotation, indexing are done with the
help of the GATE platform (Cunningham 2002). A
web interface was developed. It enables the writers
to create articles. Results of the analysis are
presented in it. They can be validated/corrected by
the writer.
Analysis
Some post processes are applied into articles, such
as tokenizers, sentence splitters, POS taggers,
gazetteers (which use the knowledge base lexical
resources as a dictionary) before JAPE patterns
matching engine. In the first prototype, we used
handmade lexico-semantic patterns. The aim is to
extract important entities (explicit pieces of
information like Persons, Organizations, places and
dates). Results of analysis are hand-checked,
corrected and validated. Implicit pieces of
information are specifically handmade.
Population
For each article analyzed, an instance of the concept
article is created in the knowledge base, so as to
represent the article. Automatic analysis, correction
done by hand and specifications of not automatically
funded criteria are used to characterize the content
of each article. Relations are created in the
knowledge base between the article’s instance, and
criterion’ instances (result of the analysis). Instances
of criteria and relations with the instances of articles
permit to index the article and create a semantic
representation understandable by the machine.
3.2.2 Profiles Indexing
In the company, sellers are in charge of
understanding the needs of each customer. Several
phone calls are necessary so as to acquire future
customers. During the phone call the seller proposes
a free trial period. This helps to create a first
handmade profile for each customer by an expert,
and avoids the problem of a cold start, common to
all content-based recommender systems.
The profile indexing process is the same as the
articles. A profile instance is created in the
knowledge base. Relations are created between the
profile instance and criteria instances. A web
interface was developed to enable sellers to define /
change the user profile. The choices are reflected in
the Knowledge base that permits to index the profile,
and create a semantic representation understandable
by the machine.
3.3 Recommendation
The recommendation task is mainly based on the
AMethodtoManagetheDifferenceofPrecisionbetweenProfilesandItemsforRecommenderSystem-AppliedUpona
NewsRecommenderSystemusingSVMApproach
467
comparison between the profile and available items.
The knowledge base is used by the system like an
index, and profiles and articles are presented by a set
of instances and relations. For each article validated
by writers, the full content is inserted in the database
and a representation of the article is created in the
knowledge base. An instance of the concept article
and an instance of each relation “isAbout” is created
to link the article with its criteria. For each profile
made by the sellers, the representation is created in
the knowledge base. An instance of the concept
profile is created, and each “isInterestedIn” relation
between the profile instance and its criteria are
instanced. We present the comparison method used
in the following section.
4 COMPARISON
The comparison task enables to evaluate the
pertinence of an article for a profile, via a similarity
measure between them.
Definition 3 similarity: SIM
x, y
:II→0,1 is a
function to measure the degree of similarity between
x and . The similarity function can satisfy some
properties:
∀x, y ∈ ISIM
x, y
0Positiveness
∀x, y ∈ ISIM
x, x
0 Reflexivity
The last one is the symmetry,∀x, y ∈
ISIM
x, y
SIM
y, x
but in our context we want
an asymmetric function, because we consider that
comparing profiles and articles is not the same as
comparing two articles.
4.1 VSM
An approach based on the vector space model
(Salton 1970) was used in the prototype. Articles
and profiles are represented by vectors on a space
wherein each dimension is a potential instance of
criteria. Several methods can be used to compare
vectors; the most common is the cosine similarity.
SIM
,
cos
.
|
|
|
|
(1)
An article can be defined like a vector of
instances of Entities and criteria. For the
recommendation task in the prototype only instances
of criteria are used.
,
,…
Wherein, an article is represented by a vector
composed by a set of instances
.
∈
Instances
are instances of concepts belonging
to the set of concepts C
defined by indexing criteria.
The set
contains all instances of all concepts in the
setC
.
⊆
The set
is a subset of the set of all instances
of the Knowledge Base.
A profile can be defined as a vector of instances
of criteria.
,
,…
Where, a profile is represented by a vector
composed of a set of instances
.
In our implementation, one vector for each
criterion is used. This enables to weigh or use
different similarity measures (e.g. cosine, jacquard,
Euclidian) for each criterion. For example, location,
theme and sectors are much more important than
project and organization and so highly weighed.
SIMF
,
∑
SIM
,
∑
(2)
SIM
,
Is the similarity between profile
and article
and
the coefficient for a specific
criterion. ∀
,
∈
,
,
,
,…
,
And
,
,
,
,…
,
. One or more concepts are
defined for each criterion.
Methods from the information retrieval fields can
be used to enhance the recommendation. One of the
first systems using external knowledge to improve
the understanding of user needs is (Voorhees 1994).
The Voorhees approach used WordNet (Fellbaum
1998) to provide a query expansion. We can
translate this kind of method to recommender
systems, instead of query expansion; we can name
this method ‘profile expansion’. Middleton
(Middleton et al. 2004) uses this method without
naming it. Ijntema (Ijntema et al. 2010) also uses it,
but unlike Middleton, he uses more powerful
ontology relations (not just is_a) to expand the user
profile. In our system, the profile expansion takes
the form of adding instances, e.g. if the user’ U
profile shows an interest for company Co and in the
knowledge base a symmetric relation like
is_aSubsidiaryOf is instanced with another company
Co’, it is possible to add the other company to his
profile. The Expanded VSM is developed in the
following section.
4.2 Expanded Vectors
The 4.1 section presents an implementation of the
similarity measure between two instances, by the
creation of vectors of related instances. It was
explained by Voorhees (Voorhees 1994) in the VSM
that all dimensions are orthogonal and so, all
elements of each vector are considered as
independent. That is not really the case for the
lexical used by Voorhees, and instances used in our
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
468
system. In Voorhees’s method, the metronomic and
synonymic relations defined in WorldNet are used to
add related lexical to the vector. In our case,
relations between individuals exist, because we used
a Knowledge base to manage the domain
knowledge, and our criteria are hierarchically
defined in it.
Meronomic relations exist between instances of
Location. With Voorhees’s method, it seems logical
to expand the profile (the query for Voorhees) with
all sub instances enclosed by the most general
instance. For example, if the profile is interested in
Bourgogne (Region, biggest Administrative
division), it seems logical to add Cote d’Or and
Yonne (Departement, sub administration division of
Region) which are two departments within the
Bourgogne region, and Dijon, Beaune, Chenove,
Auxerre, which are cities within these departments.
But, in the real case with this method, it is necessary
to add four departments and its 2047 cities to the
vector. The similarity between a profile interested in
Bourgogne and an article about Dijon will be very
low with this method.
So our method is to expand profile and articles,
not only the profile and limiting the size of vector,
instances added are including instances, not included
(by meronomic relations). So our method is
analogue to graph-based methods which search the
common ancestor. For example, if the profile is
interested by Dijon, Cote d’Or and Bourgogne, these
are added to the vector. But if the profile is
interested just by Bourgogne, nothing is added.
With this method the first drawback is solved,
synonymic and meronomic relations between
instances are managed, but the second still is not.
Our similarity function is symmetric, but we want to
compare a profile and an article to different things.
The precision of an article must not have the same
consequences as the precision of a profile. This
problem is the subject of the following section.
4.3 Expanded Vectors
The previous section looked at how to take into
account the relations between instances in the VSM.
This section is about the managing of the difference
of precision between profiles and articles. To solve
this problem we used an intermediate vector for each
criterion, a sub vector
composed of common
instances of the article
and profile
vectors for
the criteria.
Definition 4. Precision: In the hierarchy of concepts,
more generals concepts enclosed more specifics one.
In the hierarchy of instances it is the same. Instances
Figure 1: Examples of profiles and articles for one
criterion.
from the top of the hierarchy are less specific than
instances from the bottom.
If the Article is about an instance form the
bottom and the Profile is interested in top instance
(of the same branch), the similarity must be higher
than if the Article is about a top instance and the
Profile is inserted in an instance from the bottom of
the hierarchy because there is a loss of precision if
the profile is more specific than the article, so the
article is less relevant.
,
∩
,
is the sub set of common elements of the set
of instances related to the profile
,
and the article
,
. ∀
,
∈S
,
,
,
,…
,
The vector
is composed by elements of the set
.
,
,
,
,
,
,
,
(3)
It is possible to weigh differently the precision of
the profile and the precision of the article with this
method. In our implementation we used ′
,
1
and ′
,
4, because we consider that the
difference of precision of the profile mustn’t
influence the final note over 20%. However, it is
possible to change the value, and it is also possible
to use different values according to the criterion.
SIMF
,
∑
,
∑
(4)
The final similaritySIMF
,
value is the sum
of the similarity measure for each criterion. This
measure is used for the ranking of articles proposed
to the user according to his profile.
With the method the similarity value for the case
A.2 (figure 1.) is higher than the case A.1, because
in the case A.2, a and p have more common
ancestors than in the case A.2. Moreover, the cases
B.1 and B.2 figure the problem of precision. With
our asymmetric method the value of similarity
AMethodtoManagetheDifferenceofPrecisionbetweenProfilesandItemsforRecommenderSystem-AppliedUpona
NewsRecommenderSystemusingSVMApproach
469
between a and p in the case B.1 is higher than in the
case B.2 because the user needs are specific and the
article information (about this criterion) very general
relative to the user needs. So the article is less
relevant for the user. The following section presents
the conclusion and future work.
5 CONCLUSIONS
In this paper we have presented the adaptation of a
standard VSM recommender system to our specific
method of indexing (e.g. articles and profiles are
semantically defined in the knowledge base via
relations with the domain knowledge already
defined in it). We first presented the context, our
goal, and the existing approaches, then we explained
the architecture in detail. Finally we explained the
specific task of comparison that we adapted to our
case.
ACKNOWLEDGEMENTS
This project is founded by the company Actualis
SARL and the financing CIFRE research grant from
the French agency ANRT.
REFERENCES
Billsus, D., Pazzani, M.J., 1999. A Personal News Agent
that Talks, Learns and Explains. In: The Third Annual
Conference on Autonomous Agents, ACM, pp. 268–
275.
Ahn, J., Brusilovsky, P., Grady, J., He, D., Syn, S.Y.,
2007. Open User Profiles for Adaptive News Systems:
Help or Harm? In: 16th international conference on
World Wide Web, ACM, pp. 11–20
Cunningham, H., 2002 GATE, A General Architecture for
Text Engineering. Computers and the Humanities 36
pp. 223–254
Ehrig, M., Haase, P., Stojanovic, N., Hefke, M., 2005
Similarity for Ontologies A Comprehensive
Framework. ECIS. Regensburg, Germany.
Fellbaum, C., ed., 1998. WordNet: An Electronic Lexical
Database. MIT Press, Cambridge, MA
Gabrilovich, E., Dumais, S., Horvitz, E., 2004.
Newsjunkie: providing personalized newsfeeds via
analysis of information novelty. In WWW, pp. 482–
490.
Getahun, F., Tekli, J., Richard, C., Viviani, M.,
Yetongnon, K., 2009. Relating RSS News/Items. In:
9th International Conference on Web Engineering,
Springer, pp. 442–452
IJntema, W., Goossen, F., Frasincar, F., Hogenboom, F.,
2010 Ontology-Based News Recommendation. In:
International Workshop on Business intelligence and
the Web, BEWEB, EDBT/ICDT Workshops. pp. 16:1
16:6. ACM, New York, USA
Li, L., Wang, D., Li, T., Knox, D., Padmanabhan, B.,
2011. SCENE: A scalable two-stage personalized
news recommendation system. In Proc. the 34th
Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval,
Beijing, China, Jul. 25-29, 2011, pp.124-134.
Liu, J., Dolan, P., Pedersen, E.R., 2010. Personalized news
recommendation based on click behavior. In
Proceedings of the 15th International Conference on
Intelligent User Interfaces, pp. 31–40. ACM.
Middleton, S.E., Shadbolt, N.R., Roure, D.C.D., 2004.
Ontological User Profiling in Recommender Systems.
ACM Transactions on Information Systems 22, pp.
54–88
Nageswara Rao, K., and Talwar, V.G., 2008. Application
Domain and Functional Classification of
Recommender Systems A Survey, Journal of Library
& Information Technology, Vol. 28, No. 3: 17-35.
Piskorski, J., Tanev, H., Wennerberg, P.O., 2007.
Extracting Violent Events From OnLine News for
Ontology Population. In: 10th International
Conference on Business Information Systems, BIS.
Lecture Notes in Computer Science, vol. 4439, pp.
287-300. Springer-Verlag Berlin Heidelberg.
Popov, B., Kiryakov, A., Kirilov, A., Manov, D.,
Ognyanoff, D., Goranov, M., 2003. KIM Semantic
Annotation Platform.
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P.,
Riedl, J., 1994. GroupLens: An open architecture for
collaborative filtering of netnews. In Proceedings of
the 1994 ACM Conference on Computer Supported
Cooperative Work, Chapel Hill, North Carolina,
United States. ACM Press, New York, pp. 175-186.
Salton, G., 1970. The SMART retrieval system :
Experiments in automatic document processing.
Prentice Hall
.
Stumme, G., Ehrig, M., Handschuh, S., Hotho, A.,
Maedche, A., Motik, B., Oberle, D., Schmitz, C.,
Staab, S., Stojanovic, L., Stojanovic, N., Studer, R.,
Sure, Y., Volz, R., Zacharias, V., 2003. The Karlsruhe
view on ontologies. Technical report, University of
Karlsruhe, Institute AIFB.
Tanev, H., Piskorski, J., Atkinson, M., 2008. Real-Time
News Event Extraction for Global Crisis Monitoring.
In: 13th International Conference on Natural
Language and Information Systems: Applications of
Natural Language to Information Systems, NLDB.
Lecture Notes in Computer Science, vol. 5039, pp.
207-218. Springer-Verlag Berlin Heidelberg
Voorhees, E., 1994. Query Expansion using Lexical-
Semantic Relations.
Werner, D., Cruz, C., Nicolle, C., 2012. Ontology-based
Recommender system of economic articles; In
Proceedings of the 2012 Webist Conference, Porto,
Portugal.
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
470
