Towards a Recommender System from Semantic Traces
for Decision Aid
Ning Wang
1
, Marie-Hélène Abel
1
, Jean-Paul Barthès
1
and Elsa Negre
2
1
UMR CNRS 7253 HEUDIASYC, University of Technology of Compiègne, Rue Roger Couttolenc, Compiègne, France
2
UMR CNRS 7243 LAMSADE, Paris-Dauphine University, Place du Maréchal de Lattre de Tassigny, Paris, France
Keywords: Traces of Interaction, Recommender System, Semantic Modelling.
Abstract: Collaboration allows integrating intellectual resources and knowledge from all participants in order to
achieve individual or collective goals. With the help of informational environments, we can better organize,
realize and record collaboration. During interactions among users in such environments, each activity
produces a set of traces. Such traces are recorded and classified, based on a model of traces and can be
exploited to improve collaboration. In this article, we propose a semantic model of traces and analyze
classified traces by means of TF-IDF. We exploit the results to offer users recommendations and decision
aid.
1 INTRODUCTION
As an important form of sharing and exchanging
information, collaboration is one of the sources of
power for human society development and progress
(Grudin, 1994). Thanks to information technology,
nowadays collaboration can be organized and
managed in an informational environment. In such
an environment, users achieve their purposes by
realizing different actions. We are interested in the
results of actions as well as in the actions
themselves. For example, not only the content of a
document a user has created matters, but also
knowing who shares and consults this document at
what time, with what kind of frequency, etc. A set of
actions, step by step, is defined as a trace (Zarka et
al., 2011). After proper modelling and analysis,
traces could in return help to indicate the strengths
and weaknesses of an individual or of a group of
users (Tomaz et al., 2011). Thus, with the
information exploited behind traces, we can improve
collaboration, as mentioned by works focusing on
the reuse of traces for different purposes such as
decision aid (García-Crespo et al., 2011), or
recommendation (Chang et al., 2013).
In this paper we propose a mechanism that
models, records and analyzes users’ traces and in
return recommends and helps users making
decisions which are personalized for either personal
purpose or for the entire group. The following tasks
are needed to achieve this objective: (i) propose a
semantic structure to record the traces, (ii) evaluate
the traces using TF-IDF and a semantic distance
among the actions, (iii) give recommendations and
provide some decision aid accordingly.
The remainder of this paper is organized as
follows. We identify various limitations of the
current studies on recommendation methods in
Section 2. In Section 3 we propose a kind of
recommender system for the need of exploiting the
traces. In Section 4 we illustrate our method by
giving a toy example. Section 5 provides
conclusions and mentions directions for future
works.
2 RELATED WORKS
Aiming at a better treatment and exploitation of
traces, we need to analyze different types of traces
as well as the structure of collaboration in a
Collaborative Working Environment (CWE). The
model of traces proposed by Li (2013) allows an
elaborated analysis of interactions among users. It
pays special attention to the exchanges of users in
informational environment. Traces then can be
exploited to feed a recommender system. Interests of
recommender systems are justified by the need to
274
Wang N., Abel M., Barthès J. and Negre E..
Towards a Recommender System from Semantic Traces for Decision Aid.
DOI: 10.5220/0005133502740279
In Proceedings of the International Conference on Knowledge Management and Information Sharing (KMIS-2014), pages 274-279
ISBN: 978-989-758-050-5
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
manage the growing amount of information
(Adomavicius 2005).
Recently, various articles were published aiming
at exploiting the traces with the help of semantics.
Chen et al (2013) present a mechanism for
personalized knowledge search and recommendation
adapting a suitable domain ontology according to the
previous browsing and reading behavior of users.
Vesin et al (2012) present a new approach for
performing effective personalization based on
semantic web technologies. Although more and
more attention is focused on exploiting implicit
information behind data (data mining), these recent
studies did not work much on modelling users’
actions. In fact a good model of organizing and
storing users’ actions can make future exploitation
of traces more efficient.
3 OUR APPROACH
Figure 1 shows the structure of our proposed
recommender system. First of all, with the model of
actions we propose, users’ actions are collected and
modelled from an interactive platform. After being
sifted by the filter of classification, we obtain
classified traces, which allows a preliminary
presentation back to the users. Alternatively, we
apply an algorithm to calculate an index indicating
the correlation between the classified traces of a
certain user and a subject. These values can lead to
useful information that are presented as personalized
recommendation, either to a group defined as a set
of users in the platform, or to an individual user.
Figure 1: The structure of our proposed recommender
system for the exploitation of semantic traces.
3.1 Modelling of Actions
We define the principal concepts as follows:
Action: an interaction or an act performed by
a user in a collaborative environment, e.g.
sending a document to other users;
Classified trace: a set of actions performed by a
user in the informational environment
classified according to the model of traces (Li,
2013);
Set of traces: a set of classified traces.
According to our definition, an action is the most
basic element forming a trace. As regarded as an
important resource for our recommender system, we
introduce the Resource Description Framework
(RDF) to model actions (Antoniou and Van, 2004).
RDF is used as a general method for conceptual
description or modelling of information that is
implemented in web resources. Figure 2 shows the
basic structure in the RDFS graph of our model. An
ellipse represents a class of resources and a rectangle
represents an object property. For example, a person
has the object property of “has_id_person” and the
range of this property is the class called “id”. In
Figure 3, the class “creation” inherits all properties
from “action”.
Figure 2: Basic structure in RDFS graph presenting an
action.
We apply the model of actions to a web-based
collaborative platform E-MEMORAe2.0 (Abel and
Leblanc, 2009). Details are shown in Figure 3.
This model of actions has two main advantages
compared to a traditional form of history or log of
users:
Actions are presented in a labeled, directed
multi-graph. In our model the actions are
represented as resources in the RDF schema
and they are no longer discrete but are
connected by properties. This allows a better
structure of storage and usage of actions. For
example, a person “Ala” chats with “Ning”.
This action can be presented by a RDF instance
showed in the lower part of Figure 3 where
“Ala” and “Ning” are two instances of the class
of resource “person”. “Chat_1” is an instance
of the class of resource “conversation” which is
linked to the action “creation”.
Interface
of Interactive
Platform
User
Filter of
classification
Classified
traces
User User
TF-IDF
Ontology of
application
Action
Recommendations
Presentation of
classified traces
Pre sentation of
Recommendations
Model
of
actions
TowardsaRecommenderSystemfromSemanticTracesforDecisionAid
275
Figure 3: The model of actions in the platform E-Memorae 2.0.
Normally different types of actions have
different importance. For example creating a
piece of Wiki is more important than
consulting it. In our model the actions are all
classified by three classes: creation,
consultation and addition which enables us to
treat different types of actions more efficiently.
3.2 Application of TF-IDF
In order to evaluate the importance of different
traces, we apply TF-IDF as the method of
evaluation. TF-IDF (Term Frequency-Inverse
Document Frequency) (Jones, 1972) is a statistical
method for weighting usually used for the research
of information from texts. This method measures the
correlation of a term in presenting a document from
Figure 4: Relation between concepts of the method TF-
IDF and the method adapted to our case.
a corpus. The weight of a term is proportionally
higher when it appears more frequently in a
document. It also varies with the frequency of the
word in the corpus. The values are used for
evaluating the reference of a document.
Here we adapt this method to our case. Figure 4
illustrates the key-concepts of the method TF-IDF
and those more relevant in our case. Typically TF-
IDF focuses on the relation between words,
documents and corpus. If a word appears more in a
document and at the same time appears less in the
other documents of the same corpus, it better
represents this document. For our research, we are
interested in evaluating the correlation between a
trace of a given user and a certain subject. We
propose to consider that if the actions of a user are
more pertinent concerning a subject, the user has
more knowledge about it. So we are able to
recommend this user as an expert in this domain. In
our case, we study the relation between actions,
traces and the set of traces in a group of users
working on the same subject.
We adapt the equation of TF-IDF:
,
,
∑
,
(1)
person do action
creation
domaine range
consultation
addition
resource
has_id_
action
has_id_
person
id
range
Ala Creation_1 Chat_1
subClassof
...
subClassof
Creats
_question
conversation document
wiki
question
...
...
annotationmeeting
domaine
has_subject subjectrangedomaine
has_date
date
range
domaine
domaine
range
model
knowledg
e
base
note
creats
_wiki
...
...
consults
_resource
adds
_resource
...
...
...
...
creats
_meeting
...
adds
_annotation
domaine
range
...
creats
_note
...
...
creats
_conversation
domaine
range
involves
_conversation
range
domaine
Ning
subClassof
... ...
...
... ... subClassof
has_space
space
domaine
range
: class
: property
Trace:
do
creats
_conversation
involves
_conversation
KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing
276
log
|
|
|
:
∈
|
(2)
Where:
,
: the number of actions concerning the
subject i performed by the user j;
∑
n
,
: the number of actions concerning all k
subjects performed by the user j;
|
|
: the number of users in a group;
|
p:t
∈p
|
: the number of users in a group
who have performed at least one action about
the subject i.
The index of TF-IDF, C
ij
indicating the
competence of user j on subject i, which can be
regarded as the relevance between a subject and a
user, is defined as follows:
,
,
(3)
3.3 Application of Semantics
When we evaluate the competence of a user on a
certain subject, we can also take into consideration
the semantic relationship among subjects. As
semantics indicates the real meaning, a smaller
semantic distance between two concepts means a
closer relation of them. For example if we need to
find an expert on "Computer", even though a person
hasn’t left a weighted trace on “Computer” directly,
he may have contributed to the subject "Tablet"
which is quite close in a semantic view. So we can
evaluate the weight of this user on the subject
"Computer" by evaluating it on the subject “Tablet”.
We explain in detail in Section 4 by giving a toy
example.
3.4 Classified Recommender System
We take the model of traces (Li, 2013) as the basis
to carry out classified recommendations. According
to this model, a trace is classified into 4 types:
Private Trace, Collaborative Trace, Collective Trace
and Personal Trace. A private trace is sent and
received by the same user. A collaborative trace has
one sender and at least two receivers. A collective
trace has many senders and receivers while a
personal trace is defined as having only one sender
with no limit on its receivers.
Our recommender system takes into consideration a
semantic model of the system along with the traces
of recorded interactions (e.g. who has shared a
document concerning the subject S with whom, e.g.,
who usually interacts with the expert John?). It aims
at realizing recommendations for a group (improve
the collaboration, identify risks, opportunities of a
set of users from a group), of an individual (how to
improve its efficiency, the organization of a user
among his tasks), for private purposes (how to
improve the private organization of a user) and for
collective purposes (how to improve the
communication inside a group, etc.).
4 EXAMPLE
Figure 5 illustrates an example of interaction on
different subjects of two groups of users with a
histogram chart. Each line represents the
collaborative trace of a user, for each subject.
Now we evaluate the competence of the user
“Ning” on the subject "WP". According to the
histogram, “Ning” has realized 13 actions among
which 1 action concerns "WP". In group 1, the
number of users is 4 among which 3 have realized at
least an action about "WP". According to our
method of evaluation, we obtain:
,
1
1
3
log
4
3
0.645
(4)
,
3
12
log
3
2
2.491
(5)
Even coming from two different groups, we can
still recommend that “Ala” is more competent than
“Ning” on “WP” relying on
,
,
.
Similarly, we evaluate the competence of the two
users on other subjects as shown in Table 1.
Table 1: The index of competence of “Ning” and “Ala”
about the subjects involved.
Ning Ala
WP 0.645 2.491
Android 0 1.661
iOS 3.224 0
Tablet 1.290 0
Computer 0.0463 0.0795
Figure 6 shows a part of the domain ontology in
Information Technology. It shows that the subjects
“Computer” and “Tablet” are subclasses of the
subject “Hardware”. Also, these subjects “iOS”,
“Android”, “WP” and “Java” are subclasses of the
concept “Software”. To compare the capability on
the subject “Java” between “Ning” and “Ala”, as we
are in lack of trace of “Ning” or “Ala” on “Java”, we
propose that such a subject be measured by the
classes (i.e., “iOS”, “Android” and “WP”) having
the same superclass. As they are closest in the view
of semantics, evidently they have closest meaning.
TowardsaRecommenderSystemfromSemanticTracesforDecisionAid
277
Figure 5: An example of collaborative interactions of two groups.
Information
Technology
Hardware Software
Computer Tablet iOS Android WP JAVA
is is is
is is
is is is
Figure 6: Semantic relationship between subjects in which
“Ning” and “Ala” are involved.
We propose:
,
,
,
,
(6)
Where:
,
: the index of competence of user U
1
about subject S, in case concerned traces is not
given;
∑
_,
: the sum of index of competence
about user U1 concerning subjects which are
siblings of subject S;
Software
iOS:
3.224
Android:
0
WP:
0.645
JAVA:
?
is is isis
Figure 7: The index of Competence about sibling concepts
of “Java” of “Ning”.
Here “Java” shares the same ancestor “Software”
with “iOS”, “Android”, and “WP”. We obtain:
_
,
=
,
+
,
+
,
= 3.869
(7)
And,
_
,
4.152
(8)
As
∑
_,
∑
_,
we can
deduce that
,
,
so that we
recommend “Ala” be more qualified than “Ning” on
the subject “Java”.
5 CONCLUSIONS AND FUTURE
WORK
Traces are important records in collaborative
environments. Fully exploitation of such information
helps us organize and improve collaboration. In this
article we propose recommendations based on
evaluation of traces using TF-IDF. Moreover we
demonstrate that we could solve the problem when
there does not exist enough relevant traces with
semantics. We illustrated our method on a toy
example.
Future works include implementing our proposal
of recommender system. In a collaborative
environment, the date and time when an action is
realized are also recorded. So, it is necessary to take
KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing
278
into consideration the fact that a recent action has
more weight than a previous action.
REFERENCES
Abel, M. H., and Leblanc, A. (2009, October). Knowledge
sharing via the E-MEMORAe2. 0 platform. In Proc of
the International Conference on Intellectual Capital,
Knowledge Management and Organizational Learning
(pp. 10-19).
Adomavicius, G., and Tuzhilin, A. (2005). Toward the
next generation of recommender systems: A survey of
the state-of-the-art and possible extensions.
Knowledge and Data Engineering, IEEE Transactions
on, 17(6), 734-749.
Antoniou, G., and Van Harmelen, F. (2004). A semantic
web primer (pp. 253-260). MIT press.
Chang, N., Irvan, M., and Terano, T. (2013). A TV
Program Recommender Framework. Procedia
Computer Science, 22, 561-570.
Chen, Y. J., Chu, H. C., Chen, Y. M., and Chao, C. Y.
(2013). Adapting domain ontology for personalized
knowledge search and recommendation. Information
and Management, 50(6), 285-303.
García-Crespo, Á., López-Cuadrado, J. L., Colomo-
Palacios, R., González-Carrasco, I., and Ruiz-Mezcua,
B. (2011). Sem-Fit: A semantic based expert system to
provide recommendations in the tourism domain.
Expert systems with applications, 38(10), 13310-
13319.
Grudin, J. (1994). Groupware and social dynamics: Eight
challenges for developers.Communications of the
ACM, 37(1), 92-105.
Jones, K. S. (1972). A statistical interpretation of term
specificity and its application in retrieval. Journal of
documentation, 28(1), 11-21.
Li, Q., Abel, M. H., and Barthès, J. P. (2012). Facilitating
Experience Sharing Groups-Collaborative Trace Reuse
and Exploitation. In Proceeding of International
Conference on Knowledge Management and
Information Sharing (pp. 21-30).
Tomaz, L. F., Nt, J. A. R., Souza, J. M., and Xexéo, G. B.
(2011, June). Bringing knowledge into recommend-
dation systems. In Computer Supported Cooperative
Work in Design (CSCWD), 2011 15th International
Conference on (pp. 246-252). IEEE.
Vesin, B., Ivanović, M., Klašnja-Milićević, A., and
Budimac, Z. (2012). Protus 2.0: Ontology-based
semantic recommendation in programming tutoring
system. Expert Systems with Applications, 39(15),
12229-12246.
Zarka, R., Cordier, A., Egyed-Zsigmond, E., and Mille, A.
(2011). Trace replay with change propagation impact
in client/server applications. Actes de IC2011, 607-
622.
TowardsaRecommenderSystemfromSemanticTracesforDecisionAid
279
