A PERSONALIZED FORUM ENVIRONMENT

Anna Stavrianou

Laboratoire ERIC, Université Lumière Lyon 2, Université de Lyon, Lyon, France

Magdalini Eirinaki

Computer Engineering Department, San Jose State University, San Jose, CA, U.S.A.

Keywords:

Forum, Recommender systems, Social networks, Web personalization.

Abstract:

Web2.0 has resulted in an increasing popularity of personalized recommender systems, especially in the con-

text of social networking applications. Although there exist design approaches available for such systems,

most of them make very explicit assumptions on the application domain as well as on the availability and data

types to be used as input. In this position paper, we discuss the requirements and challenges of Forum Rec-

ommender Systems. Such systems aim at generating automatically posting recommendations for the different

user proﬁles that deal with a forum. Despite the fact that these systems share characteristics with other social

media, they have hardly been explored due to the particularities they present in terms of structure, context

and user differences. Here, we discuss the particularities of Forum Recommender Systems and we propose

a framework that enables the gathering of proﬁle data and the generation of posting recommendations. The

proposed framework can also be adjusted to other social networks.

1 INTRODUCTION

The blogosphere, the forums, the web newsgroups,

the social network sites aggregate masses of user-

generated and personalized data. Nowadays, data

such as social network relationships (e.g., friendship)

and respective ratings/opinions are employed to rec-

ommend items (Guy et al., 2009; Konstas et al., 2009;

Massa and Avesani, 2007) or users (Kunegis et al.,

2009; Leskovec et al., 2010; Varlamis et al., 2010;

Weng et al., 2010). The recommended items may

be news stories, blog posts, or communities and the

recommended users may be bloggers, or likely-to-be-

friends, depending on the context.

A big challenge of nowadays is the design and the

evaluation of personalized systems and recommender

applications in the context of social networking me-

dia. Although many approaches exist, most of them

make speciﬁc assumptions on the application domain

as well as on the availability and type of data to be

used as input. However, questions such as, “How can

we design a system so as to enable the gathering of

proﬁle data?”, “What characteristics do we have to

take into account?”, “Howcan we evaluate such a sys-

tem?”, have not been answered in a systematic way.

In this position paper, we propose a generic frame-

work that addresses the aforementioned challenges,

focusing on Forum Recommender Systems. Al-

though, nowadays, the forum Web sites provide so-

cial networking functionality such as, who is fan of

whom, whether a user is popular, etc., this social net-

work (SN) information, although available, is hardly

used for purposes other than statistical or purely in-

formative. We focus on formulating the framework

that forums should be based on and the information

that should be gathered from forum web sites in order

to facilitate the implementation of a Forum Recom-

mender System.

We begin by deﬁning the characteristics of a Fo-

rum Recommender System, then, in Section 3, we

discuss the requirements based on the different user

types and tasks. Section 4 presents proﬁling users and

items, while in Section 5 challenges and open issues

are discussed. Related work is in Section 6, and the

conclusion in Section 7. We should note that since

forums share common characteristics with other so-

cial media, such as, user connectivity, shared author-

ship, tagging, and reviewing/commenting, the pro-

posed framework can be easily adjusted to other so-

cial networks as well.

360

Stavrianou A. and Eirinaki M..

A PERSONALIZED FORUM ENVIRONMENT.

DOI: 10.5220/0003400303600365

In Proceedings of the 7th International Conference on Web Information Systems and Technologies (WEBIST-2011), pages 360-365

ISBN: 978-989-8425-51-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

2 FORUM RECOMMENDER

SYSTEMS

Our objectiveis to deﬁne a frameworkthat enables the

generation of recommendations in forums. For this

purpose we have to consider a set of design features

which implicitly generate data that can be mined in

order to create user proﬁles and/or input in the recom-

mendation process. These are characteristics that are

common to any social networking application whose

actors author texts. Here, we focus on forums.

A forum can be represented by a graph whose

nodes show the users, and the edges may represent

various types of relationships. In the simple case, an

undirected edge between two nodes implies that two

users have posted on the same thread. Expanding to

a directed graph introduces follow-up discussions or

enriches the graph with social relationships, such as

sharing of information. By applying social network

analysis algorithms, we can mine interesting infor-

mation regarding the position and importance of each

user in the graph (e.g. by using degree/betweenness

centrality, or variations of PageRank).

Authorship is a major component of forums.

Users author postings and may also comment or aug-

ment other users’ postings too. These postings which

are on average much lengthier than any other social

network’s counterpart, are rich in terms of content,

context, they may contain different subtopics, and

also express speciﬁc sentiments/opinions.

Forums keep evolving and changing with time.

Thus, postings that may be popular at a certain time

period, may be displaced by other postings some min-

utes afterwards. Similarly, the location of the user in

the network, her role or the degree of interaction of

users may change at any time.

Based on these characteristics, we deﬁne a

“Forum Recommender System” (FRS) based on

the recommendation problem (Adomavicius and

Tuzhilin, 2005):

Deﬁnition 1. Let us consider a forum F having a set

of postings P. A Forum Recommender System is a

system that recommends a subset P

′

⊂ P of postings

to a user usr such that the utility u of the user is the

maximum:

′

usr

= argmax

′

⊂P

u(P

′

, usr) (1)

The utility function represents the satisfaction of the

user regarding how interesting a posting is.

The deﬁnition of an FRS points out the presence

of different user roles and tasks (i.e. why they need

a recommendation) in a forum. Different users have

different requirements and needs from a system that

recommends postings.

3 FRS REQUIREMENTS

In this section we discuss the requirements of an FRS

system per type of user.

3.1 End User

An end user is a user that browses and navigates the

forum. She desires to understand the gist of what has

been discussed and participate as well. Due to the

abundance of information, it is evident that the sys-

tem needs to generate recommendations of postings,

in order to facilitate navigation.

A FRS accessed by an end-user should be able to

comply with the following requirements:

Recommended items:

• The system should recommend a number of forum

postings to a user.

• The number of the recommended postings should

not exceed a certain threshold which could be ei-

ther user-deﬁned or user-friendly.

Access/Usability of recommendations:

• The recommendations should ideally be accompa-

nied by the location of the posting in a forum. If

the user is interested in the recommended posting

she should be able to follow the thread, the replies

to this posting or to what this posting replies to.

• The system should allow the user to browse the

results quickly (e.g. show the beginning of the

text content of each recommendation).

• The user should be able to access the recom-

mended postings quickly (e.g. by clicking on

them).

• The system should visualize the recommendations

in a user-intuitive way.

• The system could recommend postings from a list

of topic-related forums. Thus, the user would be

encouraged to visit different forums on the same

topic whose existence may have ignored.

These requirements apply also to the other user

roles and tasks, since efﬁcient browsing of forums is

a requirement for any type of user.

3.2 Forum Administrator

A forum administrator is responsible for the content

of the forum in terms of guiding and supervising its

A PERSONALIZED FORUM ENVIRONMENT

361

ﬂow. She wants to know how the discussion has

evolved, which postings have caused a lot of reac-

tions, in which parts of the discussion people argue,

etc. In order to assist this user, the FRS should make

available the selection of content-based criteria, as

outlined in the following requirements:

• The system should recommend postings that con-

tain certain (user-provided) keywords.

• The system should recommend postings that be-

long to speciﬁc topics.

• The system should recommend “controversial”

postings, where people seem to have disputed

over a point or a subject.

• The system should recommend postings that are

considered “interesting”, “uninteresting”, etc. In

this way, the rejection of hostile, insulting or spam

postings becomes more efﬁcient.

3.3 Analyst

Forum analysts are also interested in the content of

the forum. The challenges such users face involve lo-

cating the relevant information, and subsequently ac-

cessing and analyzing the useful portions related to

it. It is evident that for the analyst the most impor-

tant aspect of forum discussions is the opinions of the

forum participants. Thus, in addition to the aforemen-

tioned end-user requirements, this user category’s re-

quirements include the following:

• The system should recommend postings which

contain “positive” or “negative” comments for a

certain (user-deﬁned) product/subject.

• The system should recommend sets of postings

where participants seem to agree or disagree over

a speciﬁc (user-deﬁned) subject.

• The system should recommend postings authored

by forum “inﬂuencers” or experts.

4 CONSTRUCTING USER AND

ITEM PROFILES

In this Section we focus on the construction of user

and item proﬁles that characterize an FRS.

4.1 Proﬁling Non-participants

Users who have not yet participated in a forum are

more challenging to proﬁle, since we have little, or

no explicit knowledge about them.

4.1.1 Implicit Preferences

An FRS should be designed in such a way that allows

the gathering of implicit data. For instance, it may fa-

cilitate the tracking of the postings that the user clicks

on, assuming that these are postings the user selects

to read (Stavrianou, 2010).

Posting Content. By analyzing the content of the

read posts, the FRS can infer the information the user

is interested in. As a result, the recommendation list

can be updated with postings similar in content or

topic to the clicked ones.

Posting NavigationDepth. Navigating on the same

thread by clicking on various posts, increases the de-

gree of certainty about the user’s interest on the spe-

ciﬁc content/topic. This is in turn reﬂected in the rec-

ommendation process by re-ranking the results plac-

ing the most relevant ones higher on the list.

Duration. An FRS could also monitor the time the

user spends on a posting, by logging the time differ-

ence between two clicks. Of course, the time lap can

lead us to erroneous conclusions, since a user may not

necessarily be dealing with the clicked posting before

he clicks somewhere else. For this reason, the time

lap should be used cautiously and always in combina-

tion with other behavioral user attitudes.

Uninterest. An FRS can even log the postings not

chosen by the user. This information may reveal the

topics or style of postings the user is not interested in

so as to avoid their recommendation.

4.1.2 Explicit Preferences

Apart from the implicit ways to monitor a user, we

can also use explicit ones in order to build her proﬁle.

Proﬁle Preferences. The FRS could allow the user

to set certain criteria. The proﬁle characteristics

could be topic-related. For instance the user may se-

lect which thematic areas are of interest for her (e.g.

World Economy and Technology) and/or which are

not. The user may be able to update this proﬁle. The

proﬁle can be also implicitly updated by the postings

the user actually reads.

Moreover, depending on the user’s role and task,

a user could choose between spam postings, post-

ings with negative comments, etc. These preferences

could be logged in the proﬁle of the user together

with their frequencies (i.e. how often they are de-

manded), or could be stored permanently and only

changed when the user chooses to do so.

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

362

Posting Feedback/Ratings. The user can be given

the possibility to tag the postings she reads as to

whether they are interesting or not. In order to achieve

this, we need to include a ranking system in the web

site of the forum. In this case, the system should clar-

ify to the user the meaning of tagging a news story or

a product with a “thumbs up" or a “thumbs down" (or

a more extensive rating, e.g. in the scale of 1-5 stars).

Recommendation Feedback/Ratings. Finally,

once the system starts generating posting recommen-

dations, a user may provide feedback on whether

these recommendations are useful or not. This func-

tionality is similar to the one of posting feedback and

requires a similar mechanism to gather information.

4.2 Proﬁling Participants

A user who is already an author in the forum, ag-

gregates in her proﬁle the information of a non-

participant. Additional information can also be gath-

ered resulting from the user’s own activity.

Authored Content. The postings which are au-

thored by the user give an indication of the topics that

the user is interested in. Similarly, the content of the

postings to which the user has replied (or not replied),

can show the interests (or non-interests) of the user in

the particular forum.

Social Network Relationships. Regarding the

reply-activity, the behavior of an author can also be

tracked in the context of the SN. If, for example, a par-

ticipant tends to always reply to the same person, the

recommendations can be updated to include postings

sent by the particular person. These postings may be

located in another thread of the forum not yet visited

by the speciﬁc user.

Expertise/Inﬂuence. Identiﬁcation of the user’s ex-

pertise or inﬂuence (Agarwal and Liu, 2008; Estévez

et al., 2007; Kale et al., 2007; Zhang et al., 2007),

could re-arrange the recommendations so that post-

ings adjusted to the role of the participant are given

priority to in the recommendation list. An expert, for

example, maybe be more eager to answer to questions

and a non-expert will be more interested to receive an-

swers written by experts.

4.3 Proﬁling Postings

The item proﬁles of a FRS refer to the proﬁles of the

different postings. These proﬁles can mainly contain

semantic information such as keywords and opinion

polarity information.

Keywords. The keywords extracted from a posting

give information about its topic. The challenge in

the case of forum postings is that the text can be so

small that Text Mining techniques based on keywords

(Mooney and Bunescu, 2006) cannot always work ef-

ﬁciently. Additionally, there is often a forum-speciﬁc

vocabulary used.

Opinion Polarity. The opinion polarity of a posting

could automatically be retrieved by Opinion Mining

methods (Ding and Liu, 2007; Ghose et al., 2007; Hu

and Liu, 2004; Turney and Littman, 2003). Know-

ing the opinion polarity enables their recommenda-

tion to users who desire to have an opinion-oriented

overview, positive or negative, of what has been said

about a product, an idea or a topic.

Timestamp Information. The “age” of a posting

and its temporal distance between other postings in

the same or other forums, together with the opinion

mining can give an indication of the opinion ﬂow or

exchange between forum postings.

5 CHALLENGES, OPEN ISSUES

It is evident from the aforementioned analysis that the

design and implementation of a Forum Recommender

System raises many challenges. We discuss the most

important of them in this section.

User Proﬁling. Although a user proﬁle with some

basic user characteristics (e.g. age, gender) can easily

be kept into a database, a proﬁle that contains infor-

mation regarding what makes a message interesting

for a user or not, is not easily catered for. Proﬁling

information needs to be constantly updated, since the

interests of a user change, as well as her expertise.

Being interested in a topic today does not necessarily

mean that a user desires to receive recommendations

for the same topic all the time. In addition, a user that

is considered expert for answering certain questions

in a forum can be “outsmarted” by another user who

joined recently the SN.

Inﬂuencers. Inﬂuential users play a key role in

spreading information. A simple way to measure in-

ﬂuence is by using the structure of the SN and apply

measures such as, centrality or prestige (PageRank).

Lately, other activity/proﬁle parameters (e.g. number

and frequency of posts, comments, etc.) have been

studied in the context of blogs (Agarwal et al., 2008)

and social networks (Kim and Han, 2009). Measur-

ing inﬂuence in a forum presents a unique challenge;

inﬂuence is strongly related to trust: how much do we

trust the opinion of a user in a speciﬁc topic. Trust is

A PERSONALIZED FORUM ENVIRONMENT

363

context speciﬁc and this discrimination is important

in the FRS context: we might trust a user’s opinion

on world politics, but not on ﬁnancial matters. Thus,

the “inﬂuence” of a user is context-related. Another

challenge that needs to be addressed is deﬁning each

user’s “circle of trust” among the users of a forum.

Interest of a Posting. Deﬁning the interest of a

message per user or user-community is not easy. Is

it the content or the author that makes a post inter-

esting? What makes a posting more interesting than

another? From the content point of view, the opinion

presence, the way the arguments are presented as well

as the type of arguments may make it interesting. On

the other hand a post authored by an expert or an in-

ﬂuencer has an increased weight of interest indepen-

dent of the actual content. Modeling and measuring

the interest of a posting has not yet been dealt with.

Opinion Mining. The postings of a forum may con-

tain opinions about products, ideas, proposals, social

and economic changes. The presence of opinion in

forums can be used as a criterion for recommending

interesting messages. Until now the Opinion Mining

techniques deal with the identiﬁcation of the polar-

ity and its strength, but they do not, yet, consider the

opinion ﬂow between two posts or the opinion ex-

change between two people. The opinion-based in-

formation ﬂow has an impact on the way the opinion

changes and on the evolution of the discourse.

User Similarities. The most common technique

used in recommender systems is collaborative ﬁlter-

ing. The same algorithms can be applied in the con-

text of FRS. The challenge, in this case, is how to

model users and the similarity among them. A naive

way would be to assume two users to be similar if

they comment on the same posts. However, this ap-

proach excludes the notion of opinion, discussed pre-

viously. Two users are similar if they have the same

opinion on the same (or similar) topics. Thus there

are two parameters that need to be considered in the

FRS context: content as well as opinion polarity.

Visualization. One design issue is how to present

the recommendation list to the users. Navigating

quickly to the chosen recommendations that corre-

spond to the user criteria or interests becomes an is-

sue, especially when the posts are long and the related

forum contains hundreds of postings. Visualization

techniques need to be studied in order to represent

a forum together with the posts-related recommenda-

tion list, allowing at the same time the user to browse

efﬁciently the posts given the recommendations.

Evaluation. Evaluating recommender systems is

not easy (Herlocker et al., 2004). In the case of FRS,

the evaluation is an important issue since each forum

has a different distribution of users and postings, dif-

ferent content and style of language used. Implicitly,

we could monitor the behavior of users towards the

recommendation list. If they actually click on the pro-

posed postings, this could be an initial indication that

they ﬁnd them interesting, depending on how much

time they do spend on them, or whether they actually

follow the speciﬁc discussion thread. Otherwise, we

could explicitly ask users to rate the postings. Evalu-

ation techniques need to be studied carefully, since no

benchmark exists for the time being.

6 RELATED WORK

A personalized forum environment may use method-

ologies from various domains. In (Kunegis et al.,

2009; Leskovec et al., 2010), several algorithms are

proposed for the recommendation problem, based on

content similarity, social link information, and com-

mon items among users. The proposed models are

only applicable to social networking applications and

not other social media. In the case of blogs, ranking

algorithms have been suggested that exploit explicit

(Nakajima et al., 2005) and/or implicit (Kritikopou-

los et al., 2006; Adar et al., 2004) hyperlinks between

blogs. A similar effort is presented in (Weng et al.,

2010), while a more generic model has been presented

in (Varlamis et al., 2010).

Identifying inﬂuencers in a SN is often modeled

as a combinatorial optimization problem: given a

ﬁxed number of nodes ﬁnd the ones with maximum

inﬂuence over the network (Domingos and Richard-

son, 2001). The proposed approaches (Estévez et al.,

2007; Kimura et al., 2008; Kempe et al., 2003) are

based on the link structure, and do not consider pa-

rameters, such as activity, rate of updates, and trust

among users. Link analysis techniques (Song et al.,

2007) and activity-related parameters have been used

in order to identify inﬂuencers in blogs (Agarwal

et al., 2008) and social networks (Kim and Han,

2009). Recently, trust has also been introduced in

the context of recommender systems (Golbeck, 2006;

Golbeck, 2005; Massa and Avesani, 2007), and trust

propagation has been studied in the case of virtual

communities (Guha et al., 2004; O’Donovan, 2009;

Ziegler, 2009).

The need of the use of opinion mining techniques

is evident due to the opinion that resides inside review

sites, blogs and forums. The majority of approaches,

such as (Hu and Liu, 2004; Ding and Liu, 2007), use

a seed list of adjectives and they attempt to identify

the relation between the words in a text and those of

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

364

the seed list. An original approach is proposed in

(Ghose et al., 2007) where the opinion is inferred by

observing the effect of user comments found in a rep-

utation system on the prices of the products sold.

7 CONCLUSIONS

Forum Recommender Systems may use knowledge

and techniques from various research ﬁelds such as

the generation of recommendations in social net-

works, the presence of inﬂuence, the trust propaga-

tion. Although much work has been done in identify-

ing and incorporating these notions in other types of

social media, there does not exist an in-depth study

of how they can be incorporated in the context of

forums. This paper provides the researchers with a

generic framework and outlines the main challenges

and open areas that still need to be explored.

REFERENCES

Adar, E., Zhang, L., Adamic, L., and Lukose, R. (2004).

Implicit structure and the dynamics of blogspace. In

Workshop on the Blogging Ecosystem, WWW.

Adomavicius, G. and Tuzhilin, A. (2005). Towards the

next generation of recommender systems: a survey

of the state-of-the-art and possible extensions. IEEE

Transactions on Knowledge and Data Engineering,

17(6):734–749.

Agarwal, N. and Liu, H. (2008). Blogosphere: research is-

sues, tools, and applications. SIGKDD Explor. Newsl.,

10(1):18–31.

Agarwal, N., Liu, H., Tang, L., and Yu, P. S. (2008). Iden-

tifying the inﬂuential bloggers in a community. In

WSDM.

Ding, X. and Liu, B. (2007). The utility of linguistic rules

in opinion mining. In SIGIR-07.

Domingos, P. and Richardson, M. (2001). Mining the net-

work value of customers. In SIGKDD, pages 57–66.

Estévez, P. A., Vera, P. A., and Saito, K. (2007). Select-

ing the most inﬂuential nodes in social networks. In

IJCNN.

Ghose, A., Ipeirotis, P., and Sundararajan, A. (2007). Opin-

ion mining using econometrics: a case study on repu-

tation systems. In ACL.

Golbeck, J. (2005). Computing and applying trust in web-

based social networks. PhD Dissertation.

Golbeck, J. (2006). Trust on the world wide web: a survey.

Found. Trends Web Sci., 1(2):131–197.

Guha, R., Kumar, R., Raghavan, P., and Tomkins, A.

(2004). Propagation of trust and distrust. In WWW.

Guy, I., Zwerdling, N., Carmel, D., Ronen, I., Uziel, E.,

Yogev, S., and Ofek-Koifman, S. (2009). Personal-

ized recommendation of social software items based

on social relations. In RecSys, pages 53–60.

Herlocker, J., Konstan, J. A., Terveen, L., and Riedl,

J. (2004). Evaluating collaborative ﬁltering recom-

mender systems. ACM TOIS.

Hu, M. and Liu, B. (2004). Mining and summarizing cus-

tomer reviews. In SIGKDD, pages 168–177.

Kale, A., Kolari, P., Java, A., Finin, T., and Joshi, A. (2007).

Modeling trust and inﬂuence in the blogosphere using

link polarity. In ICWSM 2007.

Kempe, D., Kleinberg, J., and Tardos, E. (2003). Maximiz-

ing the spread of inﬂuence through a social network.

In SIGKDD, pages 137–146.

Kim, E. and Han, S. S. (2009). An analytical way to ﬁnd in-

ﬂuencers on social networks and validate their effects

in disseminating social games. In ASONAM 2009.

Kimura, M., Yamakawa, K., Saito, K., and Motoda, H.

(2008). Community analysis of inﬂuential nodes for

information diffusion on a social network. In IJCNN.

Konstas, I., Stathopoulos, V., and Jose, J. M. (2009). On

social networks and collaborative recommendation. In

SIGIR, pages 195–202.

Kritikopoulos, A., Sideri, M., and Varlamis, I. (2006). Blo-

grank: ranking blogs based on connectivity and simi-

larity features. In AAA-IDEA’06. ACM.

Kunegis, J., Lommatzsch, A., and Bauckhage, C. (2009).

The slashdot zoo: mining a social network with nega-

tive edges. In WWW, pages 741–750.

Leskovec, J., Huttenlocher, D. P., and Kleinberg, J. M.

(2010). Predicting positive and negative links in on-

line social networks. In WWW, pages 641–650.

Massa, P. and Avesani, P. (2007). Trust-aware recommender

systems. In RecSys, pages 17–24.

Mooney, R. and Bunescu, R. (2006). Mining knowledge

from text using information extraction. SIGKDD Ex-

plor. Newsl., 7(1):3–10.

Nakajima, S., Tatemura, J., Hino, Y., Hara, Y., and Tanaka,

K. (2005). Discovering important bloggers based on

analyzing blog threads. In Workshop on the Blogging

Ecosystem: Aggregation, Analysis and Dynamics.

O’Donovan, J. (2009). Capturing trust in social web appli-

cations. In Golbeck, J., editor, Computing with Social

Trust, chapter 9. Springer London.

Song, X., Chi, Y., Hino, K., and Tseng, B. (2007). Identify-

ing opinion leaders in the blogosphere. In CIKM.

Stavrianou, A. (2010). Modeling and mining of web dis-

cussions. PhD Dissertation.

Turney, P. and Littman, M. (2003). Measuring praise and

critisism: inference of semantic orientation from as-

sociation. ACM TOIS, 21(4):315–146.

Varlamis, I., Eirinaki, M., and Louta, M. (2010). A study

on social network metrics and their application in trust

networks. In ASONAM 2010.

Weng, J., Lim, E.-P., Jiang, J., and He, Q. (2010). Twitter-

rank: Finding topic-sensitive inﬂuential twitterers. In

WSDM.

Zhang, J., Ackerman, M., and Adamic, L. (2007). Expertise

networks in online communities: Structure and algo-

rithms. In WWW, pages 221–230.

Ziegler, C.-N. (2009). On propagating interpersonal trust

in social networks. In Computing with Social Trust.

Springer.

A PERSONALIZED FORUM ENVIRONMENT

365