Long Term Goal Oriented Recommender Systems

Amir Hossein Nabizadeh

1,3

, Al

ıpio M

ario Jorge

1,3

and Jos

e Paulo Leal

2,3

LIAAD, INESC TEC, Porto, Portugal

CRACS, INESC TEC, Porto, Portugal

Faculdade de Ci

encias, Universidade do Porto, Porto, Portugal

Keywords:

Recommender System, Collaborative Filtering, Content Based Filtering, Persuasive Recommender System,

Learning Design, Matrix Factorization, Long Term Recommender System.

Abstract:

The main goal of recommender systems is to assist users in ﬁnding items of their interest in very large collec-

tions. The use of good automatic recommendation promotes customer loyalty and user satisfaction because it

helps users to attain their goals. Current methods focus on the immediate value of recommendations and are

evaluated as such. This is insufﬁcient for long term goals, either deﬁned by users or by platform managers.

This is of interest in recommending learning resources to learn a target concept, and also when a company is

organizing a campaign to lead users to buy certain products or moving to a different customer segment. There-

fore, we believe that it would be useful to develop recommendation algorithms that promote the goals of users

and platform managers (e.g. e-shop manager, e-learning tutor, ministry of culture promotor). Accordingly, we

must deﬁne appropriate evaluation methodologies and demonstrate the concept on practical cases.

1 INTRODUCTION

Current recommender systems focus on the immedi-

ate value of recommendations. This is insufﬁcient for

achieving long term goals. For that, we need Long

Term Recommender Systems (LTRS) that are able

to guide the users to predeﬁned areas in item space

and/or to achieve other types of goals. In such a sce-

nario, user guidance would be achieved by generat-

ing a sequence of relevant recommendations through

time. For example, in the case of music, companies

may utilize these systems in order to guide the users

from a preferred music genre to a target genre. There-

fore, LTRS can be used to gradually inﬂuence users’

interests through time. Companies can use LTRS to

improve their proﬁt on selected products (e.g. new

products or products of a new segment). Another ex-

ample for the potential application of LTRS is in e-

learning. In this case, learning objects can be recom-

mended to students with a higher level objective in

view. By applying this system in the scope of learn-

ing, the activities can be more productive and less

time consuming for both the student and teacher.

In a more abstract way, the main question we pro-

pose is: how can we generate recommendation se-

quences that successfully conduct the user to a goal

(e.g. a target area of the item space), while satisfy-

ing user requirements? A goal can be deﬁned as a

pre-determined area in the item space of interest to

both user and platform manager. To attain a long term

goal, a recommendation algorithm must act strategi-

cally and not merely tactically.

The quality of a LTRS can be measured on how

it can inﬂuence users’ decisions and guide the users

towards a predeﬁned target area. Although there are

several techniques to evaluate the accuracy of RS such

as Precision, Recall or MSE, these are not enough

to assess the strategic capabilities of a recommenda-

tion system. Therefore, we argue that complmentary

means of evaluation will be needed for LTRS.

One important feature of LTRS is the ability to

persuade users. Persuasive systems have been pro-

posed by Fogg (Fogg, 2002). The aim is to use com-

puters to positively inﬂuence how users think and

act. Persuasive technology emphasizes on the social

role of computers. Since then, a few studies applied

this technology for recommender systems (Yoo et al.,

2012), which mainly focused on psychological as-

pects. We believe that the use of persuasiveness prin-

ciples can improve the effectiveness of recommenda-

tions in order to guide the users towards the long term

goals.

LTRS must also be able to handle issues such as

Sparsity and Scalability (Burke, 2002). Matrix Fac-

552

Hossein Nabizadeh A., Mário Jorge A. and Paulo Leal J..

Long Term Goal Oriented Recommender Systems.

DOI: 10.5220/0005493505520557

In Proceedings of the 11th International Conference on Web Information Systems and Technologies (WEBIST-2015), pages 552-557

ISBN: 978-989-758-106-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

torization (MF) algorithms have been successful in

dealing with both problems (Gillis, 2012). Therefore,

we believe that MF approaches can be use as a core

of LTRS.

Finally, we plan to adopt Learning Design (LD)

principles in order to generate recommendation se-

quences. In the e-learning ﬁeld, the main goal of

LD is to generate a suitable learning path (a sequence

of objects) for users. Several researchers applied

these principles for recommenders (Learning Design

Recommender Systems (LDRS)). LDRS have sev-

eral advantages such as: ﬁnding suitable learning ob-

jects, propose a well-deﬁned order of objects (a path)

and recommending not only based on the similar-

ity among objects. Current LDRS are focused on

e-learning applications and require an explicit target

item (competency) (Durand et al., 2013). Different

evaluation measures are used to evaluate the results

of LDRS.

In this position paper, we propose the idea of Long

Term Recommender Systems that guide users toward

a predeﬁned goal by generating relevant recommen-

dations. LTRS will be supported by persuasive tech-

nology, LDRS principles and MF algorithms. In ad-

dition, we plan to design a general evaluation frame-

work in order to assess the results of LTRS.

The remainder of this paper is structured as fol-

lows. Section 2 describes the related work of this

study. The research methodology is detailed in Sec-

tion 3 and then we conclude.

2 RELATED WORK

The main goal of recommender systems is to assist

users in ﬁnding items of their interest in large col-

lections. Items can be movies, news, articles, music,

places to visit, etc. The creation of the World Wide

Web and its later shift to the Web 2.0 in the 90s, made

users face the issue of data overload (Kantor et al.,

2011). Since then, the quantity of data and informa-

tion has been dramatically increasing daily.

Due to the data overload issue, ﬁnding the in-

terests of customers efﬁciently is a critical problem.

One of the tools which help users in ﬁltering informa-

tion are search engines. To personalize the informa-

tion ﬁltering process, the community created recom-

mender systems (RS). The main functionality of such

a system is to sort and ﬁlter the items according to

an acquired proﬁle of each user (Resnick and Varian,

1997).

There are many techniques and algorithms that

can be applied by recommendation systems to gen-

erate the recommendations. These techniques can be

classiﬁed in 3 main categories:

• Content based recommendation (CBF): which

recommends new items to users that are compara-

ble in content with items that a user has purchased

already (Balabanovi

c and Shoham, 1997).

• Collaborative Filtering (CF): recommends items

based on other users that have similar interests

or other items that have the same characteristics

(Balabanovi

c and Shoham, 1997).

• Hybrid method: is a combination of CBF and CF.

(Melville and Sindhwani, 2010).

2.1 Persuasive Recommendation System

LTRS are aimed at generating recommendations that

satisfy users’ requirements and persuade users to fol-

low them. However, if the users do not follow

them, the main goal cannot be accomplished. Persua-

sive technology as proposed by Fogg in 2002 (Fogg,

2002), applies computers to inﬂuence users’ thoughts

and actions. In RS ﬁeld, this technology focuses on

psychological aspect of recommendations and clari-

ﬁes how recommendations can be represented that in-

ﬂuence users more.

Persuasive RS are based on two theories: Media

equation theory (Reeves and Nass, 1997) and Com-

munication persuasion paradigm (O’Keefe, 2002).

According to communication persuasion paradigm,

the scope that a person inﬂuence others depends on

(1) form and content, (2) source, (3) the receiver char-

acteristics, (4) contextual factor (O’Keefe, 2002). If

we see the system (in our case a RS) as a person that

we communicate with (media equation theory), sys-

tem can be seen as a source, user as a receiver and

recommendations as messages. The whole process of

recommending is about a speciﬁc context. Recom-

mendations persuade receivers whether to continue

using the system or not (Yoo et al., 2012)

Figure 1: Conceptual framework of persuasive RS (Yoo

et al., 2012).

2.1.1 Source Factors

Source is a system that generates recommendations.

LongTermGoalOrientedRecommenderSystems

553

According to (Xiao and Benbasat, 2007), the factors

of source that have effect on persuasiveness of RS are:

type of recommender (CBF, CF, Hybrid); inputs such

as user preferences elicitation methods; ease of gen-

erating the recommendations and giving more control

to users during their interaction with system; process

features (like how system generates the recommen-

dations or providing information about response time

of system; embodied agent features: recommender

systems usually include virtual character conducting

the users through the process. It can be assumed that

recommendations are more convicing if system is per-

soniﬁed. Some of these features are anthropomor-

phism, agent demographics, style of speech and hu-

mor that inﬂuence the persuasiveness of recommen-

dations.

2.1.2 Message Factors

According to the previous research (Cosley et al.,

2003; Sinha and Swearingen, 2001) which are con-

ducted on the same ﬁeld content (such as discrep-

ancy, speciﬁcity, sidedness) and format of messages

(text, video, audio) are as important as the recommen-

dation system (source) and have signiﬁcant inﬂuence

on users’ evaluation and behavior.

2.1.3 Receiver Factors

User or receiver features that inﬂuence persuasiveness

of recommendations such as familiarity, involvement

and knowledge, are detailed as follow:

1. Knowledge: when users do not have sufﬁcient in-

formation about items prefer to use a website or

system which is equipped with a recommendation

system (Doong and Wang, 2011; Perera, 2000).

2. Involvement: when users explicitly participate in

the preference elicitation process, a system can

generate more accurate recommendations (Zanker

et al., 2006; Drenner et al., 2008).

3. Familiarity: users with former experience with

recommenders can trust the better and results

of RS can be more convincing (Swearingen and

Sinha, 2002).

4. Demographic Cues: genders are different in

acceptance of recommendations. For instance,

women assess the quality of recommendations

to a greater extent than men (Doong and Wang,

2011). The users’ culture can also be considered

as a demographic cue (Chen and Pu, 2008). For

example, when a recommendation system sug-

gests alcoholic drinks to a person who has restric-

tions due to his culture, it causes the user to give

a lower rate to the recommendations.

2.2 Learning Design Recommendation

Systems (LDRS)

In the area of e-learning, Learning Design (LD) is an

activity to build an effective learning path (a set of

connected learning objects) by ﬁnding suitable learn-

ing objects (Durand et al., 2013). A learning object

is any reusable digital resource which supports the

learning process (Wiley, 2003). Several approaches

applied LD in order to recommend an efﬁcient learn-

ing path. In other words, proposing and recommend-

ing an appropriate order of learning objects can be

considered as the main functionality of LDRS (Car-

chiolo et al., 2010). The main advantages of these

systems are the ability to ﬁnd adequate learning ob-

jects (items), build an efﬁcient path and avoid to gen-

erate recommendations only based on the similarity

among objects.

Different methods and techniques have been ap-

plied to propose learning design recommenders. For

example, Ullrich and Melis applied Hierarchical Task

Network (HTN) to build an adaptive structured course

generation framework for different goals (Ullrich and

Melis, 2009). Sicilia et al. also used HTN to de-

sign learning scenarios in educational systems (Sicilia

et al., 2006).

Vassileva and Deters utilized decision rules in

a tool that generates individual courses. Their

tool exploits on previous knowledge of a user and

user’s goals. This tool can be updated dynami-

cally based on user progress (Vassileva and Deters,

1998). Markov decision (Durand et al., 2011), fuzzy

petri nets (Huang et al., 2008), and production rules

(Karampiperis and Sampson, 2005) are other tech-

niques that are applied in the LD ﬁeld.

Although all above studies covered LD, none built

their models for a big set of objects. More recently a

LD approach based on graph theory (Durand et al.,

2013) has been tested on a large set of learning ob-

jects. The approaches mentioned compute the set of

possible paths and suggest one path in a single static

recommendation. When the user fails to proceed in

the recommended path, the recommender suggests

another one. Current LDRS lack a general evaluation

framework that enables the comparison of different

approaches.

Since Long Term Recommender Systems are con-

cerned with guiding the user towards a goal, Learn-

ing Design principles and existing approaches are

highly relevant for their development. LTRS gener-

alize some of the LD principles to generate long term

recommmendations in any domain and not only to e-

learning.

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

554

2.3 Matrix Factorization (MF)

Recommender systems typically face the problems of

Sparsity and Scalability. The ﬁrst one is related to

the fact that each user interacts with a very small frac-

tion of the items. The second is caused by the in-

creasingly high volumes of data found in practical

applications (Burke, 2002). Collaborative Filtering

(CF) techniques have been successful in dealing with

both problems. The majority of CF techniques are

based on the Matrix Factorization (MF) (Gillis, 2012).

Since we intend to generalize our strategy for differ-

ent scopes such as music (in case of music, we always

face the Sparsity), core of our strategy will be based

on matrix factorization approach.

Matrix Factorization (MF) discovers latent rela-

tions between users and items in a ratings matrix (Ko-

ren, 2008; Parambath, 2013). In the following we de-

scribe the basic MF setting. Suppose that R is a matrix

with size n × m (n user and m items) that entries are

ratings. By applying MF technique on the matrix:

R ≈

R = A.B (1)

In (1), A represents a user matrix with size n × q

and B with size m × q portrays item matrix. Value

q represents the number of latent factors which are

learned from past responses of user. Interpretation of

factors is not easy and change tremendously depend-

ing on q selection. The following equation shows dot

product can be applied to predict the rate of user u to

item i:

= A

(2)

Minimization of regularized squared error for

known values in R is performed as training:

min

A,B.

∑

(u,i)∈D

(

−A

)+λ(||A

+||B

) (3)

The parameter (||A

+ ||B

) is used to avoid

over ﬁtting by penalizing the high dimension param-

eters.

3 A NEW LINE OF RESEARCH

Recommender systems usually focus on satisfying

current users’ requirements and are evaluated as they

are. In this position paper, we argue for the impor-

tance of the study of LTRS that guide the users to a

predeﬁned goal in item space. The users are guided

toward goal by generating a sequence of relevant rec-

ommendations through time. We intend to design and

develop a strategy to generate the recommendations

and a framework in order to evaluate the success of

our strategy. The proposed strategy is applicable in

different domains such as music, E-learning, etc.

To learn about long term interaction between users

and recommender systems, we are currently analyz-

ing the log data of a music recommendation system

in order to see how recommendations affect the evo-

lution of users (how they response to the recommen-

dations, their current activities and interests and etc.).

We collect activity data from the recommender ser-

vice (e.g. recommendations generated, recommenda-

tions followed). Later, when our strategic recommen-

dation method is running, we will introduce it into the

recommendation service and monitor the effects of its

usage with respect to the goals of this study. We will

also look for a second application set up in the area of

e-learning to explore more the long term goal recom-

mender.

The data feed deﬁned in the previous step will be

used in a continuous streaming fashion to characterize

user behavior over time and to test the predictability

of users trajectories in the item space. The knowl-

edge acquired in this task will be important for the

development of our strategic recommendation algo-

rithm. It will also be of potential importance to other

researchers interested in user behavior and character-

ization.

The trajectory characterization step, we intend to

deﬁne a strategy that is able to learn from user activ-

ity and make a series of recommendations taking into

account well deﬁned long term goals and user satis-

faction. Learning design principles will be used to

learn from the collected users’ data in order to gener-

ate more effective recommendations to guide users.

In addition, the recommendations will be based on

a matrix factorization approach which is detailed in

Section 2.3. We will use distance based reasoning to

make sense of the space of items and represent user’s

trajectories and goal in that space. We will also ex-

ploit other data to improve recommendations, namely

item features, and user-item interaction ratings (pref-

erence rating or test results in the case of e-learning).

RS researchers evaluate their proposals using the

following approaches:

• Information Retrieval (IR) approaches such as

Precision and Recall

• Machine Learning (ML) approaches such as

RSME, MAE

• Decision Support System (DSS) approaches such

as customer satisfaction and user loyalty

Although many researchers used IR and ML mea-

sures in order to evaluate the recommenders (Yoo

LongTermGoalOrientedRecommenderSystems

555

et al., 2012), we need to continually measure users in-

teraction with system and DSS evaluation approaches

can provide more appropriate evaluation for LTRS.

Finally, we design appropriate evaluation mea-

sures and methodologies in order to assess the suc-

cess of the proposed methodology. This is a particu-

larly challenging task since, to be appropriate, evalu-

ation must be performed with live recommendations

on real situations. We will deﬁne goals for test users

and assess the success of the methodology in conduct-

ing users to the goals. Results will be compared with

a control group of users (A/B test). Ofﬂine, online

and user study are the methods that we intend to use

in evaluation phase.

Figure 2 provides a conceptual view of our idea.

It shows an item space (a set of objects with different

characteristics) that is included our target user who is

interested in speciﬁc type of objects (gray highlighted

area). Our strategy guides the target user toward goal

(green highlighted area) step by step while dynami-

cally assess how far the target user is from the tar-

get area (calculate the distance between the current

position of the target user and target area after each

recommendation). In this example, the goal of each

recommendation is to broaden the preference area of

target user until he starts to use the items in the tar-

get area (or decreasing the distance between the tar-

get user position area and target area (goal) by every

recommendation).

Figure 2: Conceptual view of LTRS.

4 CONCLUSION

Recommendation systems normally focus on the im-

mediate needs of users and are evaluated as such.

This is insufﬁcient for long term goals recommenders.

Generating strategy for long term goals is of interest

in recommending learning resources to learn a con-

cept, and also when a company attempts to convince

users to buy certain products.

In this position paper, to face the Sparsity and

Scalability problems, we propose the use of matrix

factorization algorithm to handle both issues. Also, in

order to have more conductive and effective recom-

mendations, learning design principles and persuasive

technology will be utilized. Learning design is se-

lected to learn from the users’ activities and since the

LTRS’s recommendations must persuade the users,

we apply the persuasive technology.

To evaluate LTRS, we will require appropriate

methods to assess the success of strategic recommen-

dations, since existing measures such as Precision,

and Recall are clearly not enough. In any case, ofﬂine

and online evalution must be complemented with user

studies.

ACKNOWLEDGEMENTS

This study is supported by ”NORTE-07-0124-

FEDER-000059” which is ﬁnanced by the North Por-

tugal Regional Operational Programme (ON.2 O

Novo Norte), under the National Strategic Reference

Framework (NSRF), through the European Regional

Development Fund (ERDF), and by national funds,

through the Portuguese funding agency, Fundac¸

para a Ci

encia e a Tecnologia (FCT).

REFERENCES

Balabanovi

c, M. and Shoham, Y. (1997). Fab: content-

based, collaborative recommendation. Communica-

tions of the ACM, 40(3):66–72.

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User modeling and user-adapted in-

teraction, 12(4):331–370.

Carchiolo, V., Longheu, A., and Malgeri, M. (2010).

Reliable peers and useful resources: Searching for

the best personalised learning path in a trust-and

recommendation-aware environment. Information

Sciences, 180(10):1893–1907.

Chen, L. and Pu, P. (2008). A cross-cultural user evaluation

of product recommender interfaces. In Proceedings of

the 2008 ACM conference on Recommender systems,

pages 75–82. ACM.

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

556

Cosley, D., Lam, S. K., Albert, I., Konstan, J. A., and Riedl,

J. (2003). Is seeing believing?: how recommender

system interfaces affect users’ opinions. In Proceed-

ings of the SIGCHI conference on Human factors in

computing systems, pages 585–592. ACM.

Doong, H.-s. and Wang, H.-c. (2011). Do males and fe-

males differ in how they perceive and elaborate on

agent-based recommendations in internet-based sell-

ing. Electronic Commerce Research and Applications,

10(5):595–604.

Drenner, S., Sen, S., and Terveen, L. (2008). Crafting the

initial user experience to achieve community goals. In

Proceedings of the 2008 ACM conference on Recom-

mender systems, pages 187–194. ACM.

Durand, G., Belacel, N., and LaPlante, F. (2013). Graph the-

ory based model for learning path recommendation.

Information Sciences, 251:10–21.

Durand, G., Laplante, F., and Kop, R. (2011). A learning

design recommendation system based on markov de-

cision processes. In KDD-2011: 17th ACM SIGKDD

conference on knowledge discovery and data mining.

Fogg, B. J. (2002). Persuasive technology: using computers

to change what we think and do. Ubiquity, 2002(De-

cember):5.

Gillis, N. (2012). Sparse and unique nonnegative matrix

factorization through data preprocessing. The Journal

of Machine Learning Research, 13(1):3349–3386.

Huang, Y.-M., Chen, J.-N., Huang, T.-C., Jeng, Y.-L., and

Kuo, Y.-H. (2008). Standardized course generation

process using dynamic fuzzy petri nets. Expert Sys-

tems with Applications, 34(1):72–86.

Kantor, P. B., Rokach, L., Ricci, F., and Shapira, B. (2011).

Recommender systems handbook. Springer.

Karampiperis, P. and Sampson, D. (2005). Adaptive

learning resources sequencing in educational hyper-

media systems. Educational Technology & Society,

8(4):128–147.

Koren, Y. (2008). Factorization meets the neighborhood:

a multifaceted collaborative ﬁltering model. In Pro-

ceedings of the 14th ACM SIGKDD international con-

ference on Knowledge discovery and data mining,

pages 426–434. ACM.

Melville, P. and Sindhwani, V. (2010). Recommender sys-

tems. In Encyclopedia of machine learning, pages

829–838. Springer.

O’Keefe, D. J. (2002). Persuasion: Theory and research,

volume 2. Sage.

Parambath, S. A. P. (2013). Matrix factorization methods

for recommender systems. Master’s thesis, Depart-

ment of Computing Science, Umea University, Swe-

den.

Perera, R. E. (2000). Optimizing human-computer inter-

action for the electronic commerce environment. J.

Electron. Commerce Res., 1(1):23–44.

Reeves, B. and Nass, C. (1997). The media equation: How

people treat computers, television, new media like real

people places. Computers & Mathematics with Appli-

cations, 33(5):128–128.

Resnick, P. and Varian, H. R. (1997). Recommender sys-

tems. Communications of the ACM, 40(3):56–58.

Sicilia, M.-A., S

anchez-Alonso, S., and Garc

ıa-

Barriocanal, E. (2006). On supporting the process of

learning design through planners. Virtual campus,

pages 81–89.

Sinha, R. R. and Swearingen, K. (2001). Comparing recom-

mendations made by online systems and friends. In

DELOS workshop: personalisation and recommender

systems in digital libraries, volume 106.

Swearingen, K. and Sinha, R. (2002). Interaction design

for recommender systems. In Designing Interactive

Systems, volume 6, pages 312–334. Citeseer.

Ullrich, C. and Melis, E. (2009). Pedagogically founded

courseware generation based on htn-planning. Expert

Systems with Applications, 36(5):9319–9332.

Vassileva, J. and Deters, R. (1998). Dynamic courseware

generation on the www. British Journal of Educa-

tional Technology, 29(1):5–14.

Wiley, D. A. (2003). Connecting learning objects to in-

structional design theory: A deﬁnition, a metaphor,

and a taxonomy.

Xiao, B. and Benbasat, I. (2007). E-commerce product rec-

ommendation agents: use, characteristics, and impact.

Mis Quarterly, 31(1):137–209.

Yoo, K.-H., Gretzel, U., and Zanker, M. (2012). Persuasive

Recommender Systems. Springer.

Zanker, M., Bricman, M., Gordea, S., Jannach, D., and

Jessenitschnig, M. (2006). Persuasive online-selling

in quality and taste domains. In E-Commerce and Web

Technologies, pages 51–60. Springer.

LongTermGoalOrientedRecommenderSystems

557