UNSUPERVISED ADAPTATION OF THE USER INTERESTS

Lucas Marin, David Isern and Antonio Moreno

Intelligent Technologies for Advanced Knowledge Acquisition (ITAKA) Research Group

Departament d’Enginyeria Inform`atica i Matem`atiques, Universitat Rovira i Virgili, Tarragona, Catalonia, Spain

Keywords:

Recommender system, Fuzzy systems.

Abstract:

One of the main problems in recommender systems is to ensure the quality of the user proﬁle. This issue is

particularly challenging if the user preferences may vary in time. This paper proposes a novel unsupervised

algorithm to adapt dynamically the user proﬁle, taking into account the interaction of the user with the system.

The paper discusses the inﬂuence of the basic parameters of the adaptation algorithm and presents some

promising preliminary results.

1 INTRODUCTION

Nowadays, the knowledge society allows the access

to unlimited sources of information that help us in

our daily activities. In this context, when a user has

to select one alternative from a set of possibilities, an

interesting facility is the inclusion of an user proﬁle

in order to sort all the possibilities according to his

interests. A user proﬁle includes the user interests

(preferences) about the criteria that deﬁne those al-

ternatives. The maintenance of this proﬁle is made

through the collection of relevant information about

the user’s daily work. Successful interpretation of

these inputs is required to tailor systems to each indi-

vidual’s behaviour, habits and knowledge (Montaner

et al., 2003). In addition, users interests about an issue

may evolve in time due to changes in their personal

conditions (e.g., change of location, having a child).

Learning and adapting proﬁles are widely studied

research areas. The adaptation requires the collection

of information about how the user is interacting with

the system, called relevance feedback. This informa-

tion can be given explicitly by the user and implic-

itly by observing the user behaviour. Some hybrid

systems combine implicit and explicit approaches

(Nichols, 1997). On the one hand, explicit feedback

is obtained when users are required to explicitly eval-

uate items. On the other hand, implicit feedback is

obtained by monitoring the user actions and automati-

cally inferring the user preferences. Explicit feedback

has some serious limitations, being the main one that

users are reluctant to spend time giving explicit in-

formation. In this paper we present an algorithm that

adapts the personal proﬁle of the users in an implicit

way. We assume that the system is used regularly,

so that it can learn how the preferences of the deci-

sion maker evolve. The main goal is to obtain the ap-

propriate recommendation over time by adapting the

preference functions.

Figure 1: Architecture of the recommender framework.

As depicted in Fig. 1, the recommender system re-

ceives a set of alternatives composed by several crite-

ria. The ﬁrst step of the recommender is to rate all

the alternatives taking into account the user proﬁle

in order to prioritize them according to his particu-

lar circumstances (this step is out of the focus of this

paper but the reader can ﬁnd more information in (Is-

ern et al., 2010)). Then, the user selects the most ap-

propriate alternative from this sorted list. Afterwards,

the adapting algorithm collects the information pro-

vided by this selection, the set of non-selected alter-

natives, and all past selections, to infer which changes

can be made to the current proﬁle. As it will be de-

scribed later, we have implemented a Web-based plat-

form that permits to simulate all those stages and ex-

tract some conclusions.

337

Marin L., Isern D. and Moreno A..

UNSUPERVISED ADAPTATION OF THE USER INTERESTS.

DOI: 10.5220/0003180803370342

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 337-342

ISBN: 978-989-8425-40-9

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

2 RELATED WORK

As stated in (Porcel et al., 2009; Resnick and Var-

ian, 1997), there are four main approaches to rec-

ommendation: content-based systems, collaborative

ﬁltering, demographic systems and knowledge-based

systems. The main focus of this paper is to present

a knowledge-based system as depicted in Fig. 1. The

whole recommender framework includes three main

parts: representing the user interests through criteria,

rating and ranking the alternatives, and adapting the

user proﬁle through the experience.

Concerning the representation of the user inter-

ests, several works attempt to achieve this goal in-

cluding history-based models (Montaner et al., 2003),

semantic networks (Minio and Tasso, 1996), and clas-

siﬁers (Boone, 1998). These models deal with a ﬁxed

number of attributes, and the aggregation of infor-

mation included in the alternatives is also difﬁcult to

model.

The rating and ranking of alternatives is called ag-

gregation of evaluations. As summarised in (Resnick

and Varian, 1997), this stage generally concerns the

evaluation of a resource, that can be done with the

collaboration of other users (voting), or using combi-

nations of evaluations with content-analysis. The use

of the fuzzy linguistic approach has provided success-

ful results modelling uncertain, vague and imprecise

information.

The online adaptation of the user proﬁle by learn-

ing from the interaction made by the user with the

system is tackled in (Castellano et al., 2010). The

user proﬁle describes the preferences of the user on

a set of criteria with fuzzy sets. A matching mech-

anism that uses the membership functions associated

with the user proﬁle and the resource, permits to eval-

uate the similarity of both elements, recommending

the most similar. In addition, an adaptation mecha-

nism permits to adapt the user proﬁle taking into ac-

count the features of the selected resource. One of the

drawbacks of this framework is the tagged informa-

tion attached to each resource. This implies a subjec-

tive annotation of all the resources taking into account

a set of available criteria.

3 REPRESENTATION OF THE

USER PREFERENCES

The user proﬁle stores the user preferences about

some criteria using a linguistic domain (S).

Let C = {c

, . . . , c

} be the set of criteria con-

tained in the user proﬁle.

Let V

= {v

i,1

, . . . , v

i,card(c

)

} be the set of values for

the criterion c

. V is the set of all values

i∈c

The number of elements depends on the cardinality

of each criterion card(c

The proﬁle P associates a level of preference

within the linguistic domain S to each of those val-

ues (c

∈ C, v

i, j

∈ V

and s

∈ S). Then, P(c

, v

i, j

) = s

This representation assumes a common linguistic

domain S, whose selection supposes an important de-

cision that inﬂuences the accuracy of results. For this

reason different scenarios with both balanced and un-

balanced sets of terms have been tested.

Most of the fuzzy applications use balanced sets.

However, in some cases it is necessary to represent

fuzzy terms that have different shapes and member-

ship functions, as the unbalanced fuzzy set depicted

in Fig. 2. This fuzzy term set S with nine unbalanced

labels and triangular membership functions is used in

the examples shown in this paper. In this case, the or-

dered set S = {VL, L, M, V M, AH, H, VH, AP, P} per-

mits to distinguish with precision positive examples

(e.g., almost perfect -AP- is slightly better than very

high -VH-).

Figure 2: Fuzzy term set S with 9 unbalanced labels.

Table 1: Example set of rated and sorted alternatives.

An example of user proﬁle can be the interests of a

patient about some healthcare centres around his town

(see Table 1). In this case, the recommender system

is used to prioritize sets of proposals of medical visits

with specialists that a user (patient) can receive when

his practitioner requires a medical test. In this case of

study, each alternative contains the day of the week

proposed for the appointment, the period of the day

(morning, afternoon, night), the existence or not of a

parking area (blue zone -payment area-, no parking,

free parking), the distance from the patient’s home to

the medical centre (very near, near, far, very far), and

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

338

a global evaluation of the doctors of the medical cen-

tre (excellent, notable, good, normal, bad, very bad).

4 UNSUPERVISED ADAPTATION

ALGORITHM

The proﬁle adaptation approach is conducted by two

processes. The ﬁrst one, called on-line adaptation, is

executed every time the user asks the system for a rec-

ommendation, and just evaluates the information that

can be extracted from the current ranked alternatives.

The second one, called off-line adaptation, is triggered

after some user recommendations have been served,

and uses the information of the previous ranked alter-

natives and selections.

On-line Adaptation Process. The on-line proﬁle

adaptation process maintains the user proﬁle updated

by evaluating each recommendation and the choice

the user makes individually. The main goals of this

stage are to decrease the preference of the attribute

values that are causing non-desired alternatives to be

ranked with high scores, and also to increase the pref-

erence of the attribute values which are important for

the user but are not well judged on the current user

proﬁle. For each recommendation the system makes,

two sources of information are evaluated: the selected

alternative, which is the choice of the user, and the

alternatives that were ranked above the selected one.

Many conclusions can be extracted by evaluating this

information.

We extract a set of over ranked characteristics

by analysing the set of alternatives that were ranked

above the user ﬁnal selection. The process observes

criterion-by-criterion the repetitions of the values in

this set, but only when this set has enough ele-

ments. These two conditions introduce two parame-

ters named t and mo. The ﬁrst parameter identiﬁes

relevant characteristics, when a value that does not

appear in the user selection is repeated on the over

ranked alternatives in a percentage over t. When the

over ranked set of alternatives contains few elements,

the analysis of this set can produce erroneous eval-

uations. The parameter mo ﬁxes a minimum num-

ber of elements to perform this stage (by default, 5).

Those characteristics are used to decrease the level of

preference of those attribute values pointed out by the

characteristics. The intensity of that decrease is regu-

lated by the number of repetitions: if a characteristic

is repeated many times, we have more evidence to de-

crease the preference on that attribute value.

For instance, Table 1 shows a case in which the

user selects the sixth alternative. Taking a threshold t

of 40%, we obtain the tuple hDay of Week, Sat, 2i as

an example of over ranked characteristic, being Day

of Week the attribute name, Saturday a possible at-

tribute value, and 2 the times it appears among the set

of over ranked alternatives. Note that hDay of Week,

Thursday, 1i can not be considered as an over ranked

characteristic since it only appears once, thus it just

represents a 20% of the over ranked alternatives.

We call selection characteristics the features ex-

tracted from the user ﬁnal selection that do not ap-

pear on the set of over ranked alternatives more than

a given number of times. In this case, the use of the

repetition threshold t is opposite as when extracting

over ranked characteristics. The adaptation process

will only consider as selection characteristics the ones

which appear on the over ranked alternatives with a

percentage lower than said threshold. Selection char-

acteristics are used to increase the level of preference

of the attribute values pointed out by the character-

istics. The intensity of that increase is greater the

fewer times that characteristic appears among the over

ranked characteristics. Using the previous example,

we have more evidence to increase the preference of

the value Monday of the attribute Day of week than

the one of the value Morning of the attribute Time,

since the ﬁrst does not appear among the over ranked

alternatives whereas the second appears once.

Off-line Adaptation Process. The previous pro-

cess gives an immediate response to a single user in-

teraction with the recommender system. This stage

analyses past elements (non analysed over ranked al-

ternatives and past selections) that permits to comple-

ment the on-line stage.

When the user selects the best ranked alternatives,

there are not enough over ranked alternatives to anal-

yse. A buffer stores all alternatives non analysed in

the previous section, which are analysed a posteriori

as a whole. When the number of saved over ranked

alternatives is adequate, characteristics are extracted.

The parameter mo is also used here to describe the

minimum number of over ranked alternativesrequired

to start extracting characteristics. Those characteris-

tics are used to decrease the preferences in the same

way as it is done in the on-line process: characteristics

with a repetition value over the deﬁned threshold are

decreased. After the stored over ranked alternatives

have been treated, they are erased from the temporal

buffer.

On the other hand, user selections are stored and,

after a certain number of choices have been made,

they are evaluated. The number of selections needed

for an evaluation is described by the parameter h. By

extracting characteristics from that set of stored selec-

UNSUPERVISED ADAPTATION OF THE USER INTERESTS

339

tions, the most repeated attribute values, which are the

ones of most interest to the user, can have its prefer-

ence increased. After evaluating them, they are erased

from the buffer.

Adaptation Mechanism. After the on- and the off-

line stages, many characteristics can be found and the

system can deduce a large amount of changes to do

in the proﬁle. Many of those changes, also named

adaptations, can be incorrect. We have addressed this

problem by restricting the number of adaptations that

can be made to the proﬁle for each execution of the

adaptation process. The parameter pc limits the num-

ber of increases and decreases which can be made to a

proﬁle per adaptation step (a value of 2 allows a total

of 4 changes: 2 increases and 2 decreases).

When a large number of adaptations is being con-

sidered, only the more evident ones are performed.

The ones with more repetitions will be decreased. For

the increases, the evidence is measured by counting

the number of over ranked alternatives which do not

have those characteristics, so the ones with a greater

counter will be increased. A signed counter regulates

the ﬁnal increase/decrease of the preference of the

proﬁle. When the adaptation process detects an ev-

idence for a possible adaptation, it increases/reduces

a counter, initially set to 0, associated to that attribute

value. When this counter reaches the parameter k, the

increase/reduction is ﬁnally performed over the pref-

erence in the proﬁle and the counter is reset to 0.

5 EVALUATION

Web-based Platform. In order to test and evaluate

the recommender framework presented in this paper,

a Web platform has been implemented

. It permits

the deﬁnition of a problem (set of user criteria, aggre-

gation and adaptation parameters), the deﬁnition of an

initial proﬁle, the deﬁnition of an ideal proﬁle that we

aim to learn, the random generation of a corpus of al-

ternatives for the problem, the simulation of tests of

user interaction with the system, and the visualisation

of the evolution of the user proﬁle.

The accesses to the platform are regulated by a

user name and a password, allowing the maintenance

of multiple users with different proﬁles for each rec-

ommendation domain. This approach permits to anal-

yse the variability of results under different circum-

stances in the aggregation and adaptation stages.

Users can upload data ﬁles with information about

the alternativesthat must be analysed in order to make

Website: http://itaka2-deim.urv.cat/recommender/

a decision. The platform stores that information in a

database, letting users deﬁne a decision making prob-

lem on those alternatives.

General Performance. In order to evaluate the

framework, we propose to iteratively measure the dis-

tance between the current user proﬁle and the ideal

proﬁle that we want to learn. Iteratively and automat-

ically, the system selects the alternative that best ﬁts

the ideal proﬁle, which is the one that the user which

we want to have its proﬁle learnt would choose. Then,

the selection and the over ranked alternatives are used

by the adaptation processes to evaluate which changes

should be performed. To compare the results of dif-

ferent tests, several simulations have been performed

taking into account different initial proﬁles, but main-

taining the ideal user proﬁle. As it can be seen, dis-

tances between both proﬁles decrease over time.

The distance between the current (P) and the ideal

(I) proﬁles is calculated as follows:

dist(I, P) = (1)

∑

j=1

∑

k=1

|COG

(P( j, k)) −COG

(I( j, k))|

|COG

) −COG

Here, n is the number of attributes, m

is the car-

dinality of the attribute j, COG

evaluates the coor-

dinate x of the centre of gravity of a linguistic label

(Isern et al., 2010), P( j, k) and I( j, k) return the lin-

guistic preference value of the attribute j for a given

value k in proﬁles P and I respectively. With this func-

tion we obtain an average of the distances between

all pairs of labels. The distance is normalised by us-

ing the extreme values of the fuzzy set S = {s

}, i ∈

{0, . . . , T}. The distance is 0 when both proﬁles are

identical. The maximum distance is 1 when all val-

ues contained in the user and the ideal proﬁles have

just the opposite labels (s

and s

). Moreover, this

function is commutative (dist(P, I) = dist(I, P)).

The conditions of all tests were the following:

3000 alternatives initialised randomly, blocks of 15

alternatives per iteration (200 iterations), proﬁles

(ideal and initial) initialised manually, and the term

set of Fig. 2. Although we will analyse the inﬂuence

of the main parameters later, these tests havebeen per-

formed using the extraction characteristics threshold t

of 30%, with the level of evidence k +/-5 to consider

a change in a value, allowing only 2 changes (pc) in

each sense in one iteration, storing up to 5 user selec-

tions (h) and needing at least 5 over ranked alterna-

tives (mo) to extract characteristics. In all cases, three

simulations with three initial proﬁles and sharing the

ideal proﬁle have been conducted.

Fig. 3(a) compares the general performance of the

adaptation algorithm using the on-line process, and

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

340

using the on-line combined with the off-line stage.

The general performance accomplished when both

processes are working together to adapt the proﬁle is

much better than when just using the on-line process.

Moreover, Fig. 3(b) shows the labels of distance

between the values of the current and ideal proﬁles

(e.g., the distance between H (high) and P (perfect)

is 3). The ﬁgure compares how many attribute values

there are with difference 0 (correctly classiﬁed), 1, 2,

3 and more than 3 labels. It can be seen how the num-

ber of the most misclassiﬁed labels (distance 3 and

more) decreases drastically, while the set of well clas-

siﬁed attribute values (distances 0 and 1) increases.

(a) Distances between user and ideal proﬁle.

(b) Quantitate study of the labels misclassiﬁed in the

comparison of the current proﬁle with the ideal proﬁle.

Figure 3: Results of the case study.

Discussion of Results. The proposed adaptation al-

gorithm introduces several parameters that should be

properly customised. This section explains the inﬂu-

ence of some of these parameters in the ﬁnal result,

although it should be noted that exact parameter val-

ues may vary depending on the context in which the

adaptation is taking place (number of attributes, num-

ber of values per attribute, etc.).

Percentage of Over Ranked Alternatives. Over

(a) Inﬂuence of the characteristics extraction threshold.

(b) Inﬂuence of the maximum number of preference

changes at each iteration.

Figure 4: Parameters of the adaptation algorithm.

ranked characteristics are extracted by considering a

threshold deﬁned as a percentage over the number

of over ranked alternatives. Fig. 4(a) shows that af-

ter several tests (using the threshold values of 15%,

30%, 50%, and 80%), a threshold of 30% was the

one that got better results. Having a greater thresh-

old implies that the extracted characteristics need to

have a greater repetition value, thus fewer character-

istics are extracted, reducing the number of possible

UNSUPERVISED ADAPTATION OF THE USER INTERESTS

341

changes in the proﬁle. On the contrary, the extrac-

tion of valuable information is compromised as the

threshold decreases, because incorrect characteristics

(which usually have a low number of repetitions) are

extracted more easily.

Number of Changes of Preferences at each Itera-

tion (pc). On the adaptation mechanism section, a pa-

rameter for controlling how many increases/decreases

of preference can be done at each adaptation iteration

was introduced. Fig. 4(b) compares the performance

of the system with different values. Setting that value

to 2 proved to give better results than using a lower or

greater number, letting the system make up to two in-

creases and two decreases on any attribute value pref-

erence of the proﬁle. This number can vary on the do-

main that we apply the recommender framework but,

intuitively, a lower value permits to make the prefer-

ence changes for which we have a greater supporting

evidence.

Number of Evidences to Change a Preference (k).

This parameter varies the necessary number of evi-

dences to increase/decrease a preference on the pro-

ﬁle. In this case, with lower values, the number

of required evidences is lower, and the quality of

these changes is also compromised. On the other

hand, higher values for this parameter hinder making

changes on preferences, so it requires more iterations

(selections) to reach a near-ideal proﬁle. Those dif-

ferences, however, tend to decrease as the number of

iterations increases. Fig. 4(c) shows these behaviours

and how an intermediate value such as +/-5 exhibits a

good performance.

6 CONCLUSIONS AND FUTURE

WORK

The Web-based recommendersystem proposed in this

paper has been designed as a knowledge-based frame-

work that permits to solve decision problems us-

ing MCDA techniques. The separation between the

domain-related (e.g., adaptation thresholds) and the

domain-independent data (e.g., rating, ranking and

adaptation processes) permits to understand the evo-

lution of values, as well as to reuse the same methods

in a wide range of problems.

The system has been prepared to make automatic

simulations, which permit to study the inﬂuence of

several parameters of the proposed adaptation algo-

rithm.

The creation of this platform is the ﬁrst step in a

project that aims to develop techniques to adapt the

user proﬁle according to the selections made in the

recommendation process. We consider that having

automatic algorithms to update the preferences is re-

quired in many application domains, where the pref-

erences may change dynamically over time.

ACKNOWLEDGEMENTS

This work has been supported by the Universi-

tat Rovira i Virgili (predoctoral grant of L. Marin,

post-doctoral contract of D. Isern, and the project

2009AIRE-04), the Spanish Government (PlanE) and

the DAMASK project (TIN2009-11005).

REFERENCES

Boone, G. (1998). Concept features in Re:Agent, an intel-

ligent Email agent. In Proc. of AGENTS 98, pages

141–148, Minneapolis, US. ACM Press.

Castellano, G., Castiello, C., Dell’Agnello, D., Fanelli,

A. M., Mencar, C., and Torsello, M. A. (2010). Learn-

ing Fuzzy User Proﬁles for Resource Recommenda-

tion. Int. J. of Uncert., Fuzziness and Know.-Based

Syst., 18(4):389–410.

Isern, D., Marin, L., Valls, A., and Moreno, A. (2010).

The Unbalanced Linguistic Ordered Weighted Aver-

aging Operator. In Proc. of the FUZZ-IEEE 2010,

pages 3063–3070, Barcelona, Catalonia. IEEE Com-

puter Society.

Minio, M. and Tasso, C. (1996). User Modeling for Infor-

mation Filtering on Internet Services: Exploiting an

Extended Version of the UMT Shell. In Proc. of UM

96, Kailua-Kowa, US.

Montaner, M., L´opez, B., and de la Rosa, J. L. (2003). A

Taxonomy of Recommender Agents on the Internet.

Art. Intell. Rev., 19(3):285–330.

Nichols, D. M. (1997). Implicit Rating and Filtering. In

Proc. of the 5th DELOS Workshop of Filtering and

Collaborative Filtering, pages 31–36, Budapest, Hon-

gary. European Research Consortium for Informatics

and Mathematics.

Porcel, C., Moreno, J. M., and Herrera-Viedma, E. (2009).

A multi-disciplinar recommender system to advice re-

search resources in University Digital Libraries. Exp.

Syst. with Appl., 36(10):12520–12528.

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

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

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

342