V3 i35

52
International Journal of Computer Science & Applications © 2006 Technomathematics Research Foundation
Vol. 3 Issue 3, pp 52 - 64
Implementations of Web-based
Recommender Systems Using
Hybrid Methods
Janusz Sobecki
Institute of Applied Informatics
Wroclaw University of Techmology, 50-370 Wroclaw,
ul. Wyb. Wyspianskiego 27, Poland
e-mail: sobecki@pwr.wroc.pl
Abstract
Application of hybrid recommendation enables overcoming disadvantages
of the basic recommendation methods: demographic, collaborative and
content-based. In this paper the two consensus-based hybrid
recommendation methods are presented. Then some examples of their
implementations to different web-based systems are shown.
1 Introduction
With growing popularity of the web-based systems that are applied in many different
areas, they tend to deliver customized information for their users by means of utilization
of recommendation methods. We can distinguish three basic recommendation
approaches: demographic, content-based and collaborative. The demographic approach
uses stereotype reasoning, which is mainly classification problem, in its
recommendations [6] and is based on the information stored in the user profile that
contains mainly different demographic features to generate initial predictions about the
user [7]. Content-based recommendation takes descriptions of the content of the
previously evaluated items to learn the relationship between a single user and the
description of the new items [7]. In this method a user is supposed to like a new item if
the item is similar to other items that are liked by the user [1]. The collaborative
recommendations are able to deliver recommendations based on the relevance feedback
from other similar users. Its main advantages over the content-based architecture are the
following [7]: the community of users can deliver subjective data about items together
with their ratings and it is able to offer completely new items to the particular user.
All the above mentioned methods have some disadvantages. The demographic
recommendations have two basic disadvantages [7]: they may be too general and they do
not provide any adaptation to user interests changing over time. Content-based method
tends to overspecialize recommendations and it is based only on the particular user
relevance. Collaborative recommended agents have also some disadvantages: they offer
poor prediction when the number of similar users is small and there is a lack of the
transparency in the predictions.
The disadvantages of all the above mentioned recommendation methods may be

53
overcome by application of the hybrid solution. For example, the lack of adaptation in
the stereotype may be easily overcome by application of collaborative or content based
recommendation. On the other hand, disadvantages of the collaborative approach of the
insufficient number of the similar users at the early stages of the system operation may
be overcome by application of the demographic stereotype reasoning [14]. Finally some
disadvantages of content-based recommendation, such as overspecialization, may be
overcome by application of collaborative approach [14].
In the following section the user model in the hybrid recommender system is defined.
In the section three two hybrid recommendation methods are presented. The section four
contains description of different implementations of these two hybrid methods applied
for different web-based systems and finally, in the summary the efficiency of the hybrid
approach and future work is discussed.
2 User model in the web-based hybrid recommendation
A user model contains knowledge about the individual preferences which determine his
or her behavior within the system [11]. The main problems with the user modeling
encompass the user model representation and acquisition. Here the following user model
elements will be presented: user profile representation, user profile initialization,
distance and similarity functions among user profiles and user profile clustering
methods.
2.1 User profile representation
User profile should contain all necessary data to model the user in recommender system.
The user profile in the recommender systems could be represented by many different
forms: binary vectors, feature vectors, trees, decision trees, semantic networks, Bayesian
networks, etc. The feature vector is the most popular among many recommender systems
[7]. In this paper, the user profile is represented by a tuple that have similar descriptive
strength as the feature vector but it is used in many previous works on information
systems and consensus methods that is applied for the recommendation in the systems
presented in this paper. The tuple is a function p: A → V, where A is a set of attributes
and Vis a set of their elementary values and (∀a∈ A)(p(a)∈Va).
The user profile usually contains two types of data: user data and usage data [6]. The
user data may contain different user characteristics: demographic data (name, gender,
address, etc.), user knowledge, user skills and user interests. The usage data are observed
directly from the user’s interaction with web-based system. The usage data concerning
web-based system user interface recommendation may contain information on the
interface layout, the information content and its structure. The usage data may also
containusability measure value of the particular interface settings [4].
However the quality of the system interface may be difficult to measure the domain
of the HCI has worked out several methods to evaluate the user interface for example
[8]: heuristics, questionnaires and user tests.
2.2 User profiles initialization
The initial profile may be empty, especially in case of content-based recommendations.
In other approaches the initial profile is quite often created from the questionnaire that is
filled in by the user. The questionnaire usually contains information on user data that
contains different information, i.e. demographic data containing: record data (name,
address, phone number, e-mail), geographic data (city, region, country and zip-code),
user’s characteristics (sex, education, occupation), and some other customer qualifying

54
data. The user data may also contain information on users’ knowledge, their skills,
interests and preferences and also their plans and goals.
The usage data, the second element of the user model, is observed and recorded
during the whole process of user’s interactions with web-based systems. It may concern
selective operations that express users’ interests, unfamiliarity or preferences, temporal
viewing behavior, as well as ratings concerning the relevance of these elements that are
expressed by different events, such as: opening a page, purchasing a product, sending
feedback information to the system are stored. There are of course more sophisticated
and general methods for gathering such data, as for example DoubleClick mechanisms.
DoubleClick enables tracking the user web activities and serving personalized
advertisements [17] by using cookies entries with unique identifiers and placing web-
bug image on every page that is to be tracked.
The initial profile may be also modified according to the whole user population
behavior, which is the case of collaborative recommendation. However, this brings some
problems with finding similar users that could be solved by clustering methods [5].
In all implementations presented in this paper, the user profile is created from the
questionnaire that is filled in by each user during the registration process that is
obligatory for each user. This is necessary because we use demographic
recommendation. The user profile is also modified during working with the system by
each particular user. These data is then used in the collaborative and content based
recommendation.
2.3 Distance and similarity between user profiles
The distance function between values of each attribute of the user profiles is defined as a
function δat
: Va × Va →[0,1] for all a∈A. This function should be given by the system
interface designer and fulfill all the distance function conditions, but not especially all
the metrics conditions, and should be determined for each attribute and its every pair of
atom values. The distance function values could be enumerated or given in any
procedural form.
The distance between user profiles could be defined in many different ways. First,
the distance between tuples i and j could be defined as a sum of distances between values
of each attribute:
∑∈
=
Aa
ji
at
ji (a))(a),r(rd),rd(r .
Second, the root of the sum of squares of these distances, or finally third, we also can
indicate the importance of each attribute a by multiplying the distance by appropriate
factor defined as a function c: A→[0,1]:
∑∈
∗=
Aa
ji
at
ji a,rardac,rrd ))]()(()([)( .
However, it is also possible to use some similarity functions instead of distance
functions. Very popular among them is the cosine similarity function [13]:
∑ ∑
∑
= =
=
∗
∗
=
m
i
m
i
ikij
m
i
ijik
kj
uu
uu
uus
1 1
2
,
2
,
1
,,
)()(
),( ,
where uj and uk are vectors of real values of m attributes describing users j and k
respectively.

55
We can find quite many different similarity functions for example in the work [13],
the function used in the Dattola clustering algorithm [2] is shown below:
),min(),(
1
,,∑=
=
m
i
ijikkj uuuus .
In all implementations presented in this paper, we applied the third distance, the one
with a factor that defines importance of each attribute. This distance is rather easy to
implement, however quite expressive. For finding the similarity between the user profile
and items description we used the cosine similarity function that is the most popular one
especially in the area of Information Retrieval (IR).
2.4 User Profile Clustering
The clustering problem is defined as partitioning the given set of users U={u1,…,un} into
subsets of set U according to some optimization criterion. We can distinguish three
major types of clustering algorithms [3]: hierarchical, Euclidean or similar metric space
and similarity matrix. The most popular are those algorithms that belong to the family of
Euclidean or similar metric space clustering.
Then the clustering optimization criteria could be described as finding such partition
of the set U into p disjoint subsets Ci i=1,..,p of users such that the distance among all
the members of each class - d(Ci) is minimal:
∑∑= =
=
r
j
r
k
kji uudCd
1 1
),()( ,
where r=Card(Ci), or the distance between all p groups as stated below is maximal:
∑∑
∑∑
= =
=
−
=
=
=
r
l
q
k
klji
p
i
i
j
ji
p
uudCCd
where
CCdUd
1 1
1
1
1
),(),(
,),()(
andr=Card(Ci), q=Card(Cl).
The analogous criteria for the similarity functions may be also shown. Computation
complexity of the problem of partition set U with n users into two clusters is exponential
with respect to n. So, practically other sub-optimal, however more effective, algorithms
should be used. They usually are based on the selection of some initial partition as for
example in the Dattola method [2], presented below:
Start
In the beginning we divide the set U={u1,…,un} of n users into k initial classes: C1,1,
C1,2,.., C1,k. For all of the classes we calculate the centroid. The centroid of the class Cj is
denoted asOj=( oj,1,oj,2,.., oj,m} where,



−
=
=
otherwiserb
fif
o
sj
sj
sj
,
,
,
00
where bis a priori assumed constant

56
{ } sjtjsj
Cu
sisj
fmtfr
uf
ji
,,,
,,
,..,2,1:max1 −=+=
= ∑∈
Then in t-1 iteration we have the partition Ct-1,1, Ct-1,2,.., Ct-1,k with adequate centroides
Ot-1,1, Ot-1,2,.., Ot-1,k. Let T be the threshold used to construct new classes Ct,j that is
determined in the following way:
{ }







==
≥
=
−−
−
klforOasOas
andTOasa
C
ltijti
jtii
jt
,..,2,1),(max),(
),(:
,1,1
,1
,
All the user profiles that where not included into any class should be inserted into set Lt
of isolated objects. New centroides are determined as described above if the following
condition is fulfilled:
∑∑ ∈
−
∈
>
jtijti Ca
jti
Ca
jti OasOas
,,
),(),( ,1,
,
otherwise Ot,j:= Ot-1,j. Users from the set of isolated objects Lt . The iteration ends when
for particular t and all j=1,2,..,k occurs Ot,j= Ot-1,j. Then the objects from Lt are treated
as a separate class or joined to those classes to which they are the most similar.
End.
The other way to solve the clustering problem, quite similar to Dattola algorithm is
so-called Lloyd’s algorithm [5] and is one of the most popular solutions to the k-means
problem. The Lloyds’s algorithm has the following steps. First, select randomly k
elements as the starting centers of the clusters (centroides). Second, assign each element
of the set to a cluster according to the smallest distance to its centroid. Third, recompute
the centroid of each cluster, for example the average of the cluster’s elements. Finally,
repeat steps 2 and 3 until some convergence conditions have not been met (for example
centroids do not change).
This algorithm is rather simple and has an ability to reach the end when using the
above mentioned convergence condition and for configurations without equidistant
elements to more than one centroid, it takes a long time to run. First, the step 2 that has
to be performed in each iteration costs O(kdN), where d is the dimension of each element
and N is the number of elements. Second, algorithm usually needs many iterations to
terminate.
There are however quite many modification of this algorithm that run faster, for
example bisecting k-means, that begins with single cluster containing all the elements,
then splits it in two clusters and replaces it by split clusters.
Both presented here algorithms were applied in recommender systems, which
implementations will be described in the section 4.
3 Consensus-based user interface recommendation
Consensus theory has its general origins in the social sciences and in the theory of choice
in particular [9]. The social choice theory considers problems of analyzing a decision
between a collection of alternatives made by a collection of different voters with
separate opinions and the selected choice should reflect the desires of all the individual

57
voters to the possible extent [16]. The main difference between consensus theory and the
choice theory is that the former one does not necessitates the solution belonging to the
set of opinions under consideration. In this section the model of consensus and hybrid
recommendation is presented.
3.1 Model of consensus
In the consensus model we distinguish the following elements: conflict system, conflict
profiles and conflict determination. Within this model it is assumed that a real world
domain is described by means of a finite set A of attributes and a set V of attribute
elementary values [9]. Furthermore let B⊆A and a tuple of type B is a function rB:
B→Π(VB) where (∀b∈B)(rb⊆Vb), and the set of all tuples of type B is denoted by
TYPE(B).
In the scope of the consensus-based hybrid recommendation of a web-based system
user interface, as the source of opinions we may assume each client in the distributed
system, that could be also called agents, or different events of the specified user client.
The subjects of agents' interest consist of events occurring in the world, i.e. mainly
observing user behavior, interface settings and its usability values. These observations
are called events and are stored as attribute values in a tuple of some type. The conflict
system definition is a modification of the one presented in the work [10]. Here the notion
of the category is introduced for differentiation of the recommendation type:
demographic, collaborative and content-based. The second difference relies on the
identification of the information sources. In the conflict system defined in [9] agents are
the source of different (conflict) information. The conflict system defined here as the
source of conflict information identifies not only separate agents but also depending on
the category also different events from the given agent. We can distinguish some subset
T⊆Athat contains attributes for event identification.
Definition 1
A conflict system of some category c is a quadruple: Sc
= (A, X, P, Z), where:
A – is a finite set of attributes (as defined above), including attributes that identify
eachevent;
X– is a finite set of conflict carriers, X ={Π(Va): a∈A};
P – is a finite set of relations on carriers from X, each relation is of some type L
(for L⊆A and L contains attribute or attributes that enable to identify the
observation from set T⊆L);
Z – is a finite set of logic formulas for which the model is a relation system (X,P).
Relations from the set P are classified in such a way that each of them includes
relations representing similar events. Here these observations and events concern the
different aspects of the user model. A conflict situation for a given category c of the
conflict system Sc
contains information about a concrete conflict as defined below.
Definition 2
A conflict situation of a given category c csc
is a pair <P, Y→B>, where Y is a set of
attributes that have influence on the interface settings: Y⊆LT and B⊆LT and Y∩B=∅
and rY≠θfor every tuples r∈P.
A conflict situation consists of event identifiers (conflict body) which appear in
relations P (conflict content) representing the observed (or induced) knowledge of
referring to subjects represented by set B of attributes, in this case interface settings.
Expression Y→B means that in the observed events there are differences referring to

58
combinations of values of attributes from Y with values of attributes from B, and the
purpose of the consensus choice is that for a tuple type Y at most one tuple of type B
should be assigned.
For a given situation csc
, we determine the set of events which take part in the conflict
as the projection of the set of relations P to the set of attributes K,
Event(csc
) = ? K
(P), where K⊆A, and K is a key of relation P. The set of subject elements
(or subjects for short) is defined as the projection of the set of relations P to the set of
attributes Y: Subject(csc
) = ? Y
(P) where Y⊆LT. Then for each subject e∈Subject(csc
) let
us determine set with repetitions Profile(e) which include knowledge from events on
subject e∈Subject(csc
), as the set of relations that identify the given subject e reduced to
the set of attributes B∪K of for and they are included Profile(e)={rB∪K: (r∈P) ∧
(ep rA)}.
The definition of consensus is based on the definition given in [9].
Definition 3
Consensus on subject e∈Subject(csc
) of situation csc
=<P,Y→B> is a tuple (C(csc
,e))
where C(csc
,e)∈TYPE(Y∪B) that fulfils logic formulas from set Z and one of the
following postulates are fulfilled: knowledge closure and consistency, Condorcet
consistency condition for choice and convergence condition.
The following theorem should enable to determine a consensus that satisfies all the
postulates from the above definition. The proof of this theorem is some modification of
the one presented in the work [9].
Theorem 1
If there is a defined distance function d between tuples of TYPE(B), then for a given
subject e of situation csc
=<P,Y→B> tuple C(csc
,e) which satisfies conditions of
Definition 3 and minimize the expression ∑∈ )(Pr
)),(,(
eofiler
B
c
B ecsCrδ should create a
consensus satisfying all postulates from the definition 3.
When all the attributes from the user profile are independent then the consensus
determination in the Profile(e) is reduced to the determination of consensus for each
attribute in the tuple of TYPE(B). Then depending on the microstructure of attribute
values such as 1-element sets or sets of values, and macrostructure of their universe
(distance function definition) different algorithms for consensus determination could be
distinguished.
In case of the simplest microstructure when an attribute a∈A is represented by 1-
element sets of values from some set Va, the consensus determination in the profile is
based on the consensus choice function from the Theorem 1. In case of other
microstructures such as number intervals, rankings and sets the algorithms for consensus
are more complicated and they can be found in work [10].
Some examples of the conflict systems, of Event(csc
), Subject(csc
) and Profile(e) for
each category, as well as consensus determination within the profile will be shown in the
following sections.
3.2 Consensus -based recommendation methods
The first consensus-based recommendation method was presented in previous works
[10] and [14]. This method of the user interface recommendation was based on the

59
hybrid approach that was the mixture of the demographic and the collaborative
recommendation with some components of the content-based approach.
The system adaptation (see figure 1) starts with registering each new user. The
registration data are stored in the user profile. Then according to the user profile the user
is assigned to the appropriate group. With each group of users there is associated a
corresponding interface profile. The user registration is not obligatory but in this case the
default interface profile is delivered.
According to the interface profile the actual user interface content, layout and
structure is generated. The user may start to work with the system and if he or she wishes
Figure 1.Architecture of the consensus-based user interface adaptation [14]
also modify the interface settings. Finally, these settings together with usability
evaluation given by user are stored in the interface profile.
When the system registers required number of users, first users are clustered using
Dattola algorithm and then according to these clusters using consensus methods new
interface profiles for recommendation are distinguished. These procedures may be
repeated from time to time in the following cases: many new users registering to the
system or interface recommendations become poor usability ratings.
The second method of the consensus-based hybrid recommendation using
demographic, collaborative and content-based approaches applied to different
components of the user model. In this paper we will only give a short description of the
three types of recommendation, the more precise one may be found in [15].

60
In the user profile represented by the set of attributes A we can distinguish the
following subsets: the demographic attributes set D, the demographic attributes of the
centroid classes of users set N (N⊆D), the set of recommended interface settings I, the
set of the interface setting attributes made directly by the user J, the set of the actual
interface setting attributes together with usability valuation values F, the set C of
attributes associated with the content (for example visited pages, purchased or ordered
items, retrieved elements) and finally, we can distinguish some attributes used for
identification and authorization purposes T. So the set of attributes equals the sum of its
elements: A=D∪J∪I∪C∪T.
Each of the determined centroids has the corresponding interface settings, which is
assigned by the expert and is recommended to the user after registering to the system by
delivering demographic information (entering values of attributes from the setN).
In the implementation described in work [14] we considered only single expert
opinion, we can assume however that more than one expert opinion is allowed and then
determine consensus among all the opinions.
In case of the multiple expert opinions we can distinguish two situations. In the first
one, all the experts share the same centroid attribute values (concerning demographic
attributes) but give different opinions on interface settings. In the second situation, all
the experts give opinions on centroids settings and corresponding user interface
attributes values.
In both cases, however, we try to find consensus for each distinct centroid.
Concerning the second case, the number of centroid may increase significantly. When
they outnumber the desired limit we can group them, find new centroids and then find
the consensus within these groups.
The application of collaborative recommendation is possible when significant number
of users have been registered, used the system, personalized the interface and delivered
ranked the interface usability. More precisely we must have the group of similar users G
concerning values of demographic attributes from the setD.
The user groups are identified by the centroids, determined by the user clustering
described above. For each group corresponding interface settings are entered into the
same consensus profile and then the consensus is determined by finding the values of the
interface settings for which:
∑∈
∗
)(Pr
)),(,(
eofiler
B
c
Busability ecsCrr δ ,
is minimal, where attribute set B=J{usability}. We should notice that comparing to
the expression from the Theorem 1 the distance is multiplied by the value of the usability
of each particular interface.
To deliver content-based recommendation for a particular user we must have
sufficient usage data of that user and appropriate inductive rules that will transform this
data into the user interface settings.
The rules for efficient content-based recommendations strongly depend on the goals
of the web based system. For example for web-based information retrieval systems we
can consider the previous relevant items as a basis for recommendation of further
retrievals. In this case many different methods can be used: fuzzy retrieval, Bayesian
networks or other intelligent information retrieval method. For quite many systems
however, the logic used for the retrieval systems, does not hold. So, for each
recommended item, no matter if an element of interface settings or a content item, we
shall define precise relationship between the user profile (or also other users profiles)
and this particular item, which may be implemented for example as ruled based system,
Bayesian or neural networks. However the usage data may lead to many different
recommendations so there is a place for consensus determination.

61
Beside above mentioned recommendations: demographic, collaborative and content-
based, we should also mention other ones, such as: environment, situation or emotion
based. This kind of recommendations may be dealt in two different ways. The first one is
based on the expansion of the subject’s attribute set with the attribute concerning
platform, situation or emotions in standard collaborative or content consensus-based
recommendation. The second method treats these recommendations as separate ones.
The consensus system for that kind of recommendation is similar to the collaborative
one. However despite grouping users according to the demographic features we can
group them according other attributes describing computer environments, situation of
use,etc.
The result of consensus determination in all the categories: demographic,
collaborative, content-based, environment or situation based is a recommendation of the
user interface for the particular user. Obviously there can be significant differences in the
recommendation of each user interface attributes. So the question arises, which type of
recommendation should be preferred?
In the cases when only one type of recommendation is available and the others do not
deliver any (null value of the recommendation) then it is obvious that we should use the
one that is available. However usually three or more types of recommendation may
deliver different settings. In such cases specific selection rules should be applied. Some
examples that describe several implementations of web-based systems user interface and
content recommendations, will be given in the following section.
4 Implementations of web-based hybrid recommender
systems
In this section we will present six selected web-based recommender systems, first in the
subsection 4.1 three systems that are based on the combination of demographic and
collaborative filtering and then in the subsection 4.2 the other three systems that are
based on the complete hybrid recommendation.
The first car information system [14] was implemented by M. Weihberg within his
master thesis work supervised by the author of this paper, who also carried out series of
tests that were necessary to verify recommendation method. The other systems were
implemented by students of the course “Interactive web-based information systems
design” curried out in the academic years 2003/2004 and 2004/2005 supervised by the
author of this paper. The application domains of recommender systems were following:
windsurfing, digital cameras (two systems), notebooks, computer news, movies (two
systems), cooking, motor-bikes, mobile phones (two systems) and CD’s with music (two
systems). Here we will present five of them applied in the following areas: hair-dressing,
mobile-phones, cooking, movies and computer news.
4.1 Implementations using demographic and collaborative
filtering methods
In the car model information system that was started in 2003 we have chosen Renault
Megane II (c) the model of the car that was awarded with Car of The Year 2003. That
choice was made in order to attract as many users as possible and also because of the
ease of acquiring all the necessary materials to build the system. This system had no
commercial application and was used only in our laboratory. The car model presentation
system is quite simple and so is the user profile. The data stored in the user profile are
entered by the users during the registration process. The user data are reduced to only
few attributes of demographic information and one that characterizes user’s interests.
The interface profile attributes that may be personalized by the user or recommended by

62
the system, contain information concerning interface layout, music and sounds,
information content and usability factor. The attributes values of the initial centroids for
stereotype reasoning were selected by experts so that none of the stereotype had all the
extreme (maximal or minimal) values and the distance between consecutive centroids
wassimilar.
The effectiveness of the recommendation was tested in controlled conditions by 75
users that were students of masters’ degree and postgraduate studies. The users were of
different age, gender, education, musical and graphical taste, preferences concerning the
system layout and information content. The tests were conducted in three steps. In the
first step a group of the users registered themselves and they were assigned to the
appropriate group according to the smallest distance to the centroids. According to these
assignments corresponding interface profiles were delivered to them. In this step users
were not allowed to personalize their interfaces and at the end the users were asked to
fill-in the questionnaire concerning seven usability aspects, i.e. information content,
visual content, interaction etc. In the second step a new group of users was asked to
register themselves and personalize their user interface according to their preferences
starting from the settings assigned by stereotype reasoning. Then they were asked to
assess the general usability with four grades scale. At this point recommendation
procedures were carried out. In the third step the last group of users, as in the first step,
was assigned to the appropriate groups according to their distance to the centroids. Then
according to these assignments, corresponding interface profiles ware delivered to the
users. Again after working with the system and obtaining all desired information on the
car model users were asked to fill-in the usability questionnaire. Comparison of the all
the user interface usability aspects between results obtained in the first step and the third
step showed that the interface adaptation results in receiving higher scores. The raise of
marks was rather small (0.3 in average in the 10 point scale) but was encountered in all
aspects, so user interface recommendations delivered by adaptive procedures were better
than those delivered by experts, but both were rated very high - 8.42 before and 8.72
after the adaptation.
The second system that was also implemented using the first recommendation method
was “Have-a-Look!” - hair-dress information system, by K. Górka and E. Wasilewska.
The system recommends the following interface settings: layout, font size and type, text
and background color, hints, sound track and loudness. The user profile attributes
concern some demographic features such as: age, profession, interests; and some specific
system domain features such as: reason for using the system and how often the user
changes hair-dress style. The recommended user interface usability of this system was
tested using two methods questionnaire and user tests that showed that the interface is
easy to use even for not very experienced users and any major usability problems were
encountered.
The third system by M. Ciesla and A. Siodlak is presenting selected mobile phone
model. The system recommends the following interface settings: layout, colors, music
and music loudness. The user profile attributes concern only some demographic features
such as: age, gender, education and life-style. The system was also tested using
questionnaire and user tests methods. Questionnaires showed that implemented interface
recommendation delivers better settings than the ones delivered by experts however user
tests showed some usability problems that should be considered in a real system.
4.2 Implementations using hybrid recommendation methods
The following three implementations used the second hybrid recommendation method
that takes into account also content-based approach. In “The cooking assistant” by M.
Podyma & M. Swieczak recommender systems, different types of recommendation were

63
used: basic and hybrid ones. The demographic, that basing on the age and gender of the
user offers different interface settings such as: music track, volume, font size and hints;
and content settings: additional information and wine selection. These settings may be
changed by the users so it is possible to find consensus among these settings and offer
them as a new recommendation for similar users. This implementation also delivers
situation recommendation that offers receipts for breakfast, lunch or dinner according to
the daytime. Finally also content-based recommendation is applied according to the
preferences explicitly stated by the users concerning preferred cuisine and specified food
products. The system usability was tested with 12 users using the questionnaire method
and the classical usability tests [12]. The test proved that the system usability is
satisfactory.
In “The Movies” by F. Luczys & M. Filarska recommender system also different
types of recommendation were used. The system delivers information about movies
stored in the system “Stopklatka” and current cinema repertoire from that information
system. The user model is initialized during registration process, where each user is
asked to give ratings of selected films (as in MovieLens system). Then according these
ratings each user is grouped according to the stereotype reasoning specific interface
settings concerning layout, types of icons, color and background are recommended. The
user profile settings concerning different movie attributes such as: genre, director,
writing, music, cast and cast are changed according to the explicit user ratings of films or
specified attributes and implicit user searching and browsing of movies.
The system delivers reach collaborative recommendation that concerns: interface
settings according to the changes delivered by the similar users, sorting the current
cinema repertoire for unregistered users according to the preferences of all the system
users and optionally also for registered users according to the preferences of all similar
users, three extra recommended movies according to the highest ranks given by the
similar users. The content based recommendation for registered user that sorts the
cinema repertoire according to the preferences reflected in the user profile. By default
the hybrid recommendation that is mixture of collaborative and content-based repertoire
sorting is applied. The conducted usability tests containing the questionnaire method and
the classical usability tests showed some minor problems with interaction but most of the
users appreciated implemented recommendation mechanisms.
Finally, the simplest system “Comp News” by L. Sliwko & K. Jankowski
recommends computer news. At the beginning users delivers some demographic data,
preferred interface style and their interests on four topics: purchase optimization, over-
clocking, hardware modding and new products, in form of a feature vector. Then news is
sorted according to the similarity with the interest vector, which is subject to constant
changes according to reading consequent news. The user may also select an average
interest vector determined for the group of similar users. Even that simple approach was
assessed by the experimental users as being pretty useful.
5 Summary
In this paper only six hybrid recommendations implementations were shown. However,
many other web-based recommender systems were implemented by students of the
course “Interactive web-based systems design” and masters works in recent two years,
proving that it is possible to implement hybrid recommendation in many different areas.
In these implementations the consensus methods were usually used in the collaborative
recommendation, but it is also possible to apply the consensus methods for demographic,
content-based and also combined hybrid recommendations.
It is quite difficult to test recommender systems, especially in controlled condition,
because many different users are necessary to show how collaborative method operates,
as well as each user needs rather long time of working with the system to show how

64
content-based method operates. However all of the systems were tested with tens of
users (one almost 80) and some methods were tested with most exhaustive user tests
method, which methodology suggest to test only five users. These tests showed the
increase of some usability factors after application of system recommendation.
References
[1] Dastani M, Jacobs N, Jonker CM, Treur J, Modelling User Preferences and
Mediating Agents in Electronic Commerce. LNCS 1991, 2001, pp. 163-193.
[2] Dattola RT, A fast algorithm for automatic classification. Report ISR-14 to the
National Science Foundation, Section V, Cornell University, Department of
Computer Science, 1968.
[3] Estivil-Castro V, Yang J, Categorizing visitors dynamically by fast and robust
clustering of access logs. LNAI 2198, 2001, pp. 498-507.
[4] International Standard ISO 9241-11, Ergonomic requirements for office work
with visual display terminals (VDT’s) – Part 11:GuidanceonUsability, 1997.
[5] Kanungo T, Mount DM, Netanyaho NS, Piatko C, Silverman R, Wu AY, An
Efficient k-means clustering algorithm: analysis and implementation. IEEE Trans.
On Pat. Analysis And Mach. Intell., 24 (7), 2002, pp. 881-892.
[6] Kobsa A, Koenemann J, Pohl W, Personalized Hypermedia Presentation
Techniques for Improving Online Customer Relationships, The Knowledge Eng.
Review, 16(2), 2001, pp. 111-155.
[7] Montaner M,. Lopez B, de la Rosa JL, “A Taxonomy for Recommender Agents
on the Internet”, Artificial Intelligence Review, 19,2003, pp. 285-330.
[8] Newman WM, Lamming MG, Interactive system design, Addison-Wesley,
Harlow, 1996.
[9] Nguyen NT, Consensus System for Solving Conflicts in Distributed Systems,
Journal of Information Sciences, 147, 2002, pp. 91-122.
[10] Nguyen NT, Sobecki J, Using Consensus Methods to Construct Adaptive
Interfaces in Multimodal Web-based Systems, J. of UAIS., 2(4), 2003, pp. 342-
358.
[11] Papatheodorou C, Machine Learning in User Modeling, Machine Learning and Its
Applications, 2001, pp. 286-294.
[12] Pearrow M, Web Site Usability Handbook, Charles River Media, 2000.
[13] Rijsbergen C, Information Retrieval. Butterworths, London, 2nd edition, 1979.
[14] Sobecki J, Weihberg M, Consensus-based Adaptive User Interface
Implementation in the Product Promotion, in Keates S (et al.), Design for a more
inclusive world, Springer-Verlag, London, 2004, pp. 111-121.
[15] Sobecki J, Consensus-Based Hybrid Adaptation of Web Systems User Interfaces,
Journal of UCS, 11(2), 2005, pp. 250-270.
[16] Social Choice Theory, http://mathworld.wolfram.com/
SocialChoiceTheory.html, online Oct. 2004.
[17] Whalen D, The Unofficial Cookie FAQ, Version 2.54 Contributed to Cookie
Central by David Whalen,http://www.cookiecentral.com/faq/#2.7, online 2002.

V3 i35

More Related Content

What's hot (20)

Viewers also liked (6)

Similar to V3 i35 (20)

V3 i35