Marco Degemmis, Pasquale Lops, Giovanni Semeraro, M. Francesca Costabile,
Oriana Licchelli, Stefano P. Guida
Dipartimento di Informatica, University of Bari, Via E. Orabona, 4, 70125 BARI, Italy
Keywords: Hybrid recommender systems, Information Filtering, Machine learning, User profiling
Abstract: Nowadays, users are overwhelmed by the abundant amount of information deliv
ered through the Internet.
Especially in the e-commerce area, largest catalogues offer millions of products and are visited by users
having a variety of interests. It is of particular interest to provide customers with personal advice: Web
personalization has become an indispensable part of e-commerce. One type of personalization that many
Web sites have started to embody is represented by recommender systems, which provide customers with
personalized advices about products or services. Collaborative systems actually represent the state-of-the-art
of recommendation engines used in most e-commerce sites. In this paper, we propose a hybrid method that
aims at improving collaborative techniques by means of user profiles that store knowledge about user interests.
Most of the largest e-commerce Web sites is using
recommender systems to help their customers find
products to purchase. A recommender system learns
from customers and recommends products that they
will find most valuable among the available
products. Recommender systems have been
revolutionizing the way shoppers and information
seekers find what they want, because they
effectively help users in selecting items that best
meet their needs and tastes.
Such systems take input directly or indirectly from
users an
d, based on user needs, preferences and
usage patterns, they make personalized
recommendations of products or services.
Recommender systems are used to either predict
whether a particular user will like a particular item
(prediction problem), or to identify a set of N items
that will be of interest to a certain user (top-N
recommendation problem) (Sarwar, et al., 2002).
The literature on recommender systems
istinguishes primarily between the collaborative
and the content-based approaches. In the first
approach, the content (e.g. text) plays an important
role: the system suggests the items similar to those
the user liked in the past, based on the content
comparison. In contrast with the content-based
approach, a collaborative approach assumes that
there is a set of users using the system: user advice is
based on the item ratings provided by other users.
Hybrid recommender systems combining both
echniques have also been proposed to gain better
performance with fewer of the drawbacks of any
individual technique (Burke, 2002; Balabanovic and
Shoham, 1997; Konstan, et al., 1998; Pazzani,
1999). Examples of this kind of hybrid systems are
Fab (Balabanovic and Shoham, 1997) and Ringo
(Shardanand and Maes, 1995). Fab maintains user
profiles based on content analysis, and directly
compares the profiles to determine similar users for
collaborative recommendations. Items are
recommended to a user both when they score highly
against that user profile or when they are highly
rated by a user with a similar profile.
Ringo is similar to
Fab except that, during a
similarity assessment among users, the system
selects profiles of users with the highest correlation
with an individual user. Ringo compares user
profiles to determine which users have similar tastes.
Once similar users have been identified, according
to a classical collaborative approach, the system
predicts how much the user may like an item that
has not yet been rated by computing a weighted
average of all the rates given to that item by the
other users that have similar tastes.
In (Tuzhilin and Adomavicius, 1999), it is remarked
that: “In
order to provide more accurate
Degemmis M., Lops P., Semeraro G., Francesca Costabile M., Licchelli O. and P. Guida S. (2004).
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 162-169
DOI: 10.5220/0002638201620169
recommendations, it is necessary to base them on a
thorough analysis of the on-line behavior of the user
that is much broader than the behavior captured by
current content-based filtering systems”. Rules
describing the on-line behavior of a user can be
learned from the analysis of his/her transactional
history using various data mining methods and can
be included as a part of that user’s profile.
Behavioral profiles can describe much richer types
of user behavior than user profiles from the content-
based approach, but they do not provide any
recommendations by themselves. Therefore, it is
important to couple the behavioral profiling
approach with other techniques.
We consider the integration of behavioral profiles
and collaborative methods into one integral
approach. This is in line with basic principles of
marketing, according to which customer
recommendations should be based on understanding
behavior of that customer and on the preferences of
similar customers. In our approach, rules describing
the customer behavior are used in order to discover
preferences of users, such as product categories. For
example, in a book recommending context, rules
could be used in order to determine whether a user is
interested or not in a specific book category. A
simple example of such rules is: “Customers that
buy at least 3 books belonging to the horror
category are interested in that book category”.
Preferences are stored in personal profiles exploited
to group customers having the same interests. Our
idea is that profiles could drive the collaborative
method by reducing the set of users, on which the
algorithm is applied, only to users interested in the
same product categories. Profiles are inferred from
the analysis of transactional data (browsing and
purchasing history of users), without considering
any content, and are exploited to discover for each
user a set of “nearest neighbors” to compute
collaborative recommendations. An intensive
experimental session has been carried out to
compare a pure collaborative approach to
recommendation with respect to the one combined
with user profiles of users.
The paper is organized as follows: Section 2
provides a description of the most frequently
approaches used in recommender systems, i.e.
collaborative and content-based ones. It also
describes a possible way to combine the approaches
to improve the entire recommendation process.
Section 3 gives a description of the two systems,
namely User Profile Engine (UPE) and Profile
Extractor (PE), we integrated to build a hybrid
recommender called U(PE)
. Section 4 presents the
experiments performed to evaluate the possible
improvement of U(PE)
, which exploits knowledge
about the users’ behavior, with respect to UPE,
which implements a pure collaborative filtering
algorithm. Conclusions are drawn in the last Section
There are many different techniques for
implementing recommender systems (Resnick and
Varian, 1997; Schafer, Konstan and Riedl, 1999;
Terveen and Hill, 2001):
Collaborative filtering is the most successful
recommender system technology to date. The
main idea is to recommend new items of interest
for a particular user based on other users’ ratings.
These systems recommend products to a
customer based on the correlation between that
customer and other customers who showed
interests in those products, e.g. who have
purchased products from the e-commerce site.
Content-based recommender systems suggest items
based on their associated features. A pure content-
based recommender system is one in which
recommendations are made for a user based solely
on a profile built by analyzing the content of items
which that user has rated in the past.
Demographic recommender systems aim at
categorizing the user based on personal attributes
and make recommendations based on
demographic classes. The benefit of the
approach is that it may not require a history of
user ratings of the type needed by collaborative
and content-based techniques.
Knowledge-based recommenders attempt to
suggest items based on inferences about a user’s
needs/preferences. In some sense, all
recommendation techniques could be
described as doing some kind of inference.
Knowledge-based approaches have knowledge
about how a particular item meets a particular
user need, and can reason about the relationship
between a need and a possible recommendation.
2.1 Collaborative Filtering Systems
Collaborative filtering is a type of recommendation
technique that works by finding patterns of
agreement among users of the system, leveraging the
tastes and opinions about quality of all of the users
to help each user individually.
Rather than recommending items because they are
similar to items a user has liked in the past, a set of
items that other similar users have liked is
recommended. In other words, similarity of users
rather than similarity of the items are computed.
Typically, for each user a set of “nearest neighbor”
users is found whose past rates have the strongest
correlation. Rates for unseen items are predicted
based on a combination of the rates known from the
nearest neighbors. Pure collaborative
recommendations give the possibility to deal with
any kind of content. Since other users’ feedback
influenced what is recommended, there is the
potential to maintain effective performance given
fewer rates from any individual user.
Collaborative filtering has a number of advantages
over content-based methods:
The knowledge engineering problem associated
with content-based methods is relieved, since
explicit content representations are not needed.
The quality of collaborative filtering typically
increases with the size of the user population,
and collaborative recommendations benefit from
improved diversity when compared to content-
based recommendations.
However collaborative filtering does suffer from a
number of significant downsides:
It is not suitable for recommending new items
because these techniques can only recommend
items already rated by other users. If a new item
is added to the content database, there can be a
significant delay before this item will be
considered for recommendation. Essentially,
only when many users have seen and rated the
item will it find its way into enough user profiles
to become available for recommendation. This
so-called “latency problem” is a serious
limitation that often renders a pure collaborative
recommendation strategy inappropriate for a
given application domain.
Collaborative recommendation can prove
unsatisfactory in dealing with what might be
termed an “unusual user”. There is no guarantee
a set of recommendation partners will be
available for a given target user, especially if
there is insufficient overlap between the target
profile and other profiles. If a target profile
contains a small number of rates or ratings for a
set of items that nobody else has reviewed, it
may be difficult to make reliable recommendations
using the collaborative technique.
2.2 User Knowledge
A key issue in the personalization of a Web site is
the automatic construction of accurate user profiles.
A profile is a collection of information about an
individual; it permits to recognize the user, know
why he or she did something, and guess what he or
she wants to do next. User profiling is typically
either knowledge-based or behavior-based.
Knowledge-based approaches engineer static models
of users and dynamically match users to the closest
model. The knowledge about users can be acquired
in different ways. Generally speaking, it could be
acquired through questionnaires, where users select
different content types and services from a list of
predefined choices. This implies that users must
manually update their profiles when their interests
change. These limitations clearly call for alternative
methods that infer preference information implicitly
and support automated content recommendation.
Behavior-based approaches use the user’ behavior
itself as a model. Machine learning techniques are
being used to recognize the regularities in the
behavior of customers interacting with e-commerce
Web sites and to infer a model of the interests of a
user, referred to as user profile or user model. The
user model is a collection of information about an
individual and should be able to recognize the user,
know why he or she did something, and guess what
he or she wants to do next. The typical user profiling
approach for recommender systems is behavioral-
based, using a binary model (two classes) to
represent what users find interesting and
uninteresting. Machine-learning techniques are then
used to assess potential items of interest in respect to
the binary model.
2.3 Integrating Collaborative
Recommender Systems with User
User models have been used in recommender
systems for content processing and information
filtering. It could be useful to develop methods for
integrating behavioral profiling with collaborative
filtering into one integral approach. In particular, the
approach we propose integrates collaborative
techniques with user profiles inferred from the
analysis of transactional data (browsing and
purchasing history of users) without considering any
content. There are two main alternatives to
accomplish this task:
1. Profiles Drive Collaborative Methods. Profiles
are used to reduce the set of items that should be
used for computing recommendations. This
means that standard collaborative methods will
be applied, but they will work on a smaller
consideration set of data. We expect this to
increase the performances of the overall
technique in comparison to the stand-alone
collaborative filtering method.
2. Profiles Are Used After Collaborative Filtering.
Standard collaborative filtering techniques are
used to generate a preliminary set of possible
recommendations. Then, profiles are exploited to
re-rank the set of the recommended items or to
prune some of the items that were preliminarily
Our approach exploits the first alternative, but it
reduces the set of users on which the algorithm is
applied instead of reducing the set of items. In
Section 3.3 we will give more details about the
adopted approach.
In previous work, we have developed two
personalization systems, each exploiting a specific
technique for providing recommendation: UPE,
described in Section 3.1, is a recommender system
that uses filtering techniques (collaborative and
simple filtering), and PE, described in Section 3.2, is
a knowledge-based recommender system
3.1 User Profile Engine
UPE (User Profile Engine) is a recommender system
that provides personalized suggestions
(recommendations) about pages users might find
interesting in a product catalogue on the Web
(Buono, et al., 2002). The user profiles managed by
UPE have a static component and a dynamic one.
The static component consists of a set of
information that identifies each user and doesn’t
change (or change rarely). For example: name,
nationality and type of user. The information sources
come primarily from the registration forms that
some users are required to fill. The dynamic
component of user profile is the changing part of
user data. The set of user preferences is part of the
dynamic profile. UPE obtains this information by
using different type of ratings: explicit ratings, i.e.
the user explicitly indicates what he or she thinks
about an item; implicit ratings, obtained by tracking
user navigation (i.e. events as access to a Web page,
print and/or save action, etc.). Even if explicit rating
is fairly precise, it has disadvantages, such as: 1)
stopping to enter explicit ratings can alter normal
patterns of browsing and reading; 2) unless users
perceive that there is a benefit providing the rates,
they may stop providing them.
Implicit ratings are much more difficult to
determine but they have the following advantages:
1) every interaction with the system (and every
absence of interaction) can contribute to implicit
rating; 2) can be gathered for free; 3) can be
combined with several types of implicit ratings for a
more accurate rating; 4) can be combined with
explicit ratings for an enhanced rating.
Indeed, the method that is quite effective is a mixed
technique that exploits implicit and explicit ratings
and we implemented it in UPE. However, especially
in the case of sites with many pages, we can be in a
situation that some pages have not been evaluated by
the current user (neither explicit nor implicit ratings
are available). To overcome this situation, UPE uses
an algorithm of collaborative filtering. It predicts
user interests on an item not evaluated by taking into
account the historical data set on rates of a users
community stored into a database of existing rating
provided by other users (Buono, et al., 2002).
As it is well known, these algorithms are useful but
also very time consuming. With the aim to further
improve UPE performance, we have defined some
heuristics that reduce the number of users involved
in the computation of users’ preferences. Such
preferences are computed by using weights that
reflect correlation (in this case the Pearson
correlation) between pairs of users. The more
objects two users have rated similarly, the closer the
two users are. To reduce the number of
computations, UPE re-calculates only the weights
for users that at least one of the two users, during his
or her interactions with the system, has produced a
number of ratings (explicit or implicit) above a
given threshold m. Furthermore, the system re-
computes the predicted rating of a user for a certain
item by taking into account only the users, that since
the last rating updating, have generated a number of
ratings (explicit or implicit) above a threshold n.
More specifically the predicted rating is a weighted
sum of ratings of the users selected for re-
3.2 Profile Extractor
In order to provide personal recommendations based
on a comprehensive knowledge of who customers
are and how they behave, we have adopted an
approach that uses information learned from
transactional histories to construct individual
Figure 1: U(PE)
profiles. The advantage of using this technique is
that profiles generated from a huge number of
transactions tend to be statistically reliable.
The process of learning customer profiles is
performed by the PE (Profile Extractor)
personalization system (Semeraro, et al., 2003),
which employs supervised learning techniques to
automatically discover users’ preferences from
transactional data recorded during past visits to the
e-commerce Web site. In Business to Consumer
(B2C) e-commerce, items are grouped in a fixed
number of categories. For example, at Amazon.com
books in the catalogue are organized in many subject
categories. PE is able to analyze data gathered from
sources such as data warehouse or transactions, for
instance, in order to infer rules describing the
customer/user behavior. Rules are exploited to build
profiles containing preferences such as the product
categories the user is interested into.
From our point of view, the problem of learning
user’s preferences can be cast to the problem of
inducing general concepts from examples labelled as
members (or non-members) of the concepts. In this
context, given a finite set of categories of interest
C = {c
, c
, …c
}, the task consists in learning the
target concept T
“users interested in the category
”. In the training phase, each user represents a
positive example of users interested in the categories
he or she likes and a negative example of users
interested in the categories he or she dislikes. We
chose an operational description of the target
concept T
, using a collection of rules that match
against the features describing a user in order to
decide if he or she is a member of T
. Transactional
data about customers are arranged into a set of
unclassified instances (each instance represents a
customer). The subset of the instances chosen to
train the learning system has to be labeled by a
domain expert, that classifies each instance as
member or non-member of each category. The
training instances are processed by the Profile
Extractor, which induces a classification rule set for
each category of interest. More precisely, the
architecture of PE is made up of several sub-
modules: (a) XML I/O Wrapper, which is the layer
responsible for the extraction of data required for the
learning process; (b) Rules Manager, which is
implemented through one of the WEKA (Frank and
Witten, 1998) classifiers. The learning algorithm
adopted in the rule induction process is PART
(Witten and Frank, 1999), which produces rules
from pruned partial decision trees; (c) Profile
Manager, which classifies each user on the ground
of the users’ transactions and the set of rules induced
by the Rules Manager. The classifications, together
with the interaction details of users, are gathered to
form a user profile.
3.3 Integrating UPE and PE: U(PE)
Our idea is to produce a hybrid method by
integrating behavioral profiles inferred by PE and
the collaborative method implemented by UPE into
one integral approach in an attempt to demonstrate
that it outperforms the pure collaborative filtering
method. The resulting system U(PE)
(Fig. 1)
implements a cascade hybrid method: profiles
inferred by PE are exploited by the Profile Analyzer
to group customers having similar preferences. In
our case, preferences are the product categories the
customer is interested in. Our idea is that profiles
could drive the collaborative method by reducing the
set of users, on which the algorithm is applied, only
to users interested in the same product categories.
PE is applied to induce rules (describing “classes” of
users) that are exploited to build the profiles. Then,
the collaborative filtering algorithm is applied to
each group of users selected by the Profile Analyzer.
In this way, it is possible to improve computational
performance by carrying out parallel computation
for each group of users. We actually use PE to
classify registered users and assign them to the
content categories of their interest; we then apply
collaborative filtering algorithm to the users of each
class, in order to generate recommendations that fit
their interests.
We have performed two experiments in order to
compare the performance of the proposed hybrid
recommender system U(PE)
with UPE. The former
measures the evaluation of UPE implementing the
classical collaborative filtering technique (see
Section 3.1). The latter measures the evaluation of
the hybrid system U(PE)
obtained integrating the
behavioral profiles inferred by PE with the UPE
collaborative method. The performance of U(PE)
has been compared with the UPE personalization
system. For both experiments we used historical
browsing data from an Italian e-commerce company.
This dataset contains information about 380 users on
154 catalogue products; in particular, it contains
explicit rates given by users and implicit rates
computed by the system on the basis of the user
behavior. Each action performed by a user on a Web
page, for example zooming on the picture of a
product, corresponds to a rate. We divided the
dataset into a training set and a test set by using
90%/10% training/test ratio. From each user in the
test set, ratings for 25% of items were randomly
withheld. Predictions were computed for the
withheld items using each of the different
algorithms. In the first experiment, the dataset was
converted into a user-product matrix that had 380
rows (i.e., 380 users) and 154 columns (i.e., products
that were rated by at least one of the users).
Predictions were computed for the withheld items
using the pure collaborative filtering technique
implemented by UPE. In the second experiment, the
dataset was converted into 11 user-product matrices,
each corresponding to a specific product category C
in which PE classified the users. Each matrix had n
rows (i.e., the number of users that PE has classified
as interested in the category C
) and 154 columns
(i.e., products that were rated by at least one of the
users). In this case, the UPE collaborative filtering
was applied separately to each matrix. Both
experiments were repeated 5 times selecting a
different test set (the intersection of the five test sets
was empty). This procedure allows running 5
experiments that are completely different. Finally,
the results of experiment 1 were averaged over the 5
runs and ones of experiment 2 were averaged over
all categories.
The quality of the predictions was measured by
comparing the predicted values for the withheld
ratings to the actual ratings, using several metrics.
In general, recommender systems research has used
several types of measures for evaluating the success
of a recommender system. We consider only two
types of metrics for evaluating predictions and
recommendations respectively.
To evaluate an individual item prediction we used
the Mean Absolute Error (MAE) between ratings
and predictions. MAE is a measure of the deviation
of recommendations from their true user-specified
values. For the prediction of N items (p
) and a
real evaluation of a user (r
E = (|p
|) is calculated. We can compute
MAE by first summing the squared absolute errors
of the N corresponding ratings-prediction pairs and
then computing the average. Since the task was to
identify or retrieve items preferred by users from a
repository, traditional information retrieval measures
were adopted, namely Precision (Pr), Recall (Re)
(Sebastiani, 2002). We have adapted the definition
of recall and precision to our case as our experiment
is different from standard IR in the sense that we
have a fixed number of recommended items. In the
evaluation phase, the concept of relevant item is
central. An item is considered as relevant by a user if
the score he or she has given is greater than 2.5. An
item is considered as relevant by a system if the
computed numerical recommendation score is
greater than 2.5. Our goal is to look into the test set
and match items that both the system and the user
deemed relevant. Then, recall is the proportion of
relevant items that are classified as relevant, and
precision is the proportion of items classified as
relevant that are really relevant. The fact that both
measures are critical for the quality judgment leads
us to use a combination of the two. In particular, we
use the standard F1 metric (Sebastiani, 2002), which
gives equal weight to them both. We also adopted
the Normalized Distance-based Performance
Measure (NDPM) (Yao, 1995) to evaluate the
goodness of the items’ ranking calculated according
to a certain relevance measure. Specifically, NDPM
was exploited to measure the distance between the
imposed on items by the user ratings and the
ranking predicted by the system. Values range from
0 (agreement) to 1 (disagreement). Results of the
experiments are divided into two parts: quality
results and performance results. In assessing the
quality of recommendations, we first analyze the
results obtained in experiment 1 by UPE (Table 1).
Table 1 – Results obtained by UPE (averaged over 5 runs)
NDPM Recall Precision F1-measure
0.066 0.942 0.905 0.923
Notice the high accuracy that can be achieved by the
system on the whole dataset in predicting the
ranking of the products according to the customers
interests (the NDPM value is close to 0). The high
value of the F1-measure and the balance between
recall and precision demonstrates that the list of
recommendations presented to users by UPE
contains relevant items correctly ranked.
In the second experiment, we examined separately
the recommendation accuracy for users grouped
according to their behavioral profiles. F
or each
Table 2 reports the number of users classified
as interested in that category, the number of users
poorly, moderately, and strongly correlated, and the
mean correlation value computed over each pair of
A correlation coefficient between 0.3 and 0.6
reveals a moderate association, while values above
0.6 indicate a strong correlation. A coefficient at
zero, or close to zero, indicates no relationship.
In Table 3, we reported the U(PE)
results. The
values of MAE are positive given the small number
of users belonging to each category (from 35 users
in the category “kitchen utensils” to the 74 users in
the category “underwear”). Only for two
(“kitchen utensils” and “jewelry”) the value of MAE
was over 2
In particular, the computed MAE for the category
“kitchen utensils” is greater than 8. In order to fully
understand this result it is important to notice that
this category contains the smallest number of users
(35) and reported the lowest value of mean user
For the users in the category “kitchen utensils”, the
computed value was 0.42 against at least 0.49
achieved in all the other categories (see Table 2 for
more details). In general, NDPM results are very
positive (values do not exceed 0.2), showing a
strong correlation between the ranking imposed by
the users and the ranking computed by the system,
although there is a high degree of variation between
different categories. NDPM is better for users
strongly correlated and belonging to “more
populated” categories: the best values have been
found in the categories “underwear” (74 users) and
“hardware”, which show the highest values of user
correlation. For the F1 score, we consider the results
as very positive. Overall, 8 out of the 11 categories
reported values that exceed 0.80, while only for one
category (“kitchen utensils” again) the system was
not able to reach a value of at least 0.70.
The aim of the second experiment was to compare
the results obtained by U(PE)
, averaged over all the
categories, with the results obtained by UPE (see
Table 4).
As regards MAE, the value achieved by UPE is
almost five times better than the value registered for
U(PE)2. UPE outperforms U(PE)2 both for NDPM
and F1-measure.
Table 2 – Statistics on the 11 products categories
10 (13%)
14 (21%)
16 (23%)
17 (25%)
35 9 (26%) 5 (14%)
70 9 (13%) 24 (34%)
ELECTRONICS 59 9 (15%) 7 (12%)
8 (12%)
11 (16%)
INFORMATICS 65 9 (14%) 13 (20%)
BABYHOOD 70 8 (12%) 12 (17%)
41 (11%)
Table 3 – Results obtained by U(PE)
UNDERWEAR 1,473 0,040 0,879
FURNITURE 0,900 0,166 0,947
PET SUPPLIES 1,228 0,062 0,888
HOUSEHOLD ARTICLES 1,395 0,135 0,847
KITCHEN UTENSILS 8,078 0,195 0,629
SANITARY ARTICLES 1,045 0,049 0,864
ELECTRONICS 1,130 0,061 0,799
HARDWARE 0,872 0,048 0,971
JEWELRY 2,496 0,109 0,724
INFORMATICS 0,939 0,155 0,862
BABYHOOD 1,734 0,072 0,822
Table 4 – Comparison UPE vs. U(PE)
MAE NDPM F1-measure
|Diff.| UPE U(PE)
|Diff.| UPE U(PE)
0.421 1.936 1.515 0.066 0.099 0.033 0.923 0.839 0.084
This result is to be expected, as the collaborative
filtering algorithm implemented by UPE generates
recommendations based on the strength of the
association among users and it is adversely affected
by reduced training sets containing poorly correlated
users. Only 3 categories (“underwear”, “electronics”,
“babyhood”) reported at least 70% of users strongly
correlated, as in the original dataset, and that the
mean user correlation observed in each category is
always lower than in the entire dataset. Nevertheless,
the results achieved using behavioral profiles are
satisfactory: NDPM is still very close to 0 and F1-
measure shows a classification accuracy in recognizing
relevant items that is almost 84%. This means that
is able to recommend “good” items,
although the individual item prediction gets worse.
When we focus on performance issues, we find the
main advantage of grouping users according to their
behavioral profiles before computing
recommendations: the time requested by UPE to
produce recommendations on the whole dataset of
380 users was 5h 47min, while the time requested by
was 57min for computing recommendations
and 1h 27min for classifying users into 11 categories.
The total time for completing the process was 2h
Recommender systems are a powerful technology
that allows a company to get additional value from
its user database. A real problem is that these
systems are being stressed by the huge volume of
user data in existing corporate databases. A strong
research issue is to develop methods that can
improve the scalability of recommender systems,
still producing high-quality recommendations. In
this paper, we have presented a new approach for
collaborative-based recommender systems. It
integrates knowledge about customers stored in
behavioral profiles into the collaborative filtering
algorithm in order to reduce the computational time
required for generating recommendations. The final
goal of the work has been to identify some measures
for evaluating the quality of recommendations. For
this purpose, we have presented the empirical
evaluation of the U(PE)
hybrid recommender
system. Our results have highlighted the actual
improvement of the proposed hybrid approach with
respect to a pure collaborative approach. We can
conclude that the proposed technique holds the
promise of allowing collaborative-based algorithms
to scale to large data sets, still producing high-
quality recommendations.
Balabanovic, M. and Shoham, Y., 1997. Fab: Content-
Based, Collaborative recommendation.
Communications of the ACM, 40 (3), 66-72.
Buono, P., Costabile, M. F., Guida, S., Piccinno, A.,
2002. Integrating User Data and Collaborative
Filtering in a Web Recommendation Systems. In S.
Reich, M.M. Tzagarakis and P.M.E. De Bra (Eds.),
Hypermedia: Openness, Structural Awareness and
Adaptivity, LNCS, vol. 2266, 315-321. Springer, Berlin.
Burke, R., 2002. Hybrid Recommender Systems: Survey
and Experiments. In User Modeling and User-Adapted
Interaction, 12, 331-370. Kluwer Academic
Publishers, the Netherlands.
Frank, E. and Witten, I., 1998. Generating accurate rule
sets without global optimization. In 15
Conference on Machine Learning, 144–151. Morgan
Konstan, J. A., Riedl, J., Borchers, A. and Herlocker, J. L.,
1998. Recommender Systems: A GroupLens Perspective.
In Recommender Systems. Papers from 1998 Workshop.
Technical Report WS-98-08, 60-64. AAAI Press.
Pazzani, M., 1999. A Framework for Collaborative,
Content-Based and Demographic Filtering. In
Artificial Intelligence Review, 393-408.
Resnick, P. and Varian, H. R., 1997. Recommender
Systems. Communications of the ACM, 40 (3),56-58.
Sarwar, B. M., Karypis, G., Konstan, J. and Riedl, J., 2002.
Recommender Systems for Large-Scale E-Commerce:
Scalable Neighborhood Formation Using Clustering. In
Proceedings of the 5
International Conference on
Computer and Information Technology (ICCIT 2002).
Schafer, J. B., Konstan, J. and Riedl, J., 1999.
Recommender Systems in E-Commerce. In EC ’99:
Proceedings of the First ACM Conference on
Electronic Commerce, Denver, CO, 158-166.
Sebastiani, F., 2002. Machine Learning in Automated Text
Categorization. ACM Computing Surveys, 34 (1), 1–47.
Semeraro, G., Abbattista, F., Degemmis, M., Licchelli, O.,
Lops, P., Zambetta, F., 2003. Agents, Personalisation,
and Intelligent Applications. In Corchuelo, R., Ruiz
Cortés, A. and Wrembel, R. (Eds.), Technologies
Supporting Business Solutions, Part IV: Data Analysis
and Knowledge Discovery, Chapter 7, 163-186. Nova
Sciences Books and Journals.
Shardanand, U. and Maes, P., 1995. Social Information
Filtering:Algorithms for Automating’Word of Mouth’.
In Proceedings of CHI ’95, Denver, CO., 210-217.
Terveen, L. and Hill, W., 2001. Human-Computer
Collaboration in Recommender Systems. In J. Carroll
(Ed.), Human Computer Interaction in the New
Millenium, Addison-Wesley, New York.
Tuzhilin, A. and Adomavicius, G., 1999. Integrating user
behaviour and collaborative methods in recommender
systems. In
CHI' 99 Workshop Interacting with
Recommender Systems
, Pittsburgh, PA, USA.
Witten, I. H. and Frank, E., 1999. Data Mining: Practical
Machine Learning Tools and Techniques with Java
Implementations. Morgan Kaufmann Publishers, San
Yao, Y. Y., 1995. Measuring Retrieval Effectiveness
Based on User Preference of Documents. Journal of
the American Society for Information Science, 46, 2,