PROBABILITY-BASED EXTENDED PROFILE FILTERING
An Advanced Collaborative Filtering Algorithm for User-generated Content
Toon De Pessemier, Kris Vanhecke, Simon Dooms, Tom Deryckere and Luc Martens
Wireless and Cable group, Dept. of Information Technology, IBBT, Ghent University
Gaston Crommenlaan 8, Ghent, Belgium
Keywords:
Recommender system, Personalization, Collaborative filtering, Profile, User-generated Content, Algorithm.
Abstract:
The enormous offer of (user-generated) content on the internet and its continuous growth make the selection
process increasingly difficult for end-users. This abundance of content can be handled by a recommenda-
tion system that observes user preferences and assists people by offering interesting suggestions. However,
present-day recommendation systems are optimized for suggesting premium content and partially lose their
effectiveness when recommending user-generated content. The transitoriness of the content and the sparsity of
the data matrix are two major characteristics that influence the effectiveness of the recommendation algorithm
and in which premium and user-generated content systems can be distinguished.
Therefore, we developed an advanced collaborative filtering algorithm which takes into account the specific
characteristics of user-generated content systems. As a solution to the sparsity problem, inadequate profiles
will be extended with the most likely future consumptions. These extended profiles will increase the profile
overlap probability, which will increase the number of neighbours in a collaborative filtering system. In this
way, the personal suggestions are based on an enlarged group of neighbours, which makes them more precise
and diverse than traditional collaborative filtering recommendations. This paper explains in detail the proposed
algorithm and demonstrates the improvements on standard collaborative filtering algorithms.
1 INTRODUCTION
Various Web 2.0 sites (e.g. YouTube, Flickr, Digg,
Google Video. . . ) have an overwhelming bulk
of user-generated content available for online con-
sumers. Although this exploding offer can be seen
as a way to meet the specific demands and expecta-
tions of users, it has complicated the content selection
process to the extent that users are overloaded with in-
formation and risk to ‘get lost’: although there is an
abundance of content available, it is often difficult to
obtain useful and relevant content.
Traditional filtering tools, e.g. keyword-based or
filtered searches, are not capable to weed out irrele-
vant content. A second filtering based on the general
popularity (expressed by user ratings or consump-
tion patterns) can assist but requires a broad basis
of user feedback before it can make reasonable sug-
gestions. Moreover, this technique does not consider
personal preference and individual consumption be-
haviour, since only the most popular content will be
favoured by the majority of the community. This sit-
uation reinforces the role of (collaborative) filtering
tools and stimulates the development of recommen-
dation systems that assist users in finding the most
relevant content.
2 RELATED WORK
The overabundance of content and the related diffi-
culty to discover interesting content items have al-
ready been addressed in several contexts. Online
shops, like Amazon, apply collaborative filtering (CF)
to personalize the online store according to the needs
of each customer (Linden et al., 2003). Purchasing
and rating behaviour are valuable information chan-
nels for online retailers to investigate consumers’ in-
terests and generate personalized recommendations
(Karypis, 2001).
Netflix is an online, mail-based, DVD rental ser-
vice for customers in the United States. They have the
possibility to express their appreciation for a rented
movie by a star-rating mechanism on the Netflix web-
site. This simple feedback method forms the basis of
the Netflix recommender. Convinced by the poten-
tial of a good recommendation system, Netflix pub-
219
De Pessemier T., Vanhecke K., Dooms S., Deryckere T. and Martens L.
PROBABILITY-BASED EXTENDED PROFILE FILTERING - An Advanced Collaborative Filtering Algorithm for User-generated Content.
DOI: 10.5220/0002780102190226
In Proceedings of the 6th International Conference on Web Information Systems and Technology (WEBIST 2010), page
ISBN: 978-989-674-025-2
Copyright
c
 2010 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

lished a large dataset and started a competition to find
the most suitable recommendation algorithm for their
store in October 2006. In this context, many research
groups have competed to find the best movie rec-
ommendation algorithm based on the Netflix dataset
(Bell and Koren, 2007).
The introduction of digital television entailed an
increase in the number of available TV channels
and the information overload linked thereto. Con-
sequently, new standards to describe this content,
e.g. TV-Anytime (Evain and Mart
´
ınez, 2007) and
advanced electronic program guides, which simplify
the navigation and selection of TV programs, became
necessary (Lee et al., 2005). Several personalized TV
guide systems which filter and recommend TV pro-
grams according to the user’s preferences, have been
developed for set-top boxes (Zhang et al., 2005) and
personal digital recorders (Kurapati et al., 2001; Yu
and Zhou, 2004).
Besides these traditional premium content
sources (i.e. professionally-generated content),
user-generated content (like personal photos, videos,
or bookmarks) received a more prominent role on the
web in recent years. These web 2.0 applications use
more pragmatic approaches, like tagging, to annotate
content than the traditional classification systems like
TV-Anytime. Such a metadata description which is
the contribution of the whole community, also known
as a folksonomy, became very popular on the web
around 2004. Since users might tag in a different
manner and use other synonyms, the annotation
and classification of this user-generated content is
less strict, which causes an additional difficulty for
content-based recommendation systems (Golder and
Huberman, 2005).
3 COLLABORATIVE FILTERING
3.1 Traditional Collaborative Filtering
Techniques
CF techniques are the most commonly used recom-
mendation algorithms because they generally provide
better results than content-based techniques (Her-
locker et al., 1999). Most user-based CF algo-
rithms start by finding a set of customers whose pur-
chased or rated items overlap the user’s purchased and
rated items. Customers can be represented as an N-
dimensional vector of items, where N is the number
of distinct catalogue items. Purchased or rated items
are recorded in the corresponding components of this
vector. This profile vector is extremely sparse (i.e.
contains a lot of missing values) for the majority of
customers who purchased or rated only a very small
fraction of the catalogue items.
The similarity of two customers, j and k, symbol-
ized by their consumption vectors, U
j
and U
k
, can be
measured in various ways. The most common method
is to measure the cosine of the angle between the two
vectors (Sarwar et al., 2000).
Sim(
~
U
j
,
~
U
k
) = cos(
~
U
j
,
~
U
k
) =
~
U
j
·
~
U
k
||
~
U
j
|| ||
~
U
k
||
(1)
Next, the algorithm aggregates the consumed items
from these similar customers, eliminates items the
user has already purchased or rated, and recommends
the remaining items to the user (Linden et al., 2003).
An alternative for this user-based CF technique is
item-based CF, a technique that matches each of the
user’s purchased and rated items to similar items and
then combines those similar items into a recommen-
dation list. For measuring the similarity of items, the
same metrics can be employed as with the user-based
CF. Because of scalability reasons, this technique is
often used to calculate recommendations for big on-
line shops, like Amazon, where the number of users
is much higher than the number of items.
3.2 Collaborative Filtering in Sparse
Data Sets
Despite the popularity of CF, its applicability is lim-
ited due to the sparsity problem, which refers to the
situation that the consumption data in the profile vec-
tors are lacking or are insufficient to calculate reli-
able recommendations. In an attempt to provide high-
quality recommendations even when data profiles are
sparse, some solutions are proposed in literature (Pa-
pagelis et al., 2005). Most of these techniques use
trust inferences, transitive associations between users
that are based on an underlying social network, to deal
with the sparsity and the cold-start problems (Weng
et al., 2006). Nevertheless, these underlying social
networks are in many cases insufficiently developed
or even nonexistent for (new) web-based applications
that desire to offer personalized content recommenda-
tions.
Default voting is an extension to the traditional CF
algorithms which tries to solve this sparsity problem
without exploiting a social network. A default value
is assumed as ‘vote’ for items without an explicit rat-
ing or purchase (Breese et al., 1998). Although this
technique enlarges the profile overlap, it can not iden-
tify more significant neighbours than the traditional
CF approach.
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
220

A direct consequence of this sparsity problem is
that the number of similar customers, the neighbours
of the users, can be very limited in a user-based CF
technique. Furthermore, because of this sparsity, the
majority of these neighbours will also have a small
number of consumed items in their profile vectors.
Because the prospective personal recommendations
are limited to this set of consumptions of neighbours,
the variety, quality and quantity of the final recom-
mendation list might be inadequate.
A comparable reasoning is applicable to item-
based CF techniques that work on sparse profile data.
Users might have consumed a small number of items,
which in turn also have a limited number of neigh-
bouring items. Again, the CF algorithm is restricted
to a narrow set of product to generate the personal
suggestions which is disastrous for the efficiency of
the recommender.
3.3 Collaborative Filtering for
User-generated Content Systems
Because of the nature of user-generated content sys-
tems, the number of content items is in many cases
significantly bigger compared to premium content
systems. User-generated content requires less pro-
duction efforts. Accordingly, the content production
rate and the number of distinct publishers are mas-
sive. For example, YouTube enjoys 65,000 daily new
uploads (Cha et al., 2007). Due to these varied con-
tent (production) characteristics, the sparsity prob-
lem will even become worse for user-generated con-
tent. Therefore, the recommender performance might
be disappointing if the traditional CF techniques will
be ported from premium content delivery systems to
user-generated content delivery systems without any
adaptation.
We developed an advance CF algorithm that ex-
tends profiles based on the probability that an item
will be consumed in the future. These extended
profiles will increase the profile overlap probability,
which will increase the number of neighbours in a CF
algorithm. These extended profiles, in reaction to the
sparsity problem, lead to more precise and varied con-
tent recommendations.
4 PROBABILITY-BASED
EXTENDED PROFILE
One of the consequences of a sparse data matrix is
that the number of neighbours for a user/item might
be very limited in a user-based/item-based CF sys-
tem. Having no or limited neighbours leads to insuf-
ficient and imprecise recommendations. The majority
of the similarity metrics that are used in CF systems
rely on the profile overlap to determine the similar-
ity of two users/items. So, to increase the amount of
neighbours, the number of overlapping profiles has
to be increased, which can be achieved by a greater
amount of consumption behaviour. Because stimulat-
ing users to consume more content is not an option in
most cases, we opted for an artificial profile extension
based on the future consumption probability.
Our developed algorithm is an iterative process. In
the first phase, a traditional CF algorithm will try to
generate personal suggestions based on the existing
profiles. In the second phase, all the initial profiles
that do not contain a minimum number of consump-
tions will be extended. These first two stages can be
repeated in order to reach a minimum threshold for
the profile size. The extended profiles will be used
to recalculate the similarities in a third phase; and at
last, the final recommendations will be generated in
the fourth phase.
4.1 Traditional CF Recommendations
In the first phase, a traditional CF algorithm will
be used to generate standard CF recommendations.
These recommendations will be applied to extend the
sparse profiles in the next phase.
4.2 Extending Profiles
To make sparse profiles more dense, possible future
consumptions are inserted in the profile vectors, in
the second phase. These additional consumptions are
based on two information sources: the general prob-
ability and the profile-based probability that the item
will be consumed in the near future.
4.2.1 User-based
In user-based CF systems, existing user profiles will
be supplemented with the items that have the biggest
probability to be consumed by the user in the near
future. The general probability that a specific item
will be consumed by a specific user without a priori
knowledge of the user is proportionally to the current
popularity of the item. Especially for user-generated
content systems, this popularity can vary rapidly in
time. In addition, the probability that the item will be
consumed by the user can also be calculated based on
the user’s profile as a priori knowledge. This proba-
bility will be inverse proportional to the index of the
item in a personal top-N recommendation list, and can
PROBABILITY-BASED EXTENDED PROFILE FILTERING - An Advanced Collaborative Filtering Algorithm for
User-generated Content
221

be estimated by the confidence value which is calcu-
lated by the traditional CF system in phase 1. After
all, this top-N recommendation list is a prediction of
the items which the user will like/consume in the near
future.
4.2.2 Item-based
In the item-based case, item profiles, which contain
the users who consumed the item in the past, will be
supplemented with the most likely future consumers.
The general probability that a specific user will con-
sume a specific item, without any knowledge of the
item, is proportional to the present intensity of the
consumption behaviour of that user. With some a pri-
ori knowledge of that item, the calculations can be re-
peated. Then, the probability will be inverse propor-
tional to the index of the user in a top-N list of users
who are the most likely to consume the product in the
future. This list and the associated confidence values
can be generated by the results of the traditional item-
based CF (Segaran, 2007).
4.2.3 Potential Consumption Behaviour
Based on this calculated general and profile based
probability, the user or item profiles will be completed
until the minimum profile threshold is reached. How-
ever, these predicted consumptions will be marked as
uncertain in contrast to the initial assured consump-
tions. For example, for a web shop, the real purchases
correspond to a 1, which refers to a 100% guaranteed
consumption, while the potential future consumptions
are represented by a decimal value between 0 and 1,
according to the probability value, in the profile vec-
tor. This second phase can consist of several succes-
sive iterations to complete the profiles.
4.3 Recalculating the Similarities
Based on these extended profile vectors, the similar-
ities will be recalculated with the chosen similarity
metric, e.g. the cosine similarity (equation 1), in a
third phase. Because of the added future consump-
tions, the profile overlap and accordingly the number
of neighbours will be increased compared to phase 1.
For the item-based case, these similarities can e.g. be
used for a ‘related products’ section in online shops.
4.4 Generating Recommendations
To produce personal suggestions, a recommendation
vector will be generated based on these extended pro-
file vectors, in a fourth phase. For a user-based algo-
rithm, the recommendation vector, R
j
, for user j can
be calculated as:
~
R
j
=
∑
N
k=1,k6= j
~
U
k
· Sim(
~
U
j
,
~
U
k
)
∑
N
k=1,k6= j
Sim(
~
U
j
,
~
U
k
)
(2)
where U
j
and U
k
represent the consumption vectors of
users j and k, which might contain real values. Subse-
quently, the top-N recommendations are obtained by
taking the indices of the highest components of the
recommendation vector, R
j
, and eliminating the items
which are already consumed by user j in the past.
5 EVALUATION DESIGN AND
MEASUREMENT
5.1 Data Set
To estimate the effectiveness of personal recommen-
dations, two different evaluation methods are possi-
ble. On the one hand, online evaluations measure the
user interactivity (e.g. clicks, buying behaviour) with
the personal suggestions on a running service. Offline
evaluations, on the other hand, use a test set with con-
sumption behaviour which has to be predicted based
on a training set with consumption history. Although
online evaluation methods are the most close to real-
ity, we opted for an offline evaluation based on data
sets because such an evaluation is fast, reproducible
and commonly used in recommendation research.
Therefore, we compared the proposed recommen-
dation algorithm with the traditional CF algorithm
based on evaluation metrics which are generated by
an offline analysis using a data set with consumption
behaviour. Because the datasets that are commonly
used to benchmark recommendation algorithms (e.g.
Netflix, Movielens, Jester) contain to few ‘small’ pro-
files and handle only premium content, we evaluated
our algorithm on a dataset of PianoFiles
1
, a user-
generated content site that offers users the opportu-
nity to manage their collection of music sheets. The
logged consumption behaviour for constructing pro-
files consists of individual additions of music sheets
to personal collections. The full data set contains
401,593 items (music sheets), 80,683 active users and
1,553,586 distinct consumptions in chronological or-
der.
5.2 Evaluation Method
For evaluation purposes, we use 50% of the consump-
tions that are the most recent ones as the test set and
1
http://www.pianofiles.com/
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
222

use the remaining 50% of the consumption records as
the input data. In order to study the performance of
the algorithm under data of different sparsity levels,
we form ten different training sets by selecting the
first 10%, 20%, 30%, until 100% of the input data.
The recommendation algorithm uses these different
training sets in successive iterations to generate per-
sonal suggestions.
As commonly done for the evaluation of recom-
mendations under sparse data (Huang et al., 2004),
the test set is first filtered to only include consump-
tions that are possible to predict with the input data
as a priori knowledge. A consumption of a sheet that
is not contained in the input data or a consumption
of a user without any consumption behaviour in the
input data is not possible to predict with CF tech-
niques. All users in this filtered test set are included
into a set of target consumers. For each of these con-
sumers, the algorithm generates five ordered lists of
10, 20, 30, 40 and 50 recommendations respectively,
which will be compared with the test set. This off-
line evaluation methodology, in which a data set is
chronologically splitted in training set and test set, is
commonly used for evaluating recommendation algo-
rithms (Hayes et al., 2002).
6 EVALUATION
6.1 Evaluation Metrics
One of the most used error metrics is the Root Mean
Squared Error (RMSE) (Herlocker et al., 2004; Cam-
pochiaro et al., 2009), which is also adopted by the
official Netflix contest. However, the Netflix contest
is mainly focused on predicting accurate ratings for
an entire set of items, while web-based (e-commerce)
applications are most interested in providing the users
with a short recommendation list of interesting items
(Campochiaro et al., 2009). To evaluate this top-N
recommendation task, i.e. a context where we are
not interested in predicting user ratings with preci-
sion, but rather in giving an ordered list of N attrac-
tive items to the users, error metrics are not mean-
ingful. Therefore, information-retrieval classification
metrics, which evaluate the quality of a short list of
recommendations, are the most suitable.
The most popular classification accuracy met-
rics are the precision and recall (Campochiaro et al.,
2009). The precision is the ratio of the number of
recommended items that match with future consump-
tions, and the total number of recommended items.
In offline evaluations, the consumptions of the test
set represent the future consumptions, and the recom-
mendations that match with these consumptions are
called the relevant recommendations.
Precision =
# Relevant recommendations
# Recommendations
(3)
The recall stands for the ratio of the number of rele-
vant recommendations and the total number of future
consumptions. Only these future consumptions are
considered as relevant items for the end-users in off-
line evaluations.
Recall =
# Relevant recommendations
# Relevant items
(4)
It has been observed that precision and recall are in-
versely related and are dependent on the length of the
result list returned to the user (Herlocker et al., 2004).
When more items are returned, precision decreases
and the recall increases. Therefore, in order to under-
stand the global quality of a recommendation system,
we may combine precision and recall by means of the
F1-measure
F1 =
2 · Precision · Recall
Precision + Recall
(5)
6.2 Bench-marked Algorithms
Because item-based algorithms generally achieved a
very low performance on the PianoFiles dataset, we
did not include any item-based technique in the eval-
uation. This poor performance is mainly due to the
nature of the dataset, which contains a much larger
number of items than users. Therefore, forming item
neighbourhoods is actually much more difficult than
forming user neighbourhoods (Huang et al., 2004).
Furthermore, with this proportion of users and items,
there is a big risk that item-based algorithms will
trap users in a ‘similarity hole’, only giving excep-
tionally similar recommendations; e.g. once a user
added a sheet of Michael Jackson to her collection,
she would only receive recommendations for more
Michael Jackson sheets (McNee et al., 2006).
Compared to this item-based CF, a user-based
strategy achieves much better results on the Pi-
anoFiles dataset. In a first evaluation, we bench-
marked this standard user-based CF algorithm
(UBCF), which operates on the initially existing pro-
files (Segaran, 2007), against the user-based ver-
sion of our probability-based extended profile filter-
ing (UBExtended), which extends the sparse profiles
before generating the actual recommendations. For
this performance evaluation, the UBExtended algo-
rithm expands sparse profiles to a target size of 6 con-
sumptions and the cosine similarity (equation 1) is
used as a measure to compare profile vectors in both
algorithms.
PROBABILITY-BASED EXTENDED PROFILE FILTERING - An Advanced Collaborative Filtering Algorithm for
User-generated Content
223

0.04
0.05
0.06
0.07
10Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastr ainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
0.04
0.05
0.06
0.07
30Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastrainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
0.04
0.05
0.06
0.07
50Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastr ainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
Figure 1: The evaluation of the UBCF and UBExtended
algorithm based on the initial training set.
6.3 Complete Training Set
The graphs in Figure 1 illustrate that the evaluation
metrics for these two algorithms increase, while more
training data becomes available. As the size of the
training set increases, more user behaviour becomes
available, including the behaviour of new users for
which no information was available in the first part(s)
of the training set. This additional data enables the
generation of recommendations for more users, which
explains the increasing recall value.
Besides information of new users, supplementary
training data will also contain consumption behaviour
of users who have already an initial music collection.
This extra information can refine the user profiles,
which leads to a higher precision. However, after
the profile size reached a critical point, supplemen-
tary training data has no more additional information
value, which leads to a stagnating precision value.
Moreover, the recommendations for new users, which
will generally have a lower precision value because of
a limited early profile, will enhance this stagnation ef-
fect. At last, the F1 metric will follow the progress of
this precision and recall graph closely because of its
definition. Besides these general trends, the graphs
indicate that the UBExtended algorithm exceeds the
standard UBCF in all three evaluation metrics (pre-
cision, recall and F1) and for different sizes of the
recommendation list. This improvement is especially
noticeable for small training sets, which mainly con-
sist of sparse user profiles.
6.4 Filtered Training Set: Sparse
Profiles Only
To illustrate this superiority of the UBExtended al-
gorithm for sparse profiles, a second evaluation is
performed. To scrutinize the recommender perfor-
mance for the subset of users with a sparse profile,
the training set is submitted to an extra filter. This fil-
ter removes all the users with more than x consump-
tions from the training set to simulate the situation of
a very novel content delivery system without ‘well-
developed’ user profiles. In accordance with the first
evalution in which we extended sparse profiles to a
target size of 6 consumptions, we chose in this sec-
ond analysis for a filter that removes all users with a
profile that is larger than this target size. In this way,
a standard UBCF that operates on a dataset with only
sparse profiles (profile size ≤ 6 consumptions), will
be compared with the UBExtended algorithm which
broadens these profiles to the target size (profile size
= 6 consumptions) before generating recommenda-
tions. Given the long-tail distribution of the profile
size in content delivery systems (Figure 2), this sub-
set of sparse-profile consumers, constitutes a consid-
erable segment of the system users.
Since, the filter modifies the training and test sets,
the absolute values of this evaluation can not be com-
pared to the absolute values of the first evaluation,
which is based on the unfiltered data sets. How-
ever, the differences between the UBCF and UBEx-
tended algorithm in this second evaluation, illustrated
in Figure 3, compared with the differences between
these algorithms in the first evaluation (Figure 1), con-
firm that the performance improvement of the UBEx-
tended algorithm increases for more sparse datasets.
Finally, the graphs of this second evaluation show
that for small training sets the precision might slightly
fluctuate due to unsufficient data and many new users.
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
224

20000
25000
30000
Number
ofusers
Histogramofprofilesize
0
5000
10000
15000
1357911131517192123252729
Numberofconsumptions
Figure 2: Histogram of the consumption profile size for the
PianoFiles dataset.
004
0.05
0.06
0.07
0.08
10Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0
.
04
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastrainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
004
0.05
0.06
0.07
0.08
30Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0
.
04
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastrainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
004
0.05
0.06
0.07
0.08
50Recommendations
UBCFPrecision
UBCFRecall
UBCF F1
0
0.01
0.02
0.03
0
.
04
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Partoftheinputdatathatisusedastrainingset
UBCF

F1
UBExtended
Precision
UBExtended
Recall
UBExtendedF1
Figure 3: The evaluation of the UBCF and UBExtended
algorithm based on the training set, which contains only
sparse profiles.
7 OPTIMISATION & DRAWBACK
Since the parameters of the UBExtended algorithm
are not yet optimized, the performance difference be-
tween the two bench-marked algorithms might even
increase considerably. The UBExtended algorithm
extends sparse profiles until each profile contains a
predefined number of consumptions. This target pro-
file size is an important parameter that has to be op-
timized in function of the performance metrics. Al-
though we have chosen a fixed value of 6 consump-
tions for the extended profiles in our evaluation, we
believe this parameter might be a function of gen-
eral dataset statistics, namely the overall sparsity of
the data matrix, the number of items and the number
of users. Moreover, the procedure of extending the
profiles, which is based on general and profile-based
influences, can be fine-tuned. An optimal balance be-
tween general and profile-based information to extend
the profiles might result in more precise recommen-
dations. Finally, there are some typical CF parame-
ters which have to be determined such as the similar-
ity measure and the number of neighbours which are
used to calculate the recommendations.
Unfortunately, the accuracy improvement ac-
quired with the UBExtended algorithm is associated
with an extra calculation cost. Compared to the stand-
ard UBCF, the UBExtended algorithm consists of 2
extra phases: extending the profiles and recalculat-
ing the similarities after this extension. Especially
the similarity calculation can be time-consuming due
to its quadratic nature and therefore may pose prob-
lems for systems that calculate the recommendations
in real-time (i.e. generating recommendations when
the web-page is requested). However, since most
recommender systems schedule the calculations and
update their recommendations periodically, the addi-
tional calculation time is no major problem.
8 CONCLUSIONS & FUTURE
WORK
In this research, we developed an advanced collabo-
rative filtering algorithm which takes into account the
specific characteristics of user-generated content sys-
tems. The algorithm extends sparse profiles with the
most likely future consumptions based on general and
personal consumption behaviour. Our experimental
study, using a dataset from the user-generated content
site PianoFiles, showed that the user-based version of
the proposed algorithm achieves better performances
than the standard user-based collaborative filtering al-
gorithm, especially on sparse data sets. These results
PROBABILITY-BASED EXTENDED PROFILE FILTERING - An Advanced Collaborative Filtering Algorithm for
User-generated Content
225

proof that there is a need to adapt traditional collab-
orative filtering techniques to the specific character-
istics, such as the sparsity, of user-generated content
websites. In future research, we will optimize the
algorithm parameters to further improve the perfor-
mance results. Besides we will investigate if the prin-
ciple of profile extension is applicable in other types
of collaborative filtering algorithms.
ACKNOWLEDGEMENTS
We would like to thank the Research Foundation -
Flanders (FWO), for the research position of Toon De
Pessemier (Aspirant FWO). We would also like to ex-
press our appreciation to Thomas Bonte, the founder
of PianoFiles, for putting the data set of his website at
our disposal.
REFERENCES
Bell, R. M. and Koren, Y. (2007). Lessons from the netflix
prize challenge. SIGKDD Explor. Newsl., 9(2):75–79.
Breese, J. S., Heckerman, D., and Kadie, C. (1998). Empir-
ical analysis of predictive algorithms for collaborative
filtering. pages 43–52. Morgan Kaufmann.
Campochiaro, E., Casatta, R., Cremonesi, P., and Turrin, R.
(2009). Do metrics make recommender algorithms?
Advanced Information Networking and Applications
Workshops, International Conference on, 0:648–653.
Cha, M., Kwak, H., Rodriguez, P., Ahn, Y.-Y., and Moon, S.
(2007). I tube, you tube, everybody tubes: analyzing
the world’s largest user generated content video sys-
tem. In IMC ’07: Proceedings of the 7th ACM SIG-
COMM conference on Internet measurement, pages
1–14, New York, NY, USA. ACM.
Evain, J.-P. and Mart
´
ınez, J. M. (2007). Tv-anytime
phase 1 and mpeg-7. J. Am. Soc. Inf. Sci. Technol.,
58(9):1367–1373.
Golder, S. and Huberman, B. A. (2005). The structure of
collaborative tagging systems. ArXiv Computer Sci-
ence e-prints.
Hayes, C., Massa, P., Avesani, P., and Cunningham, P.
(2002). An on-line evaluation framework for rec-
ommender systems. In In Workshop on Personaliza-
tion and Recommendation in E-Commerce (Malaga.
Springer Verlag.
Herlocker, J. L., Konstan, J. A., Borchers, A., and Riedl,
J. (1999). An algorithmic framework for perform-
ing collaborative filtering. In SIGIR ’99: Proceedings
of the 22nd annual international ACM SIGIR confer-
ence on Research and development in information re-
trieval, pages 230–237, New York, NY, USA. ACM.
Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl,
J. T. (2004). Evaluating collaborative filtering recom-
mender systems. ACM Trans. Inf. Syst., 22(1):5–53.
Huang, Z., Zeng, D., and H.Chen (2004). A link analy-
sis approach to recommendation with sparse data. In
AMCIS 2004: Americas Conference on Information
Systems, New York, NY, USA.
Karypis, G. (2001). Evaluation of item-based top-n recom-
mendation algorithms. In CIKM ’01: Proceedings of
the tenth international conference on Information and
knowledge management, pages 247–254, New York,
NY, USA. ACM.
Kurapati, K., Gutta, S., Schaffer, D., Martino, J., and Zim-
merman, J. (2001). A multi-agent tv recommender. In
Proceedings of the UM 2001 workshop Personaliza-
tion in Future TV, pages 13–14.
Lee, H., Yang, S.-J., Kim, J.-G., and Hong, J. (2005).
Personalized tv services based on tv-anytime for pdr.
Consumer Electronics, 2005. ICCE. 2005 Digest of
Technical Papers. International Conference on, pages
115–116.
Linden, G., Smith, B., and York, J. (2003). Amazon.com
recommendations: item-to-item collaborative filter-
ing. Internet Computing, IEEE, 7(1):76–80.
McNee, S. M., Riedl, J., and Konstan, J. A. (2006). Being
accurate is not enough: how accuracy metrics have
hurt recommender systems. In CHI ’06: CHI ’06 ex-
tended abstracts on Human factors in computing sys-
tems, pages 1097–1101, New York, NY, USA. ACM.
Papagelis, M., Plexousakis, D., and Kutsuras, T. (2005).
Trust Management. Springer Berlin / Heidelberg.
Sarwar, B., Karypis, G., Konstan, J., and Riedl, J.
(2000). Analysis of recommendation algorithms for
e-commerce. In EC ’00: Proceedings of the 2nd ACM
conference on Electronic commerce, pages 158–167,
New York, NY, USA. ACM.
Segaran, T. (2007). Programming collective intelligence.
O’Reilly.
Weng, J., Miao, C., Goh, A., Shen, Z., and Gay, R. (2006).
Trust-based agent community for collaborative rec-
ommendation. In AAMAS ’06: Proceedings of the fifth
international joint conference on Autonomous agents
and multiagent systems, pages 1260–1262, New York,
NY, USA. ACM.
Yu, Z. and Zhou, X. (Feb 2004). Tv3p: an adaptive assis-
tant for personalized tv. Consumer Electronics, IEEE
Transactions on, 50(1):393–399.
Zhang, H., Zheng, S., and Yuan, J. (May 2005). A personal-
ized tv guide system compliant with mhp. Consumer
Electronics, IEEE Transactions on, 51(2):731–737.
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
226