WEB INFORMATION RECOMMENDATION MAKING BASED
ON ITEM TAXONOMY
LI-Tung Weng, Yue Xu, Yuefeng Li and Richi Nayak
Faculty of Information Technology, Queensland University of Technology, 4001 Queensland, Australia
Keywords: Recommender System, Taxonomy, Ecommerce, Cold-start Problem.
Abstract: Recommender systems have been widely applied in the domain of ecommerce. They have caught much
research attention in recent years. They make recommendations to users by exploiting past users’ item
preferences, thus eliminating the needs for users to form their queries explicitly. However, recommender
systems’ performance can be easily affected when there are no sufficient item preferences data provided by
previous users. This problem is commonly referred to as cold-start problem. This paper suggests another
information source, item taxonomies, in addition to item preferences for assisting recommendation making.
Item taxonomy information has been popularly applied in diverse ecommerce domains for product or
content classification, and therefore can be easily obtained and adapted by recommender systems. In this
paper, we investigate the implicit relations between users’ item preferences and taxonomic preferences,
suggest and verify using information gain that users who share similar item preferences may also share
similar taxonomic preferences. Under this assumption, a novel recommendation technique is proposed that
combines the users’ item preferences and the additional taxonomic preferences together to make better
quality recommendations as well as alleviate the cold-start problem. Empirical evaluations to this approach
are conducted and the results show that the proposed technique outperforms other existing techniques in
both recommendation quality and computation efficiency.
1 INTRODUCTION
Recommender systems have been an active research
area for more than a decade, and many different
techniques and systems with distinct strengths have
been developed (Montaner et al. 2003). Among all
these different recommendation techniques,
collaborative filtering is perhaps the most successful
and widely applied technique for building
recommender systems(Deshpande and Karypis
2004; Schafer et al. 2000). In general, collaborative
filtering based recommenders recommend items that
are commonly preferred by users with similar item
preferences to a target user. Therefore, the
recommendation quality of the collaborative filtering
technique depends upon the number of users with
similar preferences to the target user. If there are
only few users in the dataset with similar
preferences to the target user, then the standard
collaborative filtering technique will not be able to
suggest quality recommendation to the user. This
issue, commonly referred to as cold-start
problem(Schein et al. 2002), usually happens when
the system is newly built (there is no initial data in
the dataset), or when there is no data available for a
new target user(Middleton et al. 2002).
A commonly used approach to alleviate the cold-
start problem is to take item content information into
consideration in recommendation making. That is,
when it is not possible to form a neighbourhood for
a target user, content based techniques can be used
to mine the item contents preferred by the target user,
and based on the preferred item contents the
recommendations can be generated by finding items
with similar contents preferred by the target user
(Burke 2002; Sarwar et al. 2000). However, because
most of the content based techniques represent item
content information as word vectors and maintain no
semantic relations among the words, therefore the
result recommendations are usually very content
centric and poor in quality(Adomavicius et al. 2005;
Burke 2002; Ferman et al. 2002; Sarwar et al. 2000).
To improve the content based techniques, the
content information for the items should be captured
in more sophisticated ways so that associations
among items can be measured by their content
semantic meanings rather than simple keywords
mappings.
20
Weng L., Xu Y., Li Y. and Nayak R. (2008).
WEB INFORMATION RECOMMENDATION MAKING BASED ON ITEM TAXONOMY.
In Proceedings of the Tenth International Conference on Enterprise Information Systems - SAIC, pages 20-28
DOI: 10.5220/0001695100200028
Copyright
c
SciTePress
In this paper, we propose a novel
recommendation making approach, namely Hybrid
Taxonomy based Recommender (HTR), which
generates item recommendations based on both
users’ item preferences and item taxonomic
preferences. The notion of item taxonomy
information is used in our system in place of
standard item content information, that is, instead of
using keywords vectors to represent items, our
system describes items based on taxonomic topics
extracted from a tree-like taxonomy structure. The
item taxonomy information is useful for
encapsulating item content semantics as it allows
items with different topics to be related if they share
common supper topics. Hence, not only the use of
item taxonomy can significantly alleviate the cold-
start problem, but it can also improve
recommendation quality by reducing the content
centric issue. The relationship between the item
preferences and the item taxonomic preferences is
also investigated in this paper. Based on our study
and experiments, we suggest that when a set of users
shares similar item preferences, they might also
share similar item taxonomic preferences. The HTR
technique utilizes the proposed relation to achieve
competitive computation efficiency and
recommendation performance. For the applicability
concern, as item taxonomy information is available
for most e-commerce sites and standardization
organizations, HTR can be easily applied and
adopted to a wide range of domains. Moreover, HTR
can also adopt the implicit user preference
information (in addition to the standard explicit user
preferences) to further enhance its recommendation
quality in cold-start environments.
2 RELATED WORK
Much research has suggested that the cold-start
problem can be alleviated by combining
collaborative filtering and content based techniques
together (Burke 2002; Ferman et al. 2002; Park et al.
2006; Schein et al. 2002). However, because part of
the recommendation process for these hybrid
recommenders is content-based, the generated
recommendations may be excessively content
centric and lack of novelty(Middleton et al. 2002;
Ziegler et al. 2004). Hence, semantic and ontology
based techniques have been suggested to improve
the recommendation generality for the content based
filtering. Middleton(Middleton et al. 2002)
suggested an ontology based recommender which
uses external organizational ontology (e.g.
publication and authorship relationships, projects
and project membership relationships, etc.) to solve
the cold start problem. However, as the Middleton’s
technique is mainly designed for recommending
research papers and documents, and also relies on a
specific organizational ontology, therefore it is not
easy to adopt this method for general recommenders.
On the other hand, Ziegler(Ziegler et al. 2004)
proposed a taxonomy-driven product recommender
(TPR), it utilizes a general tree structured product
taxonomy to enhance its recommendations. Due to
the simplicity of the taxonomy structure, Ziegler’s
technique is considered widely applicable to
different domains(Ziegler et al. 2004). To the best of
our knowledge, Middleton and Ziegler’s techniques
are the only two works bearing traits similar to the
proposed HTR technique. HTR employs similar tree
structured taxonomy to TPR, and therefore it inherits
TPR’s generality advantage. However, while TPR
only considers implicit item preferences for making
recommendations, HTR utilizes the relationship
between users’ explicit item preference and implicit
taxonomic preferences for recommendation making,
therefore yields better recommendation
performances. Moreover, HTR adopts item-based
collaborative filtering paradigm (Deshpande and
Karypis 2004) in contrast to TPR’s user-based
collaborative filtering. Item-based collaborative
filtering allows most computations to be done offline.
Therefore, the computation efficiency of online
recommendation generation can be improved.
3 PROPOSED APPROACH
The idea behind HTR is intuitive. It firstly finds a set
of users with similar preferences to a given target
user, and then extracts taxonomy topics that are
popularly and uniquely preferred by these users.
Finally, HTR estimate the target user’s preference to
a candidate item by combining user item preferences
with taxonomy topic preferences.
This section is divided into five parts. In Section
3.1, the basic system model and general notations
used throughout this paper are described. In Section
3.2, we discuss the implicit relation between users’
item preferences and taxonomic preferences. The
technique for taxonomic preference extraction is
described in Section 3.3. At last, Section 3.4 details
the proposed HTR method.
3.1 System Model
We envision a world with a set of users
WEB INFORMATION RECOMMENDATION MAKING BASED ON ITEM TAXONOMY
21
,
,…,
and a set of items
,
,…,
. Each user  is associated with a
set of rated items 

. Based on the
different rating methods, we can divide these items
into implicitly rated items 




and explicitly rated items 



. A
user can rate an item implicitly or explicitly, but not
both (i.e.



).
In explicit ratings, users express their
preferences to items in numeric form, that is, the
value 0 indicates minimal satisfaction and 1
indicates maximum satisfaction. We use
, to denote user ’s rating to item


, such that 0,1.
HTR uses taxonomy based descriptors to
describe items. Specifically,  
,
,…,
denotes a set of descriptors
characterizing any item ’s taxonomy, where
||1. A taxonomy descriptor is a sequence of
ordered taxonomic topics, denoted by 

,
,…,
where , . The topics
within a descriptor are sequenced so that the former
topics are super topics of the latter topics,
specifically,
is the direct super topic for

where 0. A super topic covers a broader
concept than its sub-topics, and a topic can have
more than one direct sub-topics. Thus, it is easy to
envision that the taxonomy topics are stored in a
tree-like structure, and the tree structure formed with
the taxonomy topics is referred as the taxonomy tree,
and all item descriptors are paths that are extracted
from the root to a leaf node on the tree.
Let be the set of all taxonomy topics,
|,
,
and : 2
be a map
from to 2
that retrieves all direct sub-topics

 for topics
. Based on , we define
a partial order on the taxonomy topic set to
differentiate between super topics and sub-topics.

,
, if

, then
and
have the
relationship , i.e.,

require that
 for all
,
, . With this
requirement and the map , we can recursively
extract the taxonomy tree structure from the set .
Moreover, as in standard tree structures, the
taxonomy tree has exactly one top-most element
with zero in-degree representing the most general
topic, it is denoted by in this paper. By contrast,
for these bottom-most elements with zero out-degree,
they are denoted by and represent the most
specific topics. In our system, for any item
descriptor
,
,…,
, it is required

and
.
3.2 Cluster-based User Neighbourhood
In HTR, cluster based neighbourhood formation is
adopted to ensure the computation efficiency. In
order to form the user neighbourhoods or clusters, a
similarity measure for computing user similarities is
essential. In HTR, we adopted the correlation
measure described in (Breese et al. 1998) to compute
the item preference similarity between two users
,
 as given in Equation (1).

,

∑

,


,

∑

,


,

(1)
where is an item rated explicitly by both
and
,
that is 





.  denotes
the average explicit ratings made by .
Based on Equation (1), can be divided into a
set of clusters 

,
,…,
, such that


 and


. For the sake of
convenience, let  denote the
cluster which contains user u. Because the clusters
are constructed based on users’ item preference
similarity, users within the same cluster will have
similar item preferences. In this paper, we take a
further investigation to suggest the following
assumption:
users within the same neighbourhood or cluster
sharing similar item preferences may share similar
taxonomic preferences and interests
The idea behind the assumption is that the users
within one cluster should have apparent similar
taxonomic focus and the taxonomic focuses of the
users in different clusters should be different. In this
paper, we use information gain to measure the
certainty of taxonomy focus of a user set, and
empirically demonstrate the validity of the above
assumption by using information gain measure.
When the information gain is high, it indicates that
the certainty of the taxonomic focuses of user
clusters is high. Therefore we can use information
gain to investigate whether different clusters have
apparent taxonomic focuses and the taxonomic
focuses are different in different user clusters. The
adapted information gain can be calculated as below:

Pr   

2
where Pr  is the probability that an item rating
is made by a user in cluster .
is the
information entropy for a given user space. The
concept of information entropy is adapted in this
paper to measure the degree of taxonomic focus in a
user set (i.e. a cluster or a neighbourhood). If the
ICEIS 2008 - International Conference on Enterprise Information Systems
22
information entropy is high for a user set, then there
is no apparent taxonomic focuses in the set (i.e.
users in the set prefer all taxonomy topics equally),
and vice versa. The information entropy formula is
depicted below:
Pr ,

Pr ,
,
3
In the entropy equation, Pr p,U
denotes the
probability that the users in the user set U
U are
interested in the taxonomy topic p. For a given
clustering UC
uc
,uc
,…,uc
, if H

are low
which means the taxonomic focuses are apparent in
cluster uc , according to Equation (2), the
information gain is high.
The effect of user clustering on taxonomy
information gain is depicted in Table 1. This result is
obtained by using k-means clustering technique to
divide 278,858 users in “Book-Crossing” dataset
(www.informatik.uni-freiburg.de/~cziegler/BX/)
into 100 clusters according to their explicit ratings.
We have tried to produce different number of
clusters for the dataset (i.e. different values for k),
and we have found by setting k to 100 (i.e. 100
clusters) can produce clusters with reasonable
qualities.
Table 1: The effect of user clustering on taxonomy
information gain.
Explicit
Ratings
Explicit +
Implicit Ratings
users clusters formed based
on user ratings
0.823 0.458
Randomly formed user
clusters (baseline)
-0.385 -0.319
Our first experiment is to show if user clusters
have stronger taxonomic focuses than the entire
dataset when only explicit ratings are considered. It
is shown in the first column of Table 1, the result
information gain is 0.823, which is a big increase
when comparing it with the information gain
obtained from the randomly formed cluster
partitions (i.e. -0.385). This result shows that, by
clustering users with their explicit ratings, each user
cluster has its own taxonomic focuses.
Because our clusters are generated based on only
explicit ratings, it might be unfair if we only
consider explicit ratings in calculating taxonomy
information gain. Hence, we further include the
implicit ratings in computing taxonomy information
gain. With identical cluster settings, we still get a
strong information gain increase (i.e. 0.458) when
comparing to the information gain obtained from the
random formed clusters (i.e. -0.319). Based on the
information gain analysis, we can conclude that
users within the same clusters not only share
similar item preferences, but they also share
similar taxonomic preferences.
3.3 Taxonomic Preferences Extraction
For each cluster , we build a cluster based
taxonomy tree similar to the global taxonomy tree
defined in Section 3.1. Formally, we define the
cluster based topic set:

|,
,

,
and



for topics 

extracts the
direct sub-topics of .
Using the similar way described in Section 3.1,
with the map

, we can construct a local
taxonomy tree from a cluster . With the local
cluster based taxonomy tree, we can then find the
frequent and distinct topics for each cluster. We
measure the distinctness of a topic within a local
cluster uc in accordance to the global user set by:


,

0, _, 
_,
_,
,
4
where _,
is the number of user ratings
to items involving taxonomy topic within a given
user set
. is a user defined constant, it is
used to filter out topics that are not popularly
interested by users. In this paper, is set to 50. So
topics need to be involved in at least 50 ratings in
order to get a reasonable score.
The higher the topic score, the higher the
possibility the taxonomy topic is unique to a cluster.
Based on the topic score, the topics with their topic
scores higher than a predefined threshold are chosen
as the hot topics for that cluster. We denote the hot
topic set by:
_
,
|

,_,

 (5)
where is the user defined threshold. In our
experiment, is set to 0.6. Figure 1 shows the
average number of topics left for each cluster for
different threshold settings.
For the “Book-Crossing” dataset there are
originally 10746 topics in the entire dataset. After
user clustering, the average number of topics per
cluster is around 3164.12. The ratio of the topic
number in the clusters out of the topic number in the
entire dataset is about 0.29. This ratio suggests that
different clusters may have very different taxonomy
topics. Moreover, after we increase the topic score
WEB INFORMATION RECOMMENDATION MAKING BASED ON ITEM TAXONOMY
23
threshold ζ, the ratio decreases drastically (e.g. when
ζ0.68, the entire dataset has 530 topics and the
average number of topics per cluster is 5.9, the ratio
is only 0.01.). This observation further strengthens
the conclusion that we made about cluster taxonomic
focuses as detailed in Section 3.2.
Figure 1: Average number of hot topics per cluster given
different minimal topic score (ζ).
3.4 Hybrid Taxonomy Recommender
In this section, we describe the proposed Hybrid
Taxonomy based Recommender (HTR) that
incorporates the hot topic set described in Section
3.3 with the item-based collaborative filtering (item-
based CF) to improve recommendation quality.
HTR generates item recommendations by
combining the estimates to item preferences and the
estimates to taxonomy preferences. We firstly
explain the item-based CF technique used in HTR to
estimate item preferences. Item-based CF
recommends item to user based on the item
similarity between and the items that have been
rated by based on user ratings to these items. The
similarity between two items is computed based on
user explicit ratings as defined below:
_
,















(6)
where
is a simplified form for ,
representing user ’s rating to item
,
is the
average rating for
’ over the users in

, and

is
the set of users who have rated both
and
.

is
defined as:

|
,



Note, it is possible that two items are never rated
by more than one user, i.e.

 . In such case,
_
,
returns a special value  which
is a label indicating “Not Computable”.
As mentioned above, the estimate of the
preference to item to user is based on the
similarities between and the items



rated by the user , where 
. In order to achieve
it, we need to find the target user’s rated items which
are computable with the target item . That is,

,



|_,

Finally, user ’s item preference prediction to
item t is computed as below:
,
_,
,
|_,|
,
(7)
where 0
,
1 .
In order to improve the recommendation quality
(especially in cold start situations), HTR also checks
whether the taxonomy of the candidate items is
preferred by the target user. We use
,
to denote
the prediction of user ’s taxonomic preference to
item t, and it can be computed as below:
,





,
,
|
|
0
0, 
(8)
where

|,
_
,
is the set of ’s topics that are hot topics of the
cluster which contains u. The idea behind the
computation of taxonomic preference score is
straightforward. We firstly check if any of the target
item t’s taxonomy topics are hot topics of the user
u’s neighbourhood (i.e. 
). If the item’s
topics are not hot topic of 
, then we
suggest that the user is not interested in the item’s
taxonomy, hence 0 will be given as the taxonomy
score. If the item’s topics are in the hot topic set,
then among these matched hot topics (
|
|
can be
greater than 1), the maximum hot topic score is
chosen as t’s taxonomy score.
It should be mentioned that the hot topics
calculated by Equation (5) represent the taxonomic
focuses of the users in a cluster. That means the
topics in represent cluster level taxonomic focuses
commonly preferred by the users in that cluster but
not particularly for any individual user. There are
two reasons for doing so. Firstly, cluster level
taxonomic preferences can be pre-computed offline,
therefore it ensures the computation efficiency of the
proposed technique. Secondly, since the cluster level
taxonomic preferences cover the taxonomic interests
of all the users in one cluster, for the target user, by
recommending items with topics commonly
preferred by the users in the cluster, the
recommender can recommend items with a wider
range of topics including the topics which may not
be particularly preferred by the target user but
preferred by the users in this cluster and thus the
recommendation quality can be improved.
0
500
1000
1500
2000
2500
3000
3500
0
0,08
0,16
0,24
0,32
0,4
0,48
0,56
0,64
0,72
0,8
0,88
0,96
averagenumberofhot
topicspercluster
minimaltopicscore
ICEIS 2008 - International Conference on Enterprise Information Systems
24
In order to recommend a set of items to a
target user , we firstly form a candidate item
list containing all items rated by ’s neighbors but
not yet rated by . Next, for each item in the
candidate list, we compute the item preference score
and taxonomic preference score for the item. The
proposed preference score for each candidate item
can then be computed by combining the item
preference score (
,
) and the item taxonomic
preference score (
,
) together. Finally, candidate
items with highest preference scores are
recommended to the user, and these recommended
items are sorted by the ranking values. The complete
algorithm is listed below:
Algorithm _,.
where  is a given target user
is the number of items to be recommended
1) SET





\

 ,
the candidate item list
2) FOR EACH 
3) SET 
,

,
1
,
4) END FOR
5) Return the top items with highest 
,
scores to .
From line (3) of the algorithm we can see that
the predicted score for an item is computed by a
linear combination of item preference score η
,
and
topic preference score
,
. The coefficient ,
computed by Equation (9) below, in the formula is
used to adjust the weights of
,
and
,
:





(9)
where 
|,|
|

|
and 01 is a user
controlled variable. ω is the ratio between the
number of the items that are commonly rated with
item t by u and other users and the number of the
items rated by u.
In Equation (9), ω reflects the quality confidence
of η
,
, because the more the target user’s past rated
items related to the target item, the higher the
accuracy of the item preference prediction (i.e. η
,
)
will be. When ω increases α will increase too, thus
η
,
will receive higher weight in the final score (i.e.
rank
,
). Variable , on the other hand, is used to
adjust the weights of ω in α, thus, if is large (e.g.
0.9)
,
will still receive high weight even is
small.
The value of is automatically adjusted along
with the change of the number of users who
commonly rated a given item . The higher the value
of the more the users who commonly rated the
item (i.e., is high which indicates a normal
situation without severe cold start problems) and
thus the item preference
,
estimated based on
these users’ rating data becomes more important and
reliable. In this case, the predicted item preference
η
,
makes more contributions to the predicted score

,
to item t than the contribution made by the
predicted taxonomic preference
,
. On the other
hand, if the value of is low (i.e. is low which
indicates a cold start situation), the taxonomic
preference prediction becomes more important and
will contribute more to the predicted score 
,
that what the predicted item preference does. This
design ensures that taxonomic preferences are used
to supplement or enrich the item preference
prediction, especially in cold start situations.
4 EXPERIMENTATION
This section presents empirical results obtained from
our experiment.
4.1 Data Acquisition
The dataset used in this experiment is the “Book-
Crossing” dataset (http://www.informatik.uni-
freiburg.de/~cziegler/BX/), which contains 278,858
users providing 1,149,780 ratings about 271,379
books. In the user ratings, 433,671 of them are the
explicit user ratings, and the rest of 716,109 ratings
are implicit ratings.
The taxonomy tree and book descriptors for our
experiment are obtained from Amazon.com.
Amazon.com’s book classification taxonomy is tree-
structured (i.e. limited to “single inheritance”) and
therefore is perfectly suitable to the proposed
technique. However, not every book in our dataset is
available in Amazon.com, and we were only able to
extract taxonomy descriptors for 270,868 books
form Amazon.com. The books without descriptors
are removed from the dataset. The average number
of descriptors per book is around 3.15, and the
taxonomy tree formed by these descriptors contains
10746 unique topics.
4.2 Experiment Framework
All recommenders being used in the experiment are
developed using the Taste (http://
taste.sourceforge.net/) framework. Taste provides a
set of standardized components for developing
recommenders, therefore it ensures the
comparability of the developed recommenders fairly.
Moreover, Taste also provides an evaluation
WEB INFORMATION RECOMMENDATION MAKING BASED ON ITEM TAXONOMY
25
framework allowing researchers or developers to
evaluate the performances of their recommenders
with a standardized test bed easily and effectively.
In this experiment we constructed 7 different
recommenders, and they are listed in Table 2.
Table 2: List of experimental recommenders.
Type / Name Descriptions
Item based
Recommender
[IR]
Standard item-based CF, the detailed
algorithm is listed in (Deshpande and
Karypis 2004).
Item based
Recommender with
User Clustering
[IRC]
Standard item-based CF, however this
version only recommend items within the
candidate item list
in order to improve
computation efficiency.
Slop One
Recommender
[SO]
A well known modern item based
recommendation technique(Lemire and
Maclachlan 2005), it features on its
implementation simplicity and computation
efficiency.
Taxonomy Product
Recommender
[TPR]
A taxonomy based recommender proposed
by Ziegler(Ziegler et al. 2004). This work
uses similar taxonomy scheme to our work,
and therefore can be a good benchmark.
Item based
Recommender with
TPR
[ITR]
The combination of the item-based CF and
TPR. The hybridization scheme is identical
to HTR. The only difference is that
,
is
computed using Ziegler’s method.
Hybrid Taxonomy
Recommender
[HTR]
The proposed HTR method using users’
explicit rating data and implicit rating data
as well
Hybrid Taxonomy
Recommender (with
only explicit ratings)
[HTR_E]
The proposed HTR method using only
explicit ratings.
The purpose of conducting this test is to
ensure fair comparison with IR, IRC, SO,
which use only explicit ratings.
4.3 Evaluation Metrics
The goal of our experiment in this paper is to
compare the recommendation performances and
computation efficiencies for the recommenders
listed in Table 2.
For the recommendation quality evaluation, we
randomly divided each user
 ‘s past ratings
(i.e. 


) into two parts, one for training and
another for testing. We use
to denote
‘s
training rating data and
to denote the testing
rating data, such that




,

, and |
||
| . The testing data
actually
consists of three types of items, and they are:
z Items implicitly rated by
:
̀




z Items preferred by
:
|


,
,


z Items not preferred by
:



\ 
In the experiment, the recommenders
recommend a list of items
to
based on the
training set
, and the recommendation list
can
be evaluated with
. In order to evaluate the
performances of different recommenders based on
and
, recommendation list based evaluation
metrics such as precision and recall, Breese Score,
Half-life, and etc. (Herlocker et al. 2004; Schein et
al. 2002) can be utilized. In this paper, the precision
and recall metric is used for the evaluation, and its
formulas are listed below:

|

|
|
|
(10)

|

|
|
|
(11)
In order to provide a general overview of the
overall performances, F1 metric is used to combine
the results of Precision and Recall:
1


(12)
For the computation efficiency evaluation, the
average time required by recommenders to make a
recommendation will be compared.
4.4 Experiment Result
The test dataset is constructed by randomly choosing
10,000 users from the 278,858 users in the Book-
Crossing dataset mentioned in Section 4.1. We let
each recommenders recommend a list of k items to
these 10,000 users. We tested different values for k
ranging from 5 to 25.
The results of this part of the experiment are
shown in Figure 2, Figure 3 and Figure 4. It can be
observed from the figures that, for all the three
evaluation metrics the proposed HTR technique
achieves the best result among all the recommenders.
In the case of using only explicit rating data, the
recommendation quality of HTR (i.e. HTR_E) still
outperforms other recommenders even slightly
Figure 2: Recommender evaluation with precision metric.
0
0,05
0,1
0,15
0,2
#5 #10 #15 #20 #25
Precision
topk recommeneditems
IRC IR SO
HTR TPR HTR_E
ITR
ICEIS 2008 - International Conference on Enterprise Information Systems
26
degrading compared with using both explicit and
implicit rating data (i.e., HTR performs the best and
HTR_E performs the second best.
The standard item based CF recommender (IR)
performed similarly to the slope one recommender
(SO), however it seems that slope one recommender
is slightly better in recommending longer item lists.
In the experiment, the clustering-based CF
recommender (IRC) performed better than the
standard one (IR). The only difference between
these two recommenders is in the candidate item list
formation process. The standard item based CF uses
all items from the dataset as its candidate item list
(i.e. \

 ), whereas the clustering-based
version uses only items within a user cluster (i.e.




\

 ). Intuitionally,
the clustering-based CF might perform worse than
the standard one, because its candidate item list is
formed from a cluster which is only a subset of the
entire item set, some potential promising items
might be excluded and thus won’t be recommended.
However, based on our observation, many of these
excluded items are noises generated from the item
similarity measure (some item similarity measures
might generate prediction noise, please refer to
(Deshpande and Karypis 2004) for more
information), therefore by removing these items
from the candidate list can actually improve the
recommendation quality. The proposed HTR also
gets benefits from the clustering strategy as it
generates recommendations from the candidate item
list formed from a cluster.
We also implemented the TPR technique
proposed by Ziegler(Ziegler et al. 2004), and it
performed worst among all recommenders in our
evaluation scheme. TPR uses only implicit ratings as
its data source and generates recommendations only
based on taxonomy preferences. In order to make the
proposed HTR and Ziegler’s TPR more comparable,
we modified TPR by adding the item-based CF
component into TPR resulting in the new
recommender ITR. ITR performed better than the
standard TPR as it included the item preference
consideration in its recommendation making process.
However it is still worse than all other
recommenders (i.e., TPR performs the worst and
ITR performs the second worst). The difference
between HTR and ITR is the method to compute the
taxonomy preferences is different (they use the same
method to compute the item preferences). The result
of HTR outperforming ITR indicates that users’ item
preference is also helpful for generating users’
taxonomy preference. The proposed HTR technique
considers the item preference implication when
generating the taxonomic preferences (i.e. the
taxonomic preferences are extracted from user
clusters which is divided based on users’ item
preferences). In contrary, TPR generates users’
taxonomic preferences purely from taxonomy data
without using any of the users’ item preferences.
Figure 3: Recommender evaluation with recall metric.
Figure 4: Recommender evaluation with F1 metric.
In the experiment, the recommender with the best
computation efficiency is the clustering based CF
(IRC) as showed in Figure 5, it is much faster than
the standard CF because its candidate item list is
much smaller. The proposed HTR methods (HTR
and HTR_E) perform the third and second best, as
they added a bit computation complexity in the
taxonomic preference predictions. However, this
extra computation complexity is trivial, because
most of these computations (i.e. computing
_ for each user cluster) can be done offline.
HTR_E performed slightly better than HTR because
it uses less data (only explicit ratings) to make
recommendations. Ziegler’s TPR is computation
expensive because it needs to convert all users and
0
0,02
0,04
0,06
0,08
0,1
0,12
#5 #10 #15 #20 #25
Recall
topk recommeneditems
IRC IR SO
HTR TPR HTR_E
0
0,02
0,04
0,06
0,08
0,1
#5 #10 #15 #20 #25
F1
topk recommeneditems
IRC IR SO
HTR TPR HTR_E
ITR
WEB INFORMATION RECOMMENDATION MAKING BASED ON ITEM TAXONOMY
27
items into high dimensional taxonomy vectors. ITR
performed slightly worse than TPR because it needs
to compute extra item preference predictions using
standard CF technique. Standard CF technique is the
most inefficient one among all the recommenders,
whereas slop one recommender offers a slight
advantage in computation efficiency.
Figure 5: Average second per recommendation.
5 CONCLUSIONS
In this paper, we investigated the implicit relations
between users’ item preferences and taxonomic
preferences, suggested and also verified using
information gain that users that share similar item
preferences may also share similar taxonomic
preferences. Based on this investigation, we
proposed a novel, hybrid technique HTR to
automated recommendation making based upon
large-scale item taxonomies which are readily
available for diverse ecommerce domains today.
HTR produces quality recommendations by
incorporating both users’ taxonomic preferences and
item preferences. Moreover, it can utilize both
explicit and implicit ratings for recommendation
making, and hence they are less prone to the cold
start problem. We have compared the proposed HTR
technique with some standard benchmark techniques
such as item-based recommender and some
advanced modern techniques such as TPR (which
are related to ours). We have conducted extensive
experiments which demonstrated that the proposed
HTR outperforms other recommenders in both
recommendation quality and computation efficiency.
REFERENCES
Adomavicius, Gediminas, Ramesh Sankaranarayanan,
Shahana Sen and Alexander Tuzhilin. 2005.
"Incorporating contextual information in recommender
systems using a multidimensional approach " ACM
Trans. Inf. Syst. 23(1):103-145.
Breese, J. S., D. Heckerman and C. Kadie. 1998.
"Empirical Analysis of Predictive Algorithms for
Collaborative Filtering." In Proceedings of 14th
Conference on Uncertainty in Artificial Intelligence.
Madison, WI.
Burke, Robin. 2002. "Hybrid Recommender Systems:
Survey and Experiments." User Modeling and User-
Adapted Interaction 12(4):331-370.
Deshpande, Mukund and George Karypis. 2004. "Item-
based top-N recommendation algorithms." ACM
Transactions on Information Systems 22(1):143-177.
Ferman, A. Mufit, James H. Errico, Peter van Beek and M.
Ibrahim Sezan. 2002. "Content-based filtering and
personalization using structured metadata." In 2nd
ACM/IEEE-CS joint conference on Digital libraries
Portland, Oregon, USA.
Herlocker, Jonathan L., Joseph A. Konstan, Loren G.
Terveen and John T. Riedl. 2004. "Evaluating
collaborative filtering recommender systems." ACM
Transactions on Information Systems (TOIS) 22(1):5-
53.
Lemire, D. and A. Maclachlan. 2005. "Slope One
Predictors for Online Rating-Based Collaborative
Filtering." In 2005 SIAM Data Mining
Middleton, Stuart E, Harith Alani, Nigel R Shadbolt and
David C. De Roure. 2002. "Exploiting Synergy
Between Ontologies and Recommender Systems." In
The Semantic Web Workshop, World Wide Web
Conference.
Montaner, Miquel, Beatriz López and Josep Lluís De La
Rosa. 2003. "A Taxonomy of Recommender Agents
on the Internet." Artificial Intelligence Review
19(4):285-330.
Park, Seung-Taek, David Pennock, Omid Madani, Nathan
Good and Dennis DeCoste. 2006. "Naive filterbots for
robust cold-start recommendations." In 12th ACM
SIGKDD international conference on Knowledge
discovery and data mining, ed. ACM Press.
Philadelphia, PA, USA.
Sarwar, B., G. Karypis, J. Konstan and J. Riedl. 2000.
"Application of dimensionality reduction in
recommender systems--a case study." In ACM
WebKDD Workshop. Boston, MA, USA.
Schafer, J. Ben, Joseph A. Konstan and John Riedl. 2000.
"E-Commerce Recommendation Applications."
Journal of Data Mining and Knowledge Discovery
5:115-152.
Schein, Andrew I., Alexandrin Popescul, Lyle H. Ungar
and David M. Pennock. 2002. "Methods and metrics
for cold-start recommendations" In 25th annual
international ACM SIGIR conference on Research and
development in information retrieval. Tampere,
Finland: ACM Press.
Ziegler, Cai-Nicolas, Georg Lausen and Lars Schmidt-
Thieme. 2004. "Taxonomy-driven Computation of
Product Recommendations" In International
Conference on Information and Knowledge
Management Washington D.C., USA
0,0017
5,6664
5,0825
0,0962
1,5441
0,0473
2,0355
0
1
2
3
4
5
6
IRC IR SO HTR TPR HTR_E ITR
seconds
recommendertypes
ICEIS 2008 - International Conference on Enterprise Information Systems
28