Content uploaded by Hyung Jun Ahn
Author content
All content in this area was uploaded by Hyung Jun Ahn on Jun 02, 2015
Content may be subject to copyright.
International Journal of Electronic Commerce / Winter 2006–7, Vol. 11, No. 2, pp. 57–78.
Copyright © 2007 M.E. Sharpe, Inc. All rights reserved.
1086-4415/2007 $9.50 + 0.00.
DOI 10.2753/JEC1086-4415110203.
Utilizing Popularity Characteristics for
Product Recommendation
Hyung Jun Ahn
ABSTRACT: This paper presents a novel approach to automated product recommendation
based on the popularity characteristics of products. Popularity plays a signifi cant role in the
consumer purchasing process but has not been given much attention in recommendation
research. A three-dimensional model of popularity is used to develop popularity classes
of products. These are joined with the MovieLens dataset to create a hybrid movie recom-
mendation system that combines genre and popularity information. As compared with
collaborative fi ltering, the hybrid system shows positive results under the conditions of data
sparsity and cold-starting. Many interesting issues for further research are suggested.
KEY WORDS AND PHRASES: Automated product recommendation, cold-starting, hybrid
recommender system, naive Bayesian, popularity-based recommendation, popularity
model, sparsity.
Automated product recommendation is widely used by many Internet shop-
ping malls, where it plays a critical role in effective on-line marketing by
promoting cross-selling and up-selling of products. As electronic commerce
matures, the effectiveness of recommendation is winning recognition as a
crucial factor for organizations under growing competitive pressure.
Researchers have produced successful recommendation methods that use
data of many different kinds, including purchase history, product ratings by
buyers, product characteristics, and demographic information of shoppers
[1, 2, 5, 20]. Recommendation methods can be broadly categorized as either
content-based methods or collaborative fi ltering methods [5, 6, 8, 12, 14, 15, 17,
18, 20, 30]. Successful results have been reported, and many of these methods
have been implemented by real-world organizations. Nonetheless, they often
entail problems, such as poor recommendation quality under data sparsity
and limited ability to recommend new products or to new buyers [7, 19, 21, 29,
33]. In addition, many recommendation methods do not clearly explain to the
user why they are recommending a specifi c product. This limits the potential
uses of the recommendation results for further analysis [11].
The present paper proposes a novel approach to recommendation that
uses the popularity characteristics of products. Popularity features play an
important role in consumers’ purchasing decisions because most consumers
are infl uenced by how others feel about a product or how widely a product has
been exposed in the market [3, 16]. Despite the signifi cance of the popularity
factor, it has not been given much attention in recommendation studies. The
discussion that follows defi nes three dimensions of popularity and presents
an algorithm for partitioning products into a reasonable number of popular-
ity classes located in the three-dimensional space of popularity. The approach
is applied to the MovieLens dataset, and a hybrid recommender system for
movies is developed by combining popularity-class information with genre
information. The performance of the combined system is experimentally
03 ahn.indd 5703 ahn.indd 57 11/20/2006 10:26:06 AM11/20/2006 10:26:06 AM
58 HYUNG JUN AHN
compared with the collaborative fi ltering method under the conditions of
data sparsity and cold-starting.
Overview of Literature
Categorization of Recommender Systems
Approaches to automated product recommendation can be broadly classifi ed
as either content-based fi ltering or collaborative fi ltering. Content-based ap-
proaches use content information, or features of products, to build profi les of
products and buyers that are then used to calculate the match between a specifi c
buyer and product [14, 18, 20]. Product categories and product descriptions
have been used for content-based fi ltering in many studies [1, 2, 20, 22]. Col-
laborative fi ltering, in contrast, uses buyers’ ratings of products rather than
content information and fi nds people-to-people similarities between buyers.
Using this information, it recommends products highly rated by similar buy-
ers to a target buyer [6, 8, 12, 15, 17, 30].
A more fi ne-grained fi ve-part classifi cation has been devised by Burke
[5]. Working from an extensive literature review, he classifi es recommender
systems as either collaborative, content-based, demographic, utility-based,
or knowledge-based, depending on the types of data and processes they use.
Demographic recommender systems resemble collaborative fi ltering in that
they utilize people-to-people similarity for recommendation, but they use
demographic information, and not buyers’ rating data, for similarity calcula-
tion. Utility-based systems use buyers’ utility functions to calculate the level
of utility of a specifi c item for a specifi c buyer. Knowledge-based systems use
inferences, often adopting techniques from artifi cial intelligence, to infer a
match between buyer and product. Both utility-based systems and knowledge-
based systems use features of products rather than buyers’ ratings.
The various recommender systems exhibit different pros and cons, and this
has led to the development of hybrid systems. Burke presents seven types of
hybrid recommender systems: weighted, switching, mixed, feature combina-
tion, cascade, feature augmentation, and meta-level [5]. These methods either
combine recommendation results from different methods (weighted, switch-
ing, and mixed), utilize different types of data together (feature combination),
or pipeline two or more recommender systems (cascade, feature augmentation,
and meta-level) in various ways.
The present paper presents a hybrid recommendation method that utilizes
both buyers’ ratings and genre information about movies. Thus, it can be lo-
cated within the feature-combination category of the classifi cation outlined
above.
Collaborative Filtering
As reported by many researchers, collaborative fi ltering performs well and has
won wide acceptance. Therefore, it is reasonable to compare its performance
with that of the hybrid method proposed in this paper.
03 ahn.indd 5803 ahn.indd 58 11/20/2006 10:26:07 AM11/20/2006 10:26:07 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 59
The literature offers many variations on collaborative fi ltering [8, 9, 10, 13,
18, 32]. Most collaborative fi ltering methods predict a buyer’s rating of a spe-
cifi c product based on the way similar buyers have rated the same product.
The Pearson correlation coeffi cient and cosine measure are frequently used to
calculate similarities between buyers. Formally, the following formulae using
Pearson’s correlation coeffi cient have been widely adopted for the similarity
calculation and the prediction, respectively [24, 26]:
(
)
(
)
(
)
(
)
(
)
xh x yh y
xh x yh y
n
ui u ui u
h
xy
nn
ui u ui u
hh
rrrr
sim u u
rr rr
,,
,,
,
′
=
′′
==
−−
=
−−
∑
∑∑
1
2
2
11
for the similarity calculation between two users u
x
and u
y
where r
u,i
is the rat-
ing of product i by user u, r\
u
is the average rating of user u for all the products
rated by the user, and n′ is the number of items co-rated by both users.
()
()
(
)
()
ha h
a
m
ah ui u
h
aa u
m
ah
h
sim u u r r
pu i r
sim u u
,
,
,
,
′
=
′
=
−
=+
∑
∑
1
1
for the prediction of a rating for item i
a
by user u
a
where m′ is the number of
other users who have also rated the item.
These formulae are also used in this paper for performance comparison.
Collaborative fi ltering methods often attain successful results but are handi-
capped by not performing well under data sparsity, cold-starting, and lack of
understandability. Sparsity of data refers to the problem of data defi ciency,
which occurs quite frequently because of the diffi culty of collecting a suffi ciently
large set of rating data in most Internet shopping malls. The problem results
in impaired similarity calculation and poor recommendation performance
[29, 33]. Cold-starting refers to the diffi culty of recommending products to
a new buyer who has little or no purchasing history, or of recommending a
new product when there are very few rating records [7, 19, 21]. Collaborative
fi ltering methods have the further limitation of not adequately explaining their
recommendations [11]. Most collaborative fi ltering methods use what may be
called a “black-box” process in that it cannot explain what characteristics of
buyer and product lead to the given recommendation, mainly because of the
diffi culty of intuitively describing specifi c results of similarity and prediction
calculation. This limits the possibilities for further analysis and for more fl exible
use of the recommendation results.
Utilizing Popularity for Recommendation
Three Dimensions of Popularity and Popularity Measures
This paper proposes a new approach to automated recommendation, here-
inafter termed popularity-based recommendation (PBR), that utilizes the
03 ahn.indd 5903 ahn.indd 59 11/20/2006 10:26:07 AM11/20/2006 10:26:07 AM
60 HYUNG JUN AHN
popularity characteristics of products. Broadly speaking, there are two reasons
for adopting popularity features for recommendation. First, popularity often
represents important characteristics of a product. Many products are planned
and designed from the outset to be strategically positioned in specifi c market
segments with different numbers of potential consumers. By implication, the
developers have intentionally provided these products with different popu-
larity characteristics in order to appeal to consumers of different types. The
extent to which a product is known to consumers via various channels usually
indicates how broadly it has been advertised and promoted, and what types
of consumers are being targeted. Second, the popularity of a product greatly
infl uences consumer purchasing decisions. Different types of consumers show
dissimilar ways of information seeking and problem solving in purchasing.
For example, some consumers utilize extensive sources of information before
making decisions, whereas others rely on easy, simple, and limited sources of
information. Quite obviously the two groups will be affected in different ways
by product popularity [16]. There are also differences in motivation between
consumers choosing a product to purchase. According to Arnould, Price, and
Zinkhan, some consumers are motivated by a ”uniqueness or novelty need,”
and others by an ”affi liation motive,“ which again implies that consumers
are distinctively affected by certain popularity characteristics [3]. Despite
their signifi cance, as discussed above, recommendation systems have made
only very limited use of popularity characteristics, such as recommending
bestsellers to buyers [28].
The utilization of the concept of popularity in recommendation requires
an understanding of its three key dimensions. The fi rst dimension represents
whether consumers perceive the product to be of high value. When used alone,
this dimension represents a product’s quality or level of satisfaction rather than
its popularity, but when combined with the other two dimensions, it plays an
important role in explaining the popularity features of a product. The second
dimension represents the frequency of a product’s being purchased regardless
of its perceived value. It differs from the fi rst dimension in that many products
are mass-marketed and widely exposed but are not rated highly by consumers.
The third dimension of product popularity is the size of the strong-support
group for a product regardless of its average rating or frequency of purchase.
Early adopters in the product life-cycle model [16] and the fans of “cult”
movies are good examples of buyers who exhibit strong support for certain
products. Figure 1 defi nes the three dimensions of popularity more formally.
As shown in the fi gure, the fi rst popularity dimension, average rating (AR), is
calculated by dividing the sum of all buyer ratings of a product by the total
number of buyers who have purchased it (or rated it). The second dimension
of popularity, percentage of being rated (PR), is the total number of ratings of
a product divided by the total number of potential buyers, which represents
the product’s level of exposure. The third popularity measure, strong support
(SS), is the number of buyers who have shown strong support for the product
divided by the total number of buyers who have purchased it (or rated it). The
measure assumes that a buyer whose rating for a product exceeds a predefi ned
threshold value has shown strong support for the product.
03 ahn.indd 6003 ahn.indd 60 11/20/2006 10:26:07 AM11/20/2006 10:26:07 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 61
In light of the foregoing, the space of popularity can be defi ned as a three-
dimensional geometric space with SS, AR, and PR as its three rectangular
axes.
Dataset for Experiment
The experiments described in this paper applied the popularity model to the
MovieLens dataset, which is widely used in recommendation research [23].
Movies with fewer than fi ve ratings were removed from the dataset. This left
about 77,000 ratings for 838 movies by 943 raters. Each user had at least 20
ratings. Eighty percent of the raters, or buyers, were assigned to training, and
20 percent were assigned to test the recommendation performance (see Figure
2). The same divisions were used for the collaborative fi ltering method for
comparison. The ratings of users range from 1 to 5 in the dataset. A rating of
4 or above indicated preference for a movie, and 2 or below dislike. A rating
of 5, indicating strong preference for a movie, was used as the threshold value
to calculate the SS dimension of popularity.
Overview of Recommendation Procedure
The recommendation procedure using the popularity-based rating method is
summarized in Table 1, where each step of the procedure is described along
with the data components in Figure 3. In essence, the PBR approach can be
defi ned as fi nding out what popularity characteristics a buyer prefers and
recommending products that exhibit them.
In the method’s fi rst step, the three-dimensional space of popularity is par-
titioned into discrete popularity classes where each class is shaped as either
a cuboid (rectangular box) or a cube in the space. Each movie in the dataset
located at a point of the popularity space is assigned a popularity class. Profi les
of the movies are constructed by combining popularity-class information from
the preceding step with genre information provided by the original dataset.
Since some movies belong to more than one genre in the MovieLens dataset,
this assigns one or more <genre ID, popularity class> pairs to each movie. In
Figure 1. Three Dimensions of Popularity
03 ahn.indd 6103 ahn.indd 61 11/20/2006 10:26:07 AM11/20/2006 10:26:07 AM
62 HYUNG JUN AHN
Figure 2. Dataset for Experiments
Step Description Output Data used
1 Derivation of popularity Discrete popularity classes. Each class is All ratings for
classes either a cuboid (rectangular box) or a movies
cube in the three-dimensional popularity
space.
2 Profi ling of movies All movies in dataset are assigned Output from
one or more <genre ID, popularity class> Step 1; genre
pairs. information from
dataset
3 Construction of virtual Virtual shopping history of training Output from
baskets from ratings buyers containing <genre ID, Step 2; movies
of training buyers popularity class> pairs for movies each in (1) of Figure 3
buyer has rated. for each buyer
b
train
4 Calculation of sample Sample probabilities that represent Output from
probabilities strength of association between Step 3
<genre ID, popularity class> pairs
from the baskets.
5 Creation of preference Preference profi les for each buyer Output from
profi les for testing buyers that contain preference scores for Step 2; movies
all <genre ID, popularity class> pairs. in (2) of Figure 3
for each buyer
b
test
6 Recommendation Movies are recommended according Output from
to preference profi le constructed in Steps 2 and 5;
Step 5. movies in (3)
of Figure 3 for
each buyer b
test
Table 1. Recommendation Procedure Using PBR Method
03 ahn.indd 6203 ahn.indd 62 11/20/2006 10:26:08 AM11/20/2006 10:26:08 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 63
the third step, virtual shopping baskets are constructed for the buyers in the
training-buyer section by assuming that they have purchased or watched the
movies they rated in the dataset. The basket represents the buyer’s purchase
history and, as in the Market Basket Analysis technique [4], is used to calculate
sample probabilities for showing the strength of association between different
<genre ID, popularity class> pairs in the fourth step. In the fi fth step, preference
profi les are constructed for the testing buyers in accordance with their virtual
purchase histories (using each buyer’s ratings of training movies as the pur-
chase history). The generated profi les contain the buyer’s preference scores for
each <genre ID, popularity class> pair. The naive Bayesian method is partly used
in this step to estimate preference scores for some <genre ID, popularity class>
pairs for which a buyer has not shown any explicit preference. In the last step,
movies are recommended to the buyer according to the buyer’s preference
profi le—movies that belong to <genre ID, popularity class> pairs with higher
preferences scores in the profi le are recommended fi rst.
These steps are explained in detail in the following sections.
Partitioning the Popularity Space and
Categorization of Movies
As stated earlier, this paper presents a hybrid recommender system that uti-
lizes popularity characteristics and movie genre information. The MovieLens
dataset comprises 18 different genres, and a movie can be assigned to more
than one genre. Thus, every movie is assigned one or more pairs of <genre
ID, a popularity class>, which from here on will be denoted simply as <g
i
, p
j
>
pairs.
Popularity classes are derived by dividing each dimension of the popularity
space into four discrete unit spaces, as shown in Figure 4. Each unit space’s
length for each dimension is:
Figure 3. Breakdown of the Dataset for the Recommendation
Procedure
03 ahn.indd 6303 ahn.indd 63 11/20/2006 10:26:08 AM11/20/2006 10:26:08 AM
64 HYUNG JUN AHN
Maximum_Value_of_Dimension Minimum_Value_of_Dimension
.
−
4
This initially classifi es movies into discrete unit spaces where there are, in
all, 4 × 4 × 4 = 64 spaces. Since sample probabilities will be calculated in the
later phase for each <g
i
, p
j
> pair, and probabilistic associations among the
pairs will be derived, the number of spaces, 64, is still too large, given that
the MovieLens dataset comprises 18 genres. If there are, in all, 64 × 18 = 1152
<g
i
, p
j
> pairs for the categorization of movies, the probabilities for each pair
as well as for the associations between the pairs will be adversely affected
by the insuffi cient number of data and thus unreliable. Moreover, having
as many as 64 different categories of products in terms only of popularity is
counter-intuitive. Considering these problems of complexity and sparsity, the
paper fi xes 10 as the maximum number of popularity classes and will not ac-
cept a class with fewer than 30 movies, but bear in mind that this is a design
choice for the purposes of the paper rather than an optimum. The following
algorithm is used to partition the three-dimensional popularity space into a
small number of popularity classes (≤ 10) under the constraints given above.
Note that the size of a class or a space in the algorithm denotes the number of
movies in the training data that belong to the class or the space.
Figure 4. Partitioning of Movies Based on Popularity Measures
03 ahn.indd 6403 ahn.indd 64 11/20/2006 10:26:09 AM11/20/2006 10:26:09 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 65
Step 1. NSC (miNimum Size of a Class) = 30.
Step 2. MSC (Maximum Size of a Class) = k ; k is a small number
(e.g., 100).
Step 3. Begin with the entire space of popularity.
Step 4.1. If the given space is a unit space, register it as a fi nal class; go
to Step 5.
Step 4.2. If the size of the given space is smaller than MSC, register it
as a fi nal class; go to Step 5.
Step 4.3. If neither of the above is true, divide the given space into
all possible pairs of halves (there can be one to three pairs
of half spaces according to the shape of the given space, as
in Figure 5). Choose the largest of the half spaces in all the
pairs, and, if it is smaller than MSC, register it as a fi nal class
and go to Step 4.1 with its counterpart in the same pair. If it
is not smaller, go to Step 4.1 twice with the half space and its
counterpart in the same pair respectively.
Step 5. If any remaining subspace is unregistered, wait until it is di-
vided and registered as classes (there can be multiple fl ows
of partitioning going through Step 4.1 to Step 4.3).
Step 6. If the number of registered classes is smaller than or equal
to 10, and the smallest block is larger than NSC, stop. If not,
increase MSC by 1 and go to Step 3.
The advantages of this partitioning method can be summarized as
follows:
1. Partitioning the space into a manageable number of classes enables
the use of the popularity characteristics as discrete metrics rather
than continuous ones. This makes it easier to use the measures.
2. Grouping movies into popularity classes facilitates identifying the
relational characteristics between classes. For example, the probabil-
Figure 5. Half Spaces According to Different Types of Given Space
03 ahn.indd 6503 ahn.indd 65 11/20/2006 10:26:09 AM11/20/2006 10:26:09 AM
66 HYUNG JUN AHN
ity that a preference for one class will imply a preference for another
class can be easily estimated with sample statistics between the <g
i
,
p
j
> pairs.
Compared with clustering approaches such as the k-means method [31],
partitioning makes the classes easier to understand because each partition
is clearly sectioned with discrete boundaries in each dimension, rather than
grouped around clustering centers with unintuitive and confusing boundaries.
Furthermore, unlike clustering, partitioning does not require an assumption
about the underlying parameters of clusters, such as initial cluster centers and
number of clusters. This is a valuable property because easy understanding
of product profi ling makes for easier and wider utilization of the measures
(as will be discussed further on).
Partitioning results in the seven popularity classes shown in Figure 4. The
numbers near a class represent, respectively, the ID of the class and the number
of movies that belong to it. For example, 7(111) represents the seventh class
with 111 movies associated. Table 2 provides brief descriptions of the classes.
As can be seen, the 32 movies in the second class are highly exposed movies
with high average ratings. The 111 movies in the seventh class appear to be
less exposed but have large, strong support groups and high average ratings.
One can roughly infer that the movies in the second class are blockbusters,
whereas those in the seventh class are cult movies favored by special groups.
Figure 6 shows the distribution of movies in each class in the three-dimensional
space of popularity and in all of its two-dimensional subplanes.
Construction of Shopping Baskets for
Training Buyers
The shopping baskets for training buyers are constructed using the class
and genre information of movies to calculate the probabilities for association
among the pairs. Two types of baskets are constructed for each user: a prefer-
ence basket and a dislike basket. A buyer’s preference basket contains <g
i
, p
j
>
pairs for movies the buyer has rated 4 or above. That is, basketPref
b
= {<g
i
, p
j
>
| <g
i
, p
j
> is a profi le of a movie that buyer b has rated 4 or above}. Similarly,
a buyer’s dislike basket is defi ned as basketDislike
b
= {<g
i
, p
j
> | <g
i
, p
j
> is a
Class Description
1 Overall poorly rated
2 High exposure and highly rated
3 Low-to-medium exposure, highly rated, low-to-medium level of strong support
4 Low-to-medium exposure, rated highly, low-to-medium level of strong support
5 Low exposure, rated OK, small strong support
6 Low exposure, rated OK, medium level of strong support
7 High strong support, low-to-medium level of exposure, rated good
Table 2. Brief Description of Popularity Classes.
03 ahn.indd 6603 ahn.indd 66 11/20/2006 10:26:10 AM11/20/2006 10:26:10 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 67
profi le of a movie that buyer b has rated 2 or below}. Note that the baskets
only represent whether or not the pairs appear in the buyer’s ratings, not the
count of the pairs. In some instances, a buyer’s preference and dislike baskets
may both contain the same <g
i
, p
j
> pair. Using the baskets constructed with
the preceding method, three types of sample probabilities are calculated, as
shown in Table 3.
Constructing Testing Buyer Profi les and
Recommending Movies
A testing buyer profi le is defi ned as a vector of 3-tuples (<g
1
, p
1
, v
1
>, <g
1
, p
2
,
v
2
>, . . . , <g
n
, p
m
, v
nm
>) of length n × m when there are n genres and m popularity
classes, and v
k
represents the preference value of the buyer for each <g
i
, p
j
>.
Constructing a testing buyer’s profi le begins with 0 preference values for
all <g
i
, p
j
> pairs. That is to say, initially a testing buyer profi le vector is (<g
1
, p
1
,
0>, <g
1
, p
2
, 0>, . . . , <g
n
, p
m
, 0>). Then, for each buyer rating of a training movie
of 4 or above, 1 is added to all the <g
i
, p
j
> pairs of the profi le that correspond
to the movie. Conversely, 1 is subtracted from the values of the <g
i
, p
j
> pairs
that correspond to the movie if the rating is 2 or below. After going through
Figure 6. Scatter Plots of the Movies of Each Class
Notes: (a) two-dimensional plot with X = SS and Y = PR; (b) two-dimensional plot with X = SS
and Y = AR; (c) two-dimensional plot with X = PR and Y = AR; (d) three-dimensional plot with
X = S, Y = PR, and Z = AR.
03 ahn.indd 6703 ahn.indd 67 11/20/2006 10:26:10 AM11/20/2006 10:26:10 AM
68 HYUNG JUN AHN
all the buyer’s ratings of the training movies, for example, what is left is a
vector that may look like the following: ((<g
1
, p
1
, 3>, <g
1
, p
2
, –2>, . . . , <g
i
, p
j
,
0>, <g
i
, p
j+1
, 0> . . . , <g
n
, p
m
, 2>). As can be seen, some <g
i
, p
j
> pairs still have 0
values because the buyer has not explicitly shown any preference or dislike for
them, whereas other pairs are marked with explicit values for preferences or
dislikes. For the pairs that have not been assigned an explicit value, the naive
Bayesian method is used to calculate the preference values for them indirectly
[31]. Suppose that a buyer has shown explicit preferences for <g
l1
, p
l1
>, . . . ,
<g
ln
, p
ln
> pairs and explicit dislikes for <g
d1
, p
d1
>, . . . , <g
dm
, p
dm
> pairs. If, for
any x and y, <g
x
, g
y
>
L
denotes an event where the user likes the <g
x
, g
y
> pair,
and <g
x
, g
y
>
D
denotes an event where the user dislikes the pair, then:
(
)
(
)
(
)
(
)
11 1 1
11 1 1
11 1 1
<>< >< >< >< >
<>< >< >< ><>
=
<>< >< >< >
<>
, | , ... , , ... ,
, ... , , ... , | ,
, ... , , ... ,
,
LL L D D
i j l l ln ln d d dm dm
LLD DL
l l ln ln d d dm dm i j
LLD D
l l ln ln d d dm dm
L
ij
Pgp gp gp gp g p
Pgp gp gp g p gp
Pgp gp gp g p
Pgp
(1)
(1) represents the probability that a buyer who has shown explicit preferences
and dislikes for certain pairings of genre and popularity class will like <g
i
, p
j
>.
Since only the ordering of <g
i
, p
j
> pairs will be utilized for the recommendation,
the naive Bayesian method discards the denominator from (1). Thus:
(
)
(
)
LLD
l l ln ln d d
DL L
dm dm i j i j
Pgp gp gp
gp gp Pgp
, ... , , ...
,|, ,.
<>< >< >
<><><>
11 1 1
(2)
Assuming independence among the conditional probabilities and using the
probabilities calculated in the preceding subsection, (2) transforms to:
(
)
(
)
(
)
(
)
(
)
like l l i j
like ln ln i j dislike d d i j
dislike dm dm i j like i j
Pgpgp
Pgp gpP gp gp
PgpgpPgp
, | , ...
, | , , | , ...
,|, ,.
<><>
< ><> < ><>
< ><><>
11
11
(3)
Probability Description
P
like
(<g
i
, p
j
>) Probability that pair <g
i
, p
j
> appears in a preference basket.
P
like
(<g
i
, p
j
> | <g
k
, p
l
>) Conditional probability that pair <g
i
, p
j
> appears in preference baskets
that contain <g
k
, p
l
>, where i ≠ k or j ≠ l.
P
dislike
(<g
i
, p
j
> | <g
k
, p
l
>) Conditional probability that pair <g
i
, p
j
> appears in dislike baskets of
buyers who have <g
k
, p
l
> in their preference baskets, where i ≠ k or j ≠ l.
Table 3. Probabilities for Naive Bayesian.
03 ahn.indd 6803 ahn.indd 68 11/20/2006 10:26:11 AM11/20/2006 10:26:11 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 69
As can be seen, all the results calculated by (3) will have values between 0 and
1. Therefore, if the 3-tuples are sorted in a testing buyer profi le in descending
order of preference values, all indirectly rated <g
i
, p
j
> pairs will be located, by
the naive Bayesian method, between explicitly preferred pairs and explicitly
disliked ones, as shown in Table 4. For a given buyer, all the candidate movies
for recommendation are scored according to the preference values, and the
highly ranked candidates are fi nally recommended. For movies with more
than one genre, the average of all corresponding <g
i
, p
j
> pair scores is used
for the recommendation.
Experiments
Overview
As shown in Figure 2, 80 percent of the buyers were used to build movie
profi les, and the remaining 20 percent were used to test the recommendation
performance. The movies were also divided into a training set and a test set,
with the training-set movies used to build test-buyer profi les, and the test-set
movies used as candidates for recommendation to the buyers. Several experi-
ments were performed using the movie and test-buyer profi les constructed
through the procedure introduced earlier. First, a basic experiment was per-
formed for different proportions of divisions between training movies and test
movies. This experiment also tested whether utilizing popularity characteris-
tics resulted in improved performance as compared with baseline methods.
Second, the recommendation performance was tested under different levels
of sparsity of rating data. Third, a cold-starting experiment was performed
for imaginary new users by increasing the number of training movies for test
buyers from one to ten. In all three experiments, the results were compared
to those obtained by the collaborative fi ltering method, using the formulae
introduced earlier. The average rating of top fi ve recommended movies was
used as the measure of performance. This measure was used because the
widely used mean absolute error (MAE) requires prediction of individual
Profi ling method Genre/popularity Preference value
Explicitly preferred <g3, p2> 4
… … …
… … …
Explicitly preferred <g11, p5> 1
Indirectly rated <g7, p2> 0.78655
Indirectly rated <g2, p3> 0.23526
… … …
Explicitly disliked <g6, p7> –1
Explicitly disliked <g2, p4> –2
… … …
Table 4. Example Profi le Table for Test Buyer.
03 ahn.indd 6903 ahn.indd 69 11/20/2006 10:26:12 AM11/20/2006 10:26:12 AM
70 HYUNG JUN AHN
ratings, which the PBR method cannot generate. The average rating of top N
recommended items is a good measure of performance in this case because it
directly measures buyer satisfaction with the N recommended items.
Performance Using All Available Ratings
In this experiment, the whole set of movies was divided into M
train
percent
training movies and M
test
percent test movies (see Figure 2). Using all the
available ratings, the recommendation performance of the PBR method was
compared to (1) a random recommendation where fi ve movies were recom-
mended randomly, (2) the collaborative fi ltering method using the formulae
introduced earlier, and (3) a genre-only method applying the same approach
as the PBR method using only the genre information of movies—that is, using
only <g
i
> for movie and buyer profi les.
The proportion of M
train
was increased from 20 to 80 percent, by 20 per each
step. The results were averaged for 20 iterations. On the whole, as shown in
Figure 7, the PBR method performed less well than the collaborative fi ltering
method, but was a signifi cant improvement on the random-recommendation
and genre-only methods, which demonstrates the effect of using the popularity
characteristics. Subsequent experiments used M
train
= 40 because at that point
both collaborative fi ltering and PBR showed the best performance. This was
also a reasonable choice for the cold-starting experiment presented below,
since only a relatively small number of profi ling movies is needed to simulate
a cold-starting situation.
Sparsity
Next, the performance of the PBR method was compared with the collabora-
tive fi ltering method under different sparsity levels. The following measure
of sparsity, adopted in other articles, was used for the experiment [24, 27]:
()( )
()
Number of full ratings Actual number of ratings
Sparsity .
Number of full ratings
−
=
Ratings were randomly chosen and deleted according to the number of
required ratings for each sparsity level. The initial sparsity of the MovieLens
data with all ratings was 0.877. Thus, the sparsity level was increased from
0.88 to 0.99 by 0.01 each step. The result shown in Figure 8 demonstrates an
interesting and valuable property of the PBR method. Under low levels of
sparsity, collaborative fi ltering method outperformed PBR, but when the
sparsity level was around 0.96, PBR began to perform as well as or better than
collaborative fi ltering. Under the more severe sparsity levels, such as 0.97,
0.98, and 0.99, the PBR method clearly outperformed the collaborative fi ltering
method. The slopes of the two methods showed different patterns, and it is
evident that the performance of the PBR method dropped more gently than
03 ahn.indd 7003 ahn.indd 70 11/20/2006 10:26:12 AM11/20/2006 10:26:12 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 71
Figure 7. Performance Comparison for Different Proportions of
Profi ling Movies
y-axis represents the average rating of top fi ve recommended movies.
3.4
3.5
3.6
3.7
3.8
3.9
4
0.88 0.89 0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98
0.99
PBR
CF
Figure 8. Performance Comparison Under Different Levels of Sparsity
y-axis represents the average rating of top fi ve recommended movies and x-axis represents dif-
ferent levels of sparsity.
that of the collaborative fi ltering method. As an explanation, one may postulate
that the PBR method is more robust to sparsity because, for profi ling movies,
it uses aggregate measures of popularity that may not change sensitively as
the number of ratings decreases, whereas the similarity calculation in the col-
laborative fi ltering method uses all the ratings individually.
03 ahn.indd 7103 ahn.indd 71 11/20/2006 10:26:12 AM11/20/2006 10:26:12 AM
72 HYUNG JUN AHN
Cold-Starting for New Shoppers with
Few Purchase Records
The next experiment tested the performance of PBR under a cold-starting
situation for new buyers. The PBR method was applied to an artifi cial cold-
starting situation where new shoppers were presented to the recommender
with a very small number of purchases on record or the number of ratings in
the case of the MovieLens dataset. For each testing buyer, the number of rated
movies for profi ling was increased from one to ten, and candidate movies were
recommended using the profi les. The same procedure was also applied to the
collaborative fi ltering method for comparison, where, again, one to ten movies
were increasingly used to calculate similarity between test buyers and training
buyers. In order to get better practical insights from the experiment, it was
repeated for 12 sparsity levels ranging from 0.88 to 0.99. The result was very
positive for the PBR method, as shown in Figure 9. At all the sparsity levels
and most of the purchasing-number levels, the PBR method performed better
than, or at least equal to, the collaborative fi ltering method.
Discussion and Further Research Issues
The experiments described in this paper clearly show the advantages of the
PBR method. Although it did not perform as well as the collaborative fi ltering
method when used with full ratings, it showed superior performance under
data sparsity and cold-starting situations for new buyers. Since sparsity is
very common, and there are always new buyers or buyers with only a few
purchases on record in Internet shopping malls, these performance charac-
teristics are very signifi cant.
Apart from these advantages, PBR has several desirable properties. First, it
is computationally much less complex than collaborative fi ltering in real-time
recommendation. For a given buyer, collaborative fi ltering requires similarity
calculation with N reference buyers using M
train
movies where both N and M
train
can be huge numbers in practice. For the calculation of similarity, N × M
train
iterations are required when using the widely adopted similarity metrics, such
as the Pearson correlation coeffi cient or the cosine measure. After the similar-
ity values are ready, N iterations are again required for each candidate movie
in order to generate predictions [25]. In contrast, with the PBR method, the
profi ling of a buyer requires only M
train
iterations, and in actual recommenda-
tion, it takes constant time for each candidate movie. The complexities of the
two methods are summarized in Table 5 using the Big O notation.
Another feature of the PBR method is that the recommendation process
is much more understandable. While the collaborative fi ltering method has
only limited ability to explain its recommendation results, the PBR method
provides explicit understanding of the buyer’s preferences for all the <g
i
, p
j
>
pairs and thus of the fact that each product has been recommended accord-
ing to these preferences. This capability can enable interesting further use
of the recommendation result. For example, if the recommendation result
is unsatisfactory for buyers with explicit preferences for certain genres and
03 ahn.indd 7203 ahn.indd 72 11/20/2006 10:26:12 AM11/20/2006 10:26:12 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 73
Figure 9. Experiments Under Cold-Starting Situation for New Buyers.
X axis: number of ratings used for profi ling (BPR) and similarity calculation (CF); Y axis: average of top fi ve recommendations.
03 ahn.indd 7303 ahn.indd 73 11/20/2006 10:26:12 AM11/20/2006 10:26:12 AM
74 HYUNG JUN AHN
Required number of iterations proportional
to problem size when N = 1,000 and
M
train
= 1,000
Similarity calculation Recommendation Similarity calculation Recommendation
or training (or prediction) or training (or prediction)
Collaborative fi ltering method O (N × M
train
) O (N) 1,000,000 1,000
PBR method O (M
train
) constant time 1,000 1
Table 5. Comparison of Complexities.
03 ahn.indd 7403 ahn.indd 74 11/20/2006 10:26:14 AM11/20/2006 10:26:14 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 75
popularity classes, this may mean that there are not enough candidate prod-
ucts that correspond to the pairs, and one can replenish them strategically.
The probabilities calculated for each <g
i
, p
j
> pair can also open up possibilities
for new ways of recommendation. For example, based on the probabilities in
Figure 10(a), a simple recommendation strategy can be implemented for a
buyer group that has shown explicit preference for horror movies in popular-
ity class 3. Similarly, the probabilities in Figure 10(b) can be utilized to choose
a target group for promotion of a specifi c comedy genre movie that belongs
to popularity class 7.
If the three popularity measures for a new product can be estimated with
reasonable accuracy, the explicit understandability of the PBR method may
improve recommendation performance in another type of cold-starting situ-
ation: recommendation of new products with little or no sales history. The
estimate is feasible in practice because companies often design or introduce
products based on explicit strategy or positioning strongly related to their
popularity characteristics. One can also attempt an estimate based on people’s
opinions, as when one consults a group of experts within or outside of an
organization who have experience of marketing similar products. This would
be an interesting issue for further research.
The PBR method has some limitations that require further investigation.
This paper used 10 as the maximum number of popularity classes to limit the
number of probabilities generated, maintain a certain level of reliability, and
make the recommendation process understandable. However, changing the
maximum number of classes may improve the recommendation performance.
Moreover, PBR may not be applicable to domains where the values of the
three dimensions of popularity cannot be estimated for each product because
of insuffi cient data or for some other reason. Also, the effectiveness of the
partitioning of the popularity space may be dependent upon shopping-mall
characteristics and thus require different approaches in different domains.
Experiments with different datasets are required for generalization of the
PBR method. Finally, since the PBR method generates only the ordering of <g
i
,
p
j
> pairs, it cannot compute individual prediction values for each candidate
product for each buyer. This is a shortcoming of the PBR approach compared
with collaborative fi ltering and other recommendation methods, and as such
it presents another challenging issue for further research.
Figure 10. Examples of Probabilities Between <g
i
, p
j
> Pairs
03 ahn.indd 7503 ahn.indd 75 11/20/2006 10:26:14 AM11/20/2006 10:26:14 AM
76 HYUNG JUN AHN
In addition to the research issues raised so far, there are several other in-
teresting topics worthy of consideration. First, since the collaborative fi ltering
method performs better when the number of ratings is large enough, a hybrid
recommendation method combining collaborative fi ltering and PBR could be
successful. Second, while this study only combined genre information with
popularity features, there are many other types of data that could potentially
bring synergy when used with popularity features. For example, certain
demographic features might be effective in combination with the popularity
features because people of different age groups or different cultures may be
affected differently by popularity. Third, different recommendation methods
could be applied for products in different popularity classes in order to as-
certain which method performs best in each class, and this might lead to the
development of another hybrid recommendation system. Fourth, different
domains or products may demonstrate different effects of using popularity
characteristics. Investigating what features of domains and products determine
the effectiveness of certain popularity features would be another meaningful
research subject.
Conclusion
This paper presents a novel approach to automated product recommenda-
tion using the popularity characteristics of products. The proposed model
of popularity was combined with movie genre information to build a hybrid
method called popularity-based recommendation. When compared with the
widely used collaborative fi ltering method, the PBR system showed signifi cant
improvement under data sparsity and cold-starting conditions. This outcome
demonstrates that PBR would be of great practical value for recommendation
in many Internet shopping malls. The other benefi ts of PBR include a more un-
derstandable recommendation process and less computational complexity.
The most signifi cant academic contribution of the research described in this
paper is that it constitutes the fi rst effort, to the author’s best knowledge, to
develop a model of popularity for recommender systems. Since the model, in
essence, presents a new way of utilizing a set of meaningful product features
that have been overlooked so far, the author believes that it may open up many
interesting opportunities for future research.
REFERENCES
1. Adomavicius, G.; Sankaranarayanan, R.; Sen, S.; and Tuzhilin, A. Incor-
porating contextual information in recommender systems using a multidi-
mensional approach. ACM Transactions on Information Systems, 23, 1 (2005),
103–145.
2. Ahn, A.; Kim, J.K.; Choi , I.Y.; and Cho, Y.H. A personalised recom-
mendation procedure based on dimensionality reduction and Web mining.
International Journal of Internet & Enterprise Management, 2, 3 (2004), 280–298.
3. Arnould, E.; Price, L.; and Zinkhan, G. Consumers. New York: McGraw-
Hill, 2002.
03 ahn.indd 7603 ahn.indd 76 11/20/2006 10:26:14 AM11/20/2006 10:26:14 AM
INTERNATIONAL JOURNAL OF ELECTRONIC COMMERCE 77
4. Berry, M.J.A., and Linoff, G.S. Data Mining Techniques: For Marketing,
Sales, and Customer Relationship Management. New York: Wiley, 2004.
5. Burke, R. Hybrid recommender systems: Survey and experiments. User
Modeling and User-Adapted Interaction, 12, 4 (2002), 331–370.
6. Cohen, W.W., and Fan, W. Web-collaborative fi ltering: Recommending
music by crawling the Web. Computer Networks, 33, 1–6 (2000), 685–698.
7. Cylogy. Personalization overview. www.cylogy.com/library/personaliza-
tion_overview-kb.pdf.
8. Greco, G.; Greco, S.; and Zumpano, E. Collaborative fi ltering supporting
Web site navigation. AI Communications, 17, 3 (2004), 155–166.
9. Han, P.; Xie, B.; Yang, F.; and Shen, R. A scalable p2p recommender sys-
tem based on distributed collaborative fi ltering. Expert Systems with Applica-
tions, 27, 2 (2004), 203–210.
10. Hayes, C., and Cunningham, P. Context boosting collaborative recom-
mendations. Knowledge-Based Systems, 17, 2–4 (2004), 131–138.
11. Herlocker, J.L.; Konstan, J.A.; and Riedl, J. Explaining collaborative fi l-
tering recommendations. In E.F. Churchill, J. McCarthy, C. Neuwirth, and T.
Rodden (eds.), Proceedings of the 2000 ACM Conference on Computer Supported
Cooperative Work. New York: ACM Press, 2000, pp. 241–250.
12. Herlocker, J.L.; Konstan, J.A.; Terveen, L.G.; and Riedl, J.T. Evaluating
collaborative fi ltering recommender systems. ACM Transactions on Informa-
tion Systems, 22, 1 (2004), 5–53.
13. Hofmann, T. Latent semantic models for collaborative fi ltering. ACM
Transactions on Information Systems, 22, 1 (2004), 89–115.
14. Kim, B.-D., and Kim, S.-O. A new recommender system to combine con-
tent-based and collaborative fi ltering systems. Journal of Database Marketing,
8, 3 (2001), 244–252.
15. Konstan, J.A.; Miller, B.N.; Maltz, D.; Herlocker, J.L.; Gordon, L.R.; and
Riedl, J. Grouplens: Applying collaborative fi ltering to usenet news. Com-
munications of the ACM, 40, 3 (1997), 77–87.
16. Kotler, P., and Armstrong, G. Principles of Marketing. Englewood Cliffs,
NJ: Prentice Hall International, 1991.
17. Lee, D.-S.; Kim, G.-Y.; and Choi, H.-I. A Web-based collaborative fi ltering
system. Pattern Recognition, 36, 2 (2003), 519–526.
18. Li, Y.; Lu, L.; and Xuefeng, L. A hybrid collaborative fi ltering method for
multiple-interests and multiple-content recommendation in e-commerce.
Expert Systems with Applications, 28, 1 (2005), 67–77.
19. Maltz, D., and Ehrlich, E. Pointing the way: Active collaborative fi lter-
ing. In I.R. Katz, R. Irvin, R.L. Mack, L. Marks, M.B. Rosson, and J. Nielsen
(eds.), Proceedings of the ACM CHI 95 Human Factors in Computing Systems.
New York: ACM Press, 1995, pp. 202–209.
20. Melville, P.; Mooney, R.J.; and Nagarajan, R. Content-boosted collabora-
tive fi ltering for improved recommendations. In R. Dechter, and R. Sutton
(eds.), Proceedings of the Eighteenth National Conference on Artifi cial Intelli-
gence. Menlo Park, CA: AAAI Press, 2002, pp. 187–192.
21. Middleton, S.E.; Alani, H.; Shadbolt, N.R.; and De Roure, D.C. Exploit-
ing synergy between ontologies and recommender systems. Paper present-
ed at the Eleventh International World Wide Web Conference, Honolulu,
HI, May 7–11, 2002.
03 ahn.indd 7703 ahn.indd 77 11/20/2006 10:26:14 AM11/20/2006 10:26:14 AM
78 HYUNG JUN AHN
22. Mirzadeh, N.; Ricci, F.; and Bansal, M. Feature selection methods for
conversational recommender systems. In X. Chu (ed.), Proceedings of 2005
IEEE International Conference on e-Technology, e-Commerce and e-Service, Los
Alamitos, CA: IEEE Computer Society, 2005, pp. 772–777.
23. MovieLens. Movielens dataset. www.grouplens.org.
24. Papagelis, M.; Plexousakis, D.; Rousidis, I.; and Theoharopoulos, E.
Qualitative analysis of user-based and item-based prediction algorithms for
recommendation systems. Paper presented at the 3rd Hellenic Data Man-
agement Symposium. Athens, June 2004.
25. Papagelis, M.; Rousidis, G.; Plexousakis, D.; and Theoharopoulos, E.
Incremental collaborative fi ltering for highly-scalable recommendation al-
gorithms. In Proceedings of the 15th International Symposium on Methodologies
for Intelligent Systems. New York: ACM Press, 2005, pp. 553–561.
26. Resnick, P.; Iacovou, N.; Suchak, M.; Bergstrom, P.; and Riedl, J. Grou-
plens: An open architecture for collaborative fi ltering of netnews. In R.
Furuta and C. Neuwirth (eds.), Proceedings of ACM 1994 Conference on Com-
puter Supported Cooperative Work. New York: ACM Press, 1994, pp. 175–186.
27. Sarwar, B.; Karypis, G.; Konstan, J.A.; and Riedl, J. Item-based collabora-
tive fi ltering recommendation algorithms. In V.Y. Shen, N. Saito, M.R. Lyu,
and M.E. Zurko (eds.), Proceedings of the Tenth International World Wide Web
Conference. New York: ACM Press, 2001, pp. 285–295.
28. Schein, A.I.; Popescul, A.; and Ungar, L.H. Croc: A new evaluation
criterion for recommender systems. Electronic Commerce Research, 5, 1 (2005),
51–74.
29. Schein, A.I.; Popescul, A.; Ungar, L.H.; and Pennock, D.M. Methods
and metrics for cold-start recommendations. In K. Järvelin, M. Beauliew, R.
Baeza-Yates, and S.H. Myaeng (eds.), Proceedings of 25th Annual International
ACM SIGIR Conference on Research and Development in Information Retrieval.
New York: ACM Press, 2002, pp. 253–260.
30. Vezina, R., and Militaru, D. Collaborative fi ltering: Theoretical positions
and a research agenda in marketing. International Journal of Technology Man-
agement, 28, 1 (2004), 31–45.
31. Witten, I.H., and Frank, E. Data Mining: Practical Machine Learning Tools
and Techniques with Java Implementation. San Francisco: Morgan Kaufmann,
2000.
32. Yu, K.; Schwaighofer, A.; Tresp, V.; Xu, X.; and Kriegel, H.-P. Probabilis-
tic memory-based collaborative fi ltering. IEEE Transactions on Knowledge &
Data Engineering, 16, 1 (2004), 56–69.
33. Zeng, C.; Xing, C.-X.; Zhou, L.-Z.; and Zheng, X.-H. Similarity measure
and instance selection for collaborative fi ltering. International Journal of Elec-
tronic Commerce, 8, 4 (2004), 115–129.
HYUNG JUN AHN (hjahn@waikato.ac.nz) is a senior lecturer at the Waikato Manage-
ment School, University of Waikato, New Zealand. He received his Ph.D. in 2004 from
the Graduate School of Management, Korea Advanced Institute of Science and Tech-
nology (KAIST), and taught two courses there as an adjunct professor before joining
the Waikato Management School. His research interests include multi-agent systems,
intelligent information systems, and e-supply-chain management. He has published
in Decision Support Systems, International Journal of Cooperative Information Systems, and
Expert Systems with Applications.
03 ahn.indd 7803 ahn.indd 78 11/20/2006 10:26:14 AM11/20/2006 10:26:14 AM
