M i n i n g t h e W e b

VISCORS: A Visual-
Content Recommender
for the Mobile Web
Chan Young Kim and Jae Kyu Lee, Korea Advanced Institute of Science and Technology

Yoon Ho Cho, Kookmin University

Deok Hwan Kim, Dongyang Technical College

A n increasing selection of content is becoming available in the mobile-Web envi-

ronment, where users navigate the Web using wireless devices such as cell phones

and PDAs. The fast growth and excellent prospects of the mobile-Web content market

have attracted many content providers. However, as continuing deregulation further

Current search lowers the entry barriers for providers, increased cient search aid that suggests only the images meet-
competition is quickly eliminating profit opportuni- ing the customer’s preference is necessary.
methods for mobile- ties. To survive in this environment, providers must As a solution to these problems, we propose VIS-
offer an intelligent system that provides customers CORS (Visual Contents Recommender System). To
Web content can be with a more pleasant mobile-shopping experience. reduce customers’search effort, VISCORS combines the
A particularly popular form of mobile-Web con- two most popular information-filtering techniques:
frustrating to use. To tent is wallpaper images for cell phones.1 Although collaborative filtering and content-based image
the market for this content is growing rapidly as related retrieval. Combined, these techniques properly han-
shorten searches for technologies evolve, customers experience much frus- dle the distinct characteristics of visual content while
tration when searching for the images they want, taking into account the mobile Web’s constraints.
cell phone wallpaper owing to inefficient sequential search (see Figure 1).
When a customer logs on to an image-download site Collaborative filtering
images, VISCORS using a cell phone, the site presents the customer with Recommender systems help customers find the
a list of the best-selling or newest images. The cus- items they’d like to purchase. The most successful
combines collaborative tomer pages through the list and selects an image to recommendation technique is collaborative filter-
inspect. If the customer likes the image, he or she ing.2,3 CF identifies customers (neighbors) whose
filtering with content- might buy it. Otherwise, the customer repeats these tastes are like those of the target customer and rec-
steps until he or she stumbles over the right image or ommends items those customers have liked.
based image retrieval. gives up. With this method, the expected number of However, CF has two major shortcomings. First,
images the customer views before hitting the desired when there’s a shortage of customer ratings, it suffers
image far exceeds the acceptable level. from a sparsity problem.3–6 Most similarity measures
These difficulties are partly attributable to the cell used in CF work properly only when there’s a suffi-
phone’s characteristics. Compared to PCs, cell cient number of ratings on common items from sim-
phones have smaller screens, fewer input keys, and ilar customers. An increase in the number of customers
less sophisticated browsers. So, the user interface of and items worsens this problem because the likeli-
mobile-Web applications isn’t as friendly as that of hood of different customers rating common items
typical Web applications. Consequently, many cus- decreases. Such sparsity in ratings makes the forma-
tomers use their PCs to select images and then request tion of a neighborhood (a group of customers with
a download to their mobile devices. Nevertheless, similar tastes) inaccurate, thereby resulting in poor
searching generally remains inconvenient and com- recommendations. Second, CF suffers from a new-
plex. To make searching more acceptable, a more effi- item problem.4,6 Because CF recommends an item on

32 1541-1672/04/$20.00 © 2004 IEEE IEEE INTELLIGENT SYSTEMS

Published by the IEEE Computer Society

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
the basis of previous customers’ratings of that
item, it doesn’t recommend a newly intro-
duced item until sufficient ratings of that item
are available.
To address these shortcomings, researchers
have proposed many variations of hybrid Image Buy
approaches that combine CF with content- selection or quit
Start End
based filtering.4–6 Content-based filtering rec-
ommends items with properties similar to
those of items the target customer liked in the
past. Despite these approaches’ success in
some applications, none of them is adequate
for wallpaper image recommendation because
of visual content’s distinct characteristics. Back to the image list
A customer’s preference of images is
ambiguous and more changeable over time Figure 1. A typical user interface for downloading wallpaper images for cell phones.
than that of the usual items, because the same
customer might perceive the same image dif-
ferently at different times.7 So, CF that rec- images and their representation in low-level preferred set. CBIR can use the preferred
ommends items entirely on the basis of the features. For CBIR to handle this semantic set’s images to refine the query to learn the
customer’s past preferences yields lower- gap, it needs the ability to learn about the cus- customer’s current preference.
quality recommendations for images than it tomer’s true intention through iterative inter- Because CBIR operates generally in the
does for ordinary items. Any hybrid approach actions. The customer’s preference regard- PC-based Web environment, applications
without a countermeasure to this drawback ing presented images needs to be fed back so that use CBIR typically assume that the user
can’t give acceptable results for image rec- that CBIR learns from this preference to interface can display many images at once,
ommendation. A mechanism for learning retrieve, in the next iteration, images more eliciting multiple levels of preference feed-
about the customer’s current preference is similar to the one the customer really wants. back. However, these assumptions don’t
essential to deliver good recommendations. This learning process, preference feedback, apply to mobile-Web devices. So, CBIR that
is essential for faster search. takes into account mobile-Web devices’ user
Content-based image retrieval To express an image’s degree of prefer- interface constraints is necessary.
CBIR, the most common image retrieval ence, you can use binary weights (for exam-
technique, uses images’ visual features to ple, preferred or unpreferred) or multilevel Viscors overview
retrieve images similar to the given query.7–9 weights (for example, highly preferred, pre- VISCORS consists of a CF module and a
However, its effectiveness is limited because ferred, neutral, unpreferred, or highly unpre- CBIR module (see Figure 2). The CF mod-
of the gap between high-level concepts of ferred).7 We call a set of preferred images a ule produces the initial list of recommended

CF-based
Collaborative filtering recommendations
(CF) module
Customer decision
(further search,
go back, buy)

Initial
Mobile-Web interface

query
Purchase Customer Image
database Profile database
Mobile
Web
Mobile
Preference Visual customer
information feature
CBIR-based
Content-based image recommendations
retrieval (CBIR)
module
Customer decision
(preferred,
unpreferred, buy)

Figure 2. VISCORS (Visual Contents Recommender System) architecture.

NOVEMBER/DECEMBER 2004 www.computer.org/intelligent 33

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
M i n i n g t h e W e b

List of recommended Selected Preference feedback obtained purchase and preference informa-
images image Retrieved images
tion to reflect dynamically the customer’s
most recent preference. This is significantly
different from the customer profiles that tra-
Buy ditional CF techniques use.2,3
Start End Given the customer profile P, the CF-
Further Quit
based recommendation procedure for a target
search
customer c takes two steps.

Step 1: Customer neighborhood

Back to the CF list
formation

Back to the CF list if no images are declared as preferred We use sim(a, b) to denote the similarity
between two customers a and b. First, we
Figure 3. VISCORS iterative search process. determine the neighborhood H = {h1, h2, ...,
hm} such that c ∉ H and sim(c, h1) is the high-
est similarity, sim(c, h2) is the next highest,
images (the left screen shot in Figure 3). For in the next iteration of retrieval. These itera- and so on. We calculate the similarity using
new customers,VISCORS generates the list using tions continue until the customer finds the the Pearson-r correlation2:
the best-seller-based method; otherwise, it uses desired image or quits the application.
∑ j =1( paj − pa )( pbj − pb )
N
CF. The customer skims through the list to see Because VISCORS adds the fed-back pref-
sim ( a, b ) = corrab =
if any images are of interest. Then, the cus- erence and purchase information to the cus-
∑ j =1( paj − pa ) ∑ j =1( pbj − pb )
N 2 N 2

tomer selects an entry to view the image (see tomer profile as ratings, it alleviates the spar-
the center screen shot in Figure 3). After view- sity problem. Because the CBIR module (2)
ing the image, the customer might decide to recommends images on the basis of their
purchase it, use it as a query for a CBIR-based visual features, VISCORS can recommend where N is a total number of images, paj and
search of similar images, or go back to the CF- images with no customer ratings, thus elim- pbj are customer a’s and b’s ratings on image
generated recommendation list. inating the new-item problem. In this way, j, and pa and pb are customer a’s and b’s
If the customer decides to use the viewed our hybrid approach leads to higher-quality average ratings on all images.
image as a query, the CF module passes that recommendations.
image to the CBIR module. For all images in Step 2: Recommendation
the database, the CBIR module calculates CF-based recommendation generation
their distances from the query and generates Our customer profile is the matrix of pref-
a list of the most similar images. This module erence ratings P = (pij): PLS(c, j) denotes the purchase likeliness
retrieves k images as recommendations, pre- score of the target customer c for image j. We
sents them to the customer one by one, and + 2 if customer i has purchased image j generate a list of n images, R = {r1, r2, …,

+ 1
interactively elicits the user’s preferences. pij = 
if customer i has marked image j as preferred rn}, such that rj ∉ {the images that c has
 −1 if customer i has marked image j as unpreferred
(The right screenshots in Figure 3 illustrate  already purchased} and PLS(c, r1) is the
 0 if customer i has not seen image j
the case of k = 3, where the system is request- (1) highest PLS, PLS(c, r2) is the next highest,
ing customer feedback on three recom- and so on. We compute the PLS as
mended images.) For each of the k images, where i = 1 to M, j = 1 to N, M is the total num-
the customer must declare whether he or she ber of customers, and N is the total number of ∑ ( pij − pi ) × sim (c, i )
prefers it (that is, the CBIR module uses images. As Equation 1 shows, the rating PLS ( c, j ) = i ∈H
binary preference information). At any point matrix’s cells have four possible values. We ∑ sim (c, i ) (3)
in this presentation session, the customer can rate the previously purchased images highest i ∈H

also decide to buy an image or quit. because they should reflect the customer’s
After the customer makes all k declarations, taste the most strongly. The preferred and CBIR-based recommendation
the CBIR module updates the preference unpreferred images have ratings that are one During CBIR, VISCORS continuously
information and purchase databases with the half or a negative one half, respectively, of the refines the query to reflect the customer’s lat-
related information, for the CF module to use magnitude of the purchased ones. (VISCORS est preference. VISCORS uses a multipoint
later. If the customer declares all k images as records the initial CF-recommended image as query, because recent studies have indicated
unpreferred, the search session returns to the preferred or unpreferred depending on that such a query handles the semantic gap
CF-generated recommendation list. Other- whether the customer selects it as an initial better than a single-point query.8,9
wise, the CBIR module learns the customer’s query for CBIR.) We assign −1 to unpreferred There are two different approaches in
current preference using the preferred set and images so that the CBIR module will recom- refining a query using multiple query points.
applies this information to refine the query and mend unseen images over unpreferred ones. The first approach clusters example images
update the distance function. It then uses the The CBIR module constantly replaces rat- and uses a centroid of each cluster as a query
refined query and updated distance function ings in the customer profile with newly point.8,9 This approach works fine when

34 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
enough example images are available. How- 1. Initialization
ever, it’s difficult to use with a very small a. Set Q to the image selected from the CF list.
number of example images. b. Set the preferred set Rp and the accumulated preferred set during the search session Ra
The other approach uses every example to empty sets.
image in a query as a query point.10 This c. Apply equal weights by setting ws to 1 for s = 1, 2, …, S.
approach offers a breakthrough in cases with 2. Image retrieval
few example images. In the mobile-Web Retrieve k images, R = {r1, r2, …, rk}, such that Dist(r1, Q) is the lowest distance between
environment, having a preferred set of suf- an image and the query, Dist(r2, Q) is the next lowest, and so on.
ficient size is difficult because the small 3. Query refinement and distance function update
screens make it impossible to obtain prefer- a. For each rj in R
ence information on multiple images in one Recommend rj to the target customer c.
interaction. So, VISCORS refines a query by If rj is marked as preferred, add rj to Rp and Ra.
replacing its query points with the newly fed- Endfor
back preferred images. b. Refine Q by replacing the images in Q with the ones in Rp.
Because a query in VISCORS can have mul- c. Update Dist(x, qj) by recalculating ws for s = 1, 2, …, S using Ra.
tiple query points, the distance function d. Set Rp to an empty set.
between an image x and a query Q should e. Go to step 2 for the next iteration.
aggregate multiple distance components
from the image to related query points. VIS- Figure 4. The three steps of CBIR-based recommendation.
CORS uses this aggregate distance function:

ter reflects a customer’s current preference customer profile. Each participant had pur-
g
Dist ( x , Q ) = by allowing different weights by dimension chased an average of 7.4 images. Before the
∑ j =11 dist 2 ( x, q j )
g
(4) and emphasizing the features with smaller participants started a search session, we
variance. asked them to select a target image to search
Figure 4 shows the three steps of CBIR- for. Using the PC Web interface instead of
where g is the number of query points in a based recommendation. the cell phone interface let participants nav-
query Q, qj is the jth query point of Q, and igate freely for a target image. A search ses-
dist(x, qj) is a distance function between an Evaluating Viscors sion continued until the system returned the
image x and a query point qj. We derived To evaluate the performance of VISCORS, target image. For each session, we logged
Equation 4 from FALCON’s formula.10 This we developed a Web-based application sys- information on target images, recommended
aggregate distance function supports a dis- tem running on a PC with exactly the same images, and images marked as preferred or
junctive query that captures high-level seman- user interface as the cell-phone-based VIS- unpreferred for later analysis.
tics of images better than a conjunctive query, CORS system. We wanted to answer two main Figure 5 illustrates how a participant finds
especially when preferred images are widely questions: a target image (the bouquet of roses in Fig-
spread out in the feature space (that is, a case ure 5a) in an actual search session, for k = 3.
of a high level of heterogeneity). This enables • How much performance improvement In Figure 5b, the CF module generates a list
faster search of the desired images. does VISCORS deliver compared to other of recommendations and presents the list’s
Just as VISCORS continuously refines the recommender systems? first page. The participant views the first
query in CBIR, it also updates Equation 4 to • How does CF’s effectiveness affect over- image, determines that it isn’t similar to the
reflect the customer’s current preference. For all performance? bouquet image, and decides to return to the
this purpose, we define the dist(x, qj) in Equa- CF-generated list.
tion 4 as The experiments In Figure 5c, the participant views the sec-
For the experiments, we used the 230 wall- ond image (of lilies) on the CF list and deter-
( ) ∑ Ss =1 ws ( xs − q js )2
dist x , q j =
(5)
paper images that Korea Telecom Freetel mines that it’s somewhat similar to the tar-
(KTF), a leading Korean CDMA (code divi- get image because it consists entirely of
sion multiple access) carrier, offered at the flowers. The participant then initiates further
where S is the number of dimensions of the time of our experiment. To characterize searching using the image as a starting
feature space, ws is a weight of the sth images, VISCORS used three color moments image. In Figure 5d, VISCORS uses the lilies
dimension in the feature space, and xs and based on HSV (hue, saturation, value), a well- image to retrieve three images with similar
qjs are coordinates of an image x and a query known visual feature. That is, we calculated visual content from the image database, and
point qj on the sth dimension. The relative the mean, standard deviation, and skewness presents them in sequence. The participant
weight for the sth feature, ws in Equation 5, of HSV values of all pixels to represent declares that all three images are unpreferred
is 1/σs, where σs is a standard deviation of images as vectors in a 9D feature space. and returns to the CF-generated list.
coordinates of the sth dimension of images. The experiment involved 200 mobile-Web In Figure 5e, the participant views the third
We calculate σs using all images in the pre- customers who had previously purchased image on the CF list and determines that the
ferred set accumulated during the search wallpaper images from KTF. Their past pur- presented image of pinkish flowers is some-
session. This distance function update bet- chase information was stored as an initial what similar to the target image. The partici-

NOVEMBER/DECEMBER 2004 www.computer.org/intelligent 35

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
M i n i n g t h e W e b

(a) (b) (c) (d) (e)

(f) (g) (h)

Figure 5. A customer’s actual search session (part of some screens are in English for this article): (a) the target image; (b) first CF
view; (c) second CF view; (d) first CBIR view; (e) third CF view; (f) second CBIR view; (g) third CBIR view; (h) fourth CBIR view.

pant then initiates further searching using the seller-based system and a typical CF-based ducted two statistical tests. One was a t-test
image as a starting image. In Figure 5f, VIS- recommender system (pure-CF). The pure- comparing the average performance of all three
CORS uses the image to retrieve three images, CF procedure is identical to that of CF-based systems. The other was a two-way ANOVA
which it presents in sequence. The participant recommendation that VISCORS uses, except (analysis of variance) test with repetition to
declares the first and third as preferred because that it uses only purchase information to assess how each period and k affected VPS.
they’re images of flowers in similar colors, build a customer profile.
and declares the second as unpreferred. Each participant performed three search Results and discussion
In Figure 5g, VISCORS uses the two pre- sessions per period for five periods, with a As Table 1 shows, the average VPS of
ferred images to retrieve three similar images, different target image each session. This was VISCORS is about 38 percent lower than pure-
which it presents in sequence. The participant to see how system’s performance changes CF and 52 percent lower than the best-
declares them all as preferred. In Figure 5h, over time. For each target image, participants seller-based system, at a significance level of
VISCORS again retrieves and presents three repeated the experiment using the three dif- 1 percent. (That is, VISCORS produced per-
images in the same manner using the three ferent recommender systems. In the case of formance gains of 38 and 52 percent.) Table
preferred images. The participant declares the VISCORS, participants performed the experi- 1 also shows that the rate of improvement in
first as preferred and the second as unpre- ment for k = 3, 5, and 7 to see how changes VPS over the five periods (that is, the sys-
ferred. The third image is the target image. in k affect the overall performance. Because tem’s learning speed) is 25 percent for VIS-
For this evaluation, we devised the views the quality of CF recommendations varies CORS, 18 percent for pure-CF, and 8 percent
per success metric, which is the number of with the neighborhood’s size,3 we performed for the best-seller-based system. These
images a customer views before he or she an initial experiment of a single period to results indicate that VISCORS offers not only
purchases an image. Basically, VPS mea- determine the optimal size. A neighborhood the lowest VPS but also the fastest perfor-
sures the customer’s effort for a successful size of 30 yielded the best performance, so mance improvement. Its superior perfor-
search. We used VPS to compare VISCORS to we used that size in our other experiments. mance over pure-CF stems from its acceler-
two other recommender systems: a best- To analyze the experiment’s results, we con- ated learning of customer preference from

Table 1. Performance comparison of VISCORS and two benchmark systems. VPS (views per success) measures the number
of images a customer views before purchasing an image.
Performance (VPS) VPS reduction over
System Period 1 Period 2 Period 3 Period 4 Period 5 Average five periods (%) t value
VISCORS (V) 23.01 22.32 19.96 18.57 17.24 20.22 25
Pure-CF (C) 37.95 33.05 31.48 31.19 31.00 32.93 18 −106.77‡
(39%)* (32%) (37%) (40%) (44%) (38%)
Best-seller-based system (B) 43.98 42.64 41.12 41.89 40.66 42.06 8 -66.49‡
(48%)† (48%) (51%) (56%) (58%) (52%)

* The performance gain of VISCORS over pure-CF is (C − V)/C.

† The performance gain of VISCORS over the best-seller-based system is (B − V)/B.
‡ p < 0.01. p is the probability that the null hypothesis is true.

36 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
the additional preference-rating information
25
fed back from CBIR. This indicates that VIS-
CORS successfully overcomes CF’s sparsity 24 k=3
problem. k=5

Average views per successful search

As we mentioned before, search sessions 23 k=7
in the experiment lasted until the system 22
found the predetermined target image. How-
ever, in a real environment, a search session 21
might end when the system retrieves an
20
image somewhat close to the customer’s
desired image. This suggests that customers 19
in a real environment would be able to search
for images with much less effort than our 18
results indicate. 17
According to the ANOVA results, the varia-
tion in VPS over the five periods (that is, the 16
period effect) is significant (F = 292.46,
15
p < 0.01). The F statistic provides a test for the 1 2 3 4 5
statistical significance of the observed VIS- Period
CORS performance differences over periods. A
large value for F indicates that performance
Figure 6. The overall performance of VISCORS.
varies by period. p is the probability that the
null hypothesis—in this case, that the perfor-
mance of VISCORS doesn’t vary by period—is more preferred images. As Part A of Table 2 Third, the VPS that CBIR consumes de-
true. Figure 6 illustrates the variation as a shows, the average rate of viewed images on creases because the likelihood increases that
decreasing curve. As periods progress, more the CF list’s first page (that is, the top nine the target image is in the CF list’s upper part.
rating information becomes available. When recommended images) that become an ini- As Part C of Table 2 shows, the rate of the
the customer profile contains more ratings, tial query of CBIR increases from 0.33 in target image being on the CF list’s first page
neighborhood formation becomes more ac- Period 1 to 0.39 in Period 5. increases from 0.18 in Period 1 to 0.28 in
curate, thereby improving the quality of CF Second, the number of iterations of CBIR Period 5.
recommendations. per success decreases because the first image The two-way ANOVA results also indicate
The results in Table 2 show that better CF used in CBIR is closer to the target image in that the variation in VPS due to k (that is, the
recommendations help decrease VPS in three the feature space. As Part B of Table 2 shows, k effect) is significant (F = 4.52, p < 0.05). In
ways. First, the images on the CF list are used the average number of iterations of CBIR per addition, the interaction between the period
more often as an initial query for CBIR success decreases from 4.23 in Period 1 to and k significantly affects performance (F =
because the CF list’s upper part contains 3.66 in Period 5. 4.09, p < 0.01). Figure 6 illustrates this as the

Table 2. Collaborative filtering's effect over five periods. k is the number of retrieved images per iteration of
content-based image retrieval.
Period
Effect k 1 2 3 4 5
A. The rate of viewed images on the CF list's first page that 3 0.33 0.34 0.37 0.35 0.37
become an initial query of content-based image retrieval 5 0.31 0.34 0.36 0.37 0.40
7 0.35 0.31 0.36 0.37 0.41
Average 0.33 0.33 0.36 0.36 0.39
B. Iterations of CBIR per success 3 6.20 5.96 5.37 5.30 5.08
5 3.81 3.67 3.45 3.50 3.40
7 2.69 2.60 2.54 2.53 2.51
Average 4.23 4.08 3.79 3.78 3.66
C. The rate of the target image being found on the CF list's 3 0.18 0.20 0.24 0.25 0.28
first page 5 0.18 0.21 0.24 0.24 0.28
7 0.18 0.20 0.23 0.25 0.27
Average 0.18 0.20 0.24 0.25 0.28
D. The ratio of the image views that CF consumed to the total 3 0.56 0.56 0.55 0.54 0.53
views per success 5 0.48 0.50 0.48 0.47 0.45
7 0.45 0.47 0.45 0.44 0.42
Average 0.50 0.51 0.49 0.48 0.47

NOVEMBER/DECEMBER 2004 www.computer.org/intelligent 37

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
M i n i n g t h e W e b

the results also suggest two aspects of VISCORS

T h e A u t h o r s to consider for its real-world application.
First, the most important factors influenc-
Chan Young Kim is an associate professor of Internet business at Dongyang
ing CF’s effectiveness are the customer pro-
Technical College and is a PhD candidate in the management information
systems program at the Korea Advanced Institute of Science and Technology. file’s sparsity level and heterogeneity in cus-
His research interests include recommender systems, e-business, and mobile tomers’ buying behaviors.3,5 In cases with a
business. He received his BS in electrical engineering from Seoul National high level of sparsity or heterogeneity, CF
University and his MBA from the University of Chicago. Contact him at the performs poorly, so we recommend using a
Dept. of Internet Business, Dongyang Technical College, 62-160 Kochuk,
Kuro, Seoul, 152-714, Korea; [email protected].
higher value of k. In other cases, because k’s
impact on performance is insignificant, you
Jae Kyu Lee is a professor at the Korea Advanced Institute of Science and can select k’s value by considering other fac-
Technology’s Graduate School of Management and a director of the Inter- tors such as the customer’s burden of feed-
national Center for Electronic Commerce. His research interests include e- back in CBIR or the service provider’s mar-
commerce, the Semantic Web, recommender systems, data mining, and expert
systems. He’s the editor in chief of Electronic Commerce Research and Appli- keting strategy concerning the diversity of
cations. He received his PhD from the Wharton School, University of Penn- recommended images.
sylvania. Contact him at the Graduate School of Management, Korea Second, our experiment used a less het-
Advanced Institute of Science and Technology, 207-43 Cheongryangri, erogeneous image database consisting of a
Dongdaemun, Seoul, 130-012, Korea; [email protected].
few images belonging to a small number of
Yoon Ho Cho is an assistant professor at Kookmin University’s School of categories (for example, love, friendship, the
E-Business. His research interests include recommender systems, e-busi- bizarre, and so on). So, we used only the color
ness, mobile business, customer relationship management, and data mining. moment for image characterization even
He received his PhD in management information systems from the Korea though we knew that CBIR’s performance
Advanced Institute of Science and Technology. Contact him at the School
of E-Business, Kookmin Univ., 861-1 Jungnung, Sungbuk, Seoul, 136-702, improves as it considers more visual features.
Korea; [email protected]. However, using a single feature, as our exper-
iment did, might not give such positive results
Deok Hwan Kim is an associate professor of mobile Internet at Dongyang
in real-world applications, where the image
Technical College. He’s now visiting the University of Arizona’s Computer database’s heterogeneity will likely be higher.
Science Department. His research interests include multimedia information Therefore, in a real-world implementation of
retrieval, multimedia databases, recommender systems, and data mining. He VISCORS, we recommend expanding the fea-
received his PhD in information and communication engineering from the ture set to include other widely used features
Korea Advanced Institute of Science and Technology. Contact him at the
Dept. of Mobile Internet, Dongyang Technical College, 62-160 Kochuk, such as texture and shape.
Kuro, Seoul, 152-714, Korea; [email protected].

difference in VPS for the three values of k in

V ISCORS offers two main benefits. First,
customers can purchase content with
much less search effort and much lower con-
the early periods and as the diminishing dif- nection time because they can much more
ference in VPS over all five periods. easily find the desired content. Second,
As Part D of Table 2 shows, CF’s relative mobile-Web content providers can improve
importance to overall performance decreases their businesses’ profitability because lower
as k increases (for example, from 0.56 when customer frustration in finding desired con-
k = 3 to 0.45 when k = 7 in Period 1). As we tent increases revenue through an improved
UPCOMING mentioned previously, CF’s effectiveness purchase conversion rate. (This rate consti-
ISSUES: also improves as the periods progress. So, a tutes the number of search sessions that end
performance with a lower k should improve with a purchase divided by the total number
Intelligent faster than that with a higher k. This means of search sessions.)
Manufacturing that the k effect is obvious when CF’s per- As mobile-Web services rapidly grow,
formance is poor and gradually diminishes mobile-Web-based recommender systems for
Control as CF improves. An additional one-way other types of multimedia content, such as
ANOVA analysis confirms the claim with its music on demand or video on demand, will
Planning with results of (F = 13.54, p < 0.01) on the k effect increasingly be an area of research interest.
Templates for the first two periods and (F = 1.10, Our approach is applicable to these types of
p = 0.33) for the last three periods. content as long as you can represent the con-
Intelligent Systems From these results, we conclude that VIS- tent as a vector in the feature space, as in the
in Government CORS is a viable solution to the problems case of wallpaper images. However, users can
Visit us on the Web at encountered when customers download wall- easily provide preference feedback on wall-
paper images on the mobile Web. However, paper images after a short viewing; this might
www.computer.org/intelligent
38 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.
not be true for music or videos. So, success-
ful application of our approach to these types
For more on this or any other computing topic,
of content will require research on the proper see our Digital Library at www.computer.org/
interfaces for preference feedback. a monthly magazine of
publications/dlib. the IEEE Computer Society

Acknowledgments

h t t p : / / d s o n l i n e . c o m p u t e r. o r g
The Korea Science & Engineering Foundation’s
Postdoctoral Fellowship Program partially sup- IEEE Distributed Systems Online
ported this work.
brings you peer-reviewed articles, detailed

References tutorials, expert-managed topic areas, and

1. Korea Internet White Paper 2002, Korea Net- diverse departments covering the latest news
work Information Center, 2002.

2. U. Shardanand and P. Maes, “Social Infor- and developments in this fast-growing field.
mation Filtering: Algorithms for Automating
‘Word of Mouth,’” Proc. 1995 ACM/SIGCHI
Conf. Human Factors in Computing Systems
(CHI 95), Addison-Wesley, 1995, pp. Log on http://dsonline.computer.org
210–217.
for free access to topic areas on
3. B. Sarwar et al., “Analysis of Recommenda-
tion Algorithms for E-Commerce,” Proc.
ACM Conf. Electronic Commerce (EC 00),
ACM Press, 2000, pp. 158–167. ❍ Grid Computing
4. M. Balabanovic and Y. Shoham, “Fab: Con- ❍ Mobile & Pervasive
tent-Based Collaborative Recommendation,”
Comm. ACM, vol. 40, no. 3, 1997, pp. 66–72.
❍ Distributed Agents
5. Y.H. Cho, J.K. Kim, and S.H. Kim, “A Per-
sonalized Recommender System Based on ❍ Security
Web Usage Mining and Decision Tree Induc-
tion,” Expert Systems with Applications, vol.
23, no. 3, 2002, pp. 329–342.
❍ Middleware
6. P. Melville, R.J. Mooney, and R. Nagarajan, ❍ Parallel Processing
“Content-Boosted Collaborative Filtering for
Improved Recommendations,” Proc. 18th
Nat’l Conf. Artificial Intelligence (AAAI 02),
❍ Web Systems
AAAI Press, 2002, pp. 187–192.
❍ Real Time & Embedded
7. X.S. Zhou and T.S. Huang, “Relevance Feed-
back for Image Retrieval: A Comprehensive ❍ Dependable Systems
Review,” ACM Multimedia Systems J., vol. 8,
no. 6, 2003, pp. 536–544.
❍ Cluster Computing
8. D.H. Kim and C.W. Chung, “Qcluster: Rele-
vance Feedback Using Adaptive Clustering ❍ Distributed Multimedia
for Content-Based Image Retrieval,” Proc.
2003 ACM SIGMOD Int’l Conf. Management
of Data (SIGMOD 03), ACM Press, 2003, pp. ❍ Distributed Databases
599–610.
❍ Collaborative Computing
9. K. Porkaew, K. Chakrabarti, and S. Mehro-
tra, “Query Refinement for Multimedia Sim-
ilarity Retrieval in MARS,” Proc. 7th ACM
❍ Operating Systems
Int’l Multimedia Conf. (MM 99), Part 1, ACM
Press, 1999, pp. 235–238. ❍ Peer-to-Peer
10. L. Wu et al., “FALCON: Feedback Adaptive
Loop for Content-Based Retrieval,” Proc.
26th Int’l Conf. Very Large Data Bases
(VLDB 00), Morgan Kaufmann, 2000, pp.
297–306. To receive regular updates, email
NOVEMBER/DECEMBER 2004
[email protected]

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on November 04,2020 at 09:41:42 UTC from IEEE Xplore. Restrictions apply.