Toward a Product Search Engine based on User Reviews
Paolo Fosci and Giuseppe Psaila
Universit
`
a di Bergamo, Viale Marconi 5, I-24044 Dalmine, BG, Italy
Keywords:
Product Search Engine, User Reviews, Itemset Mining.
Abstract:
We address the problem of developing a method for retrieving products exploiting product user-reviews that
can be found on the internet. For this purpose, we introduce a ranking model based on the concept of item-
set mining of frequent terms. The prototype search engine that implements the proposed retrieval model is
illustrated, and a preliminary evaluation on a real data set is discussed.
1 INTRODUCTION
”What is a movie with great funny hilarious jokes?”
among a list of thousands of movies that could be
watched on an on demand TV? Similar questions
could be stated by a user for any kind of product or
service, but an incredible source of opinions about
products is represented by user reviews, i.e., com-
ments posted by other people.
In the era of Web 2.0, user reviews become an im-
portant source of reputation, that other users could ex-
ploit to find out the best product they are looking for,
or to avoid the worst ones. Therefore, a novel search
engine could be built, in order to exploit user reviews
to rank products: a person that wants to find prod-
ucts with given characteristics or opinions might sub-
mit a text query containing the words that better char-
acterize the properties and related opinions he/she is
looking for; as a result, the search engine produces a
ranked list, where the ranking measure is based on the
content of reviews.
This paper is a first step toward the development of
a product search engine based on users reviews. The
main idea behind our work is the way user reviews are
processed to summarize their content. Common car-
rying opinion words are often used by review writers
when they have more or less the same opinion about
a product; thus, the extraction of sets of words that
are frequently used by writers of reviews concerning
the same product can be seen as a problem of frequent
itemset mining.
Thus, the set of extracted itemsets gives a summa-
rized view of the reviews. The set of extracted itemset
can be queried to find the best products that match the
text query.
In this paper, we present the main ideas about the
exploitation of itemsets to summarize product reviews
and query the collection of products. In particular, we
present a novel retrieval model based on this concept,
whose goal is to define a suitable ranking measure
for products based on how the user query matches the
content of product reviews. At this point, an evalua-
tion is performed, where we discuss the effectiveness
of our approach.
This paper is organized as follow. Section 2
presents a short survey of related works. Section 3
adapts the concept of itemset mining to our context.
Section 4 introduces the proposed retrieval model.
Section 5 presents a preliminary evaluation on a real
data set. Finally Section 6 draws the conclusions.
2 RELATED WORK
Our research work is in the middle of several research
areas: data mining, recommendation systems, search
engines. In the last 2 decades, an incredible amount
of research works have been published in the above
mentioned areas. We do no not want to report an ex-
haustive review of the last twenty years, but focus on
the topic of our work, considering the recent works
that, we think, better represent the current state of the
art of related work.
In the area of recommendation systems, several
works exploited data mining techniques for building
a model of user preferences and use this model to rec-
ommend products to potential customers. In partic-
ular, concerning recommendation systems that incor-
porate data mining technique, we can cite (Sandvig
et al., 2007), that exploits association rule mining.
223
Fosci P. and Psaila G..
Toward a Product Search Engine based on User Reviews.
DOI: 10.5220/0004051602230228
In Proceedings of the International Conference on Data Technologies and Applications (DATA-2012), pages 223-228
ISBN: 978-989-8565-18-1
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
Several works (Hu and Liu, 2004b; Liu et al.,
2005; Kim et al., 2009) adopt association rule mining
for analyzing customer reviews and extract opinion
from them. In (Hu and Liu, 2004b), association rule
mining is used to extract, from within customer re-
views, relevant features that characterize opinions of
users about products. In (Liu et al., 2005), a system to
compare opinions about products is presented, where
product reviews reports PROs and CONs; in partic-
ular, association rule mining is exploited to assign a
positive or negative polarity to words (namely, adjec-
tives) in product reviews, and use this polarity to rank
the opinion about products (note that we are not in-
terestd in polarity, itemsets are a mean to summarize
text). The work in (Hu and Liu, 2004a) extracts, by
means of an association rule mining technique, rele-
vant features that summarize product reviews.
Other data mining techniques are used to analyze
customer reviews. For instance, in (Ly et al., 2011)
a sentence clustering technique is adopted. In (Chao-
valit and Zhou, 2005), both supervised and unsuper-
vised approaches are evalauted. A similar work is
done in (Lee et al., 2008), where key classification
techniques for opinion mining are discussed. The
work in (Dave et al., 2003) presents a technique for
semantic classifications of product reviews.
W.r.t. all these works, that adopts data mining, and
association rule mining in particular, to summarize re-
views or to summarize/discover/synthesize opinions
about products, we exploit itemset mining for a dif-
ferent purpose. In our proposal, itemset mining is the
basis for a retrieval model: words in the query are
matched against product reviews to find out the prod-
ucts with reviews that better matches the query words.
The ranking of products is performed by evaluating a
small subset of all possible itemsets and their support
is aggregated in a relevance value, used to produce the
final ranked list of products.
For the best of our knowledge, no similar work has
been done yet.
3 ITEMSET MINING AND
PRODUCT REVIEWS
Our approach to product retrieval is based on the
widely exploited notion of itemset mining. Originally
(Agrawal and Srikant, 1994), itemset mining is the
basic and computationally difficult step for associa-
tion rule mining. Here, we are not interested in as-
sociation rule mining, but only in itemset mining for
analyzing sets of words that frequently occur together
in reviews. In this Section, we first present the basic
notions of itemset mining. Then, we adapt the con-
cept to our context.
3.1 Basic Notions on Itemset Mining
The notion of itemset mining was introduced as a
fundamental part of the process for mining associa-
tion rules (Agrawal et al., 1993; Agrawal and Srikant,
1994). In a transaction database Z = {z
1
,...,z
n
}, a
transaction z is a set of items z = {i
1
,...,i
k
} sold to-
gether in transaction z.
A k−itemset I = {i
1
,...,i
k
} is a set of k items.
The support of I, denoted as s(I) is the frequency with
which the itemset occurs in the transaction database,
i.e., s(I) = |Z(I)|/|Z| (where Z(I) ⊆ Z is the set of
transactions z ∈ Z such that z ∩ I = I, i.e., the transac-
tions that contain itemset I).
Given a minimum support threshold s, an itemset I
is said large if its support is no less than the threshold
s, i.e., s(I) ≥ s.
The problem known as itemset mining is the prob-
lem of extracting large itemsets from the transaction
database Z, provided that a minimum threshold for
itemset support is defined.
Notice that this problem has been widely and ex-
tensively studied in the last two decades. As far as
the main results about efficient algorithms are con-
cerned, we just cite the fundamental paper (Agrawal
and Srikant, 1994), in which the Apriori algorithm
was defined, that has enabled itemset mining on large
databases.
3.2 Product Reviews and Itemsets
Consider now a product p (a movie, a camera, etc.)
and a set of reviews R(p) = {r
1
,...,r
k
}, where each
review is a text, i.e., a sequence of term occurrences
r
i
=< t
1
,...,t
s
>.
A set of reviews for a product p can be seen as
a transaction database, where each review r
i
∈ R(p)
can be seen as a transaction, and terms t
j
corresponds
to items. Given a minimum threshold for support, a
large k−itemset I = {t
1
,...,t
k
} summarizes a relevant
number of reviews in R(p). Consequently, R(p) can
be represented (and summarized) by means of the set
of large itemsets IS(R(p)) extracted from within re-
views, and composed of terms in reviews.
Moving from the general idea described above, we
now define the way to decline the concept for our con-
text.
Definition 1: Consider the set of terms T (R(p)) = {t}
appearing in reviews r ∈ R(p) and the set SW of stop-
words. Consider also a minimum support threshold
for single terms s
t
∈ (0,1).
The set T (R(p)) ⊆ T (R(p)) is the set of Relevant
DATA2012-InternationalConferenceonDataTechnologiesandApplications
224
Terms, such that ∀t ∈ T (R(p)), it is s(t) ≥ s
t
, s(t) ×
|R(p)| ≥ 3 and t 6∈ SW .
In other words, we consider relevant a term if it
is not a stopword and it is reasonably frequent in the
reviews. Notice that the set SW of stopwards is prior
knowledge for our approach.
Notice that condition s(t)×|R(p)| ≥ 3 discards terms
that are present in less than 3 reviews, even though
their support is greater than s
t
; in fact, terms that ap-
pear in only one single review are sporadic and not
sufficiently mediated by the community to character-
ize the product (typical situation with a small number
of reviews for a product).
Definition 2: Consider the set
T (R(p)) of relevant
words, and a minimum support threshold for com-
pound itemsets s
c
.
A k−itemset I with length k ≥ 2 is said relevant if
I ∩ T (R(p)) = I and s(I) ≥ s
c
.
The set of relevant itemsets for product reviews R(p)
is denoted as IS(R(p)).
In other words, we focus on itemsets composed
only of relevant words whose support is greater than
a minimum support threshold s
c
that is possibly dif-
ferent w.r.t. the minimum support threshold s
t
used
for determining relevant terms.
The reason why we consider two different min-
imum support thresholds for single terms and com-
pound itemsets is the following. Terms must be rele-
vant alone, thus they must be sufficiently representa-
tive of community opinions. But compound itemsets
may be rarer than single terms in reviews, so a differ-
ent minimum support threshold for compound item-
sets is considered.
1
Therefore, we can state that given two minimum
support thresholds s
t
and s
c
for, respectively, single
terms and compound itemsets, a set R(p) of product
reviews for a product p is summarized by the set of
single relevant terms T(R(p)) an by the set of relevant
compound itemsets IS(R(p)).
4 RETRIEVAL MODEL
The key element of our proposal is the Retrieval
Model, since it is used to rank products in order to let
most significant products to emerge, i.e., the products
whose reviews better match the query terms.
Consider a query q = {t
1
,...,t
n
} containing a
number of terms n ≥ 1. The query q itself is a
1
In principle, we could think about different minimum
thresholds for compound k−itemsets that somehow depend
on itemset length k. We will consider this aspect in our
future work.
n−itemset.
With I
l
(q) we denote a l−itemset composed by l
terms in q (it is 1 ≤ l ≤ n). We denote with D
q
the
set of all itemsets that can be obtained with terms in
q, and with D
q
(l) the subset of itemsets of length l;
notice that cardinality of D
q
(l) is |D
q
(l)| = (
n
l
) and
notice that |D| = 2
n
− 1, i.e., D is the power set of q
without the empty itemset.
In order to define a rank for a product based on
its reviews, we define the concept of weight for an
itemset.
Definition 3: The weight of a l−itemset is denoted
as w
q
(l). The weight of the single n−itemset q is, by
definition, w
q
(n) = 1, while for 1 ≤ l < n is w
q
(l) =
w
q
(l + 1)/(
n
l
).
The rationale behind Definition 3 is the follow-
ing. The topmost itemset, corresponding to the whole
query, is the most important one, and it gives the full
contribution to rank the product. In contrast, lower
levels must contribute a little, but not as much as the
topmost itemset; in particular, each non-topmost level
contributes as one of the itemset in the upper level.
Notice, that this way, the overall weight of item-
sets of level (n−1) is 1, exactly as the single topmost
itemset; reducing the size of itemsets, the contribution
of each level quickly decreases.
Example 1: To illustrate, consider the following ex-
ample query over a movie reviews collection: great
funny hilarious jokes. Based on the previous defi-
nitions, the weights for itemsets are determined. In
particular, the 4-itemset that coincides with the whole
query is assigned weight 1. For each of the four 3-
itemsets, the assigned weight is w
q
(3) = w
q
(4)/(
4
3
) =
1/4 = 0.25. For each 2-itemset, the weight is w
q
(2) =
w
q
(3)/(
4
2
) = 0.25/6 = 0.042; the weight at this level
dramatically decreases, so that the presence of only
two terms together in the review gives a very little
contribution to the overall product ranking.
Finally, 1-itemsets are assigned with weight w
q
(1) =
w
q
(2)/(
4
1
) = 0.042/4 = 0.01. Notice that even though
single terms are less than 2-itemsets, their contribu-
tion to the overall ranking is very low, in order to
compensate the possible very high supports of single
terms.
Figure 1 graphically illustrates the itemsets levels.
We are now ready to define the Product Relevance
Value (PRV).
Definition 4: Consider a product p, the set of re-
views R(p) and the set of relevant terms and com-
pound itemsets F(p) =
T (R(p))∪IS(p) that it is pos-
sible to extract from R(p). Consider now a query q
and the set of itemsets D
q
included in q.
TowardaProductSearchEnginebasedonUserReviews
225
# weight l itemsets
1 1.0000 4 {funny,great,hilarious,jokes}
4 0.2500 3 {funny,great,hilarious} {funny,great,jokes} {funny,hilarious, jokes} {great,hilarious, jokes}
6 0.0417 2 {funny,great} {funny,hilarious} {funny,jokes} {great,hilarious} {great,jokes} {hilarious, jokes}
4 0.0104 1 {funny} {great} {hilarious} {jokes}
Figure 1: Itemsets levels for query great funny hilarious jokes and corresponding weights.
If we denote with D
q
(p) = F(p) ∩ D
q
the set of
itemsets included in q that can be actually extracted
from within reviews of product p, the Product Rele-
vance Value (PRV) for product p is defined as
PRV
q
(p) =
∑
I∈D
q
(p)
(w
q
(|I|) × s(I))
The rationale of the above definition is the follow-
ing. For each itemset included in the query q and ac-
tually relevant in the reviews, its contribution to the
overall relevance value is its support value multiplied
by its weight. Notice that the longer the matched
itemset, the stronger its contribution; the shorter the
matched itemset, the smaller its contribution.
However, the definition of PRV does not take the
number of reviews for each product into account. For
this reason, we introduce the Adaptive Product Rele-
vance Value.
Definition 5: Consider a product p, the set of reviews
R(p), the query q and the Product Relevance Value
PRV
q
(p). The Adaptive Product Relevance Value
APRV is defined as
APRV
q
(p) =PRV
q
(p)×log
10
(|R(p)|)
The rationale behind this variation of the relevance
measure is that the same itemset I with a given sup-
port s(I) is more effective in summarizing a product
with a large number of reviews than a product with
a very small number of reviews. The logarithm al-
lows us to avoid that products with very large sets of
reviews excessively domintae the other ones.
Consequently, based on this concepts, the goal of
the retrieval task is to provide a ranked list of prod-
ucts, based either on the Product Relevance Value or
on the Adaptive Product Relevance Value.
5 EVALUATION
We performed a preliminary evaluation of the pro-
posed retrieval method. Our approach has been to de-
sign and then populate a database on our servers with
data derived from user-reviews taken from the inter-
net. Then we built a search engine that, performing a
query on the database, retrieves the information about
the items involved in the query and builds on them the
itemsets to apply our model. Below we describe the
source data set, discuss the query results based either
on PRV or on APRV, observe the effects of different
settings for minimum support thresholds.
5.1 Source Data Set
Products are described by a set of user-reviews
grabbed from the internet. The web site source is
epinion.com, from which we took only products
with at least five reviews. In particular, in this experi-
ment we considered user reviews concerning movies.
Our dataset is composed by a corpus of 324 prod-
ucts (movies). Each product has a number of reviews
that varies from 5 to 509. The total number of an-
alyzed reviews is 9701, and the average number of
reviews per movie is 30.
While collecting terms, we do not matter about
lower or upper case, and we do not consider puntu-
action or stop-words. We have a dictionary with a list
of about 600 distinct stop-words. Moreover, for each
product, we collect only terms with support greater
than or equal to threshold s
t
= 0.1. The support of
terms varies from 0.1 to 1, with 0.33 as average value.
As a result, we collected 72802 distinct frequent
terms (and a total of 709845 occurrences). The num-
ber of words in the set of review for a product p varies
from a minimum of 61 words to a maximum of 36200;
the average number of terms per product is 441.
From the point of view of the single-review, we
have to say that the average length of each review is
73 terms; the number of terms per reviews varies from
1 to 441.
5.2 Query Results
We submitted the query: ”What is a movie with great
funny hilarious jokes?”; in this query, there are just
4 query terms to be searched (”great, funny, hilar-
ious, jokes”), because the others are considered as
stop-words.
We repeated the query processing with different
minimum support thresholds s
c
: 0.0, 0.1, 0.2. The
results of the experiments are reported in Tables 1 and
Table 2.
Both Table 1.a and Table 1.b show results with
s
c
= 0.1. For each product p we report the number
of reviews (column R), the number of query terms hit
DATA2012-InternationalConferenceonDataTechnologiesandApplications
226
Table 1: Search result with MinSup: 0.1 according PRV ranking (a), and APRV ranking (b).
## Product R H IS PVR APVR
1 The Other Guys 17 4 15 1.0043 2.8454
2 Funny People 12 4 15 0.6016 1.4948
3 40-Year-Old Virgin 25 4 15 0.5179 1.6671
4 Paul 11 4 15 0.4830 1.1581
5 Eddie Murphy - Raw 10 4 13 0.2708 0.6236
6 Bridesmaids 11 4 11 0.2348 0.5631
7 Observe and Report 5 3 7 0.1792 0.2884
8 The Hangover 53 4 12 0.1773 0.7039
9 Hot Tub Time Machine 14 4 11 0.1734 0.4575
10 Get Him to the Greek 6 3 7 0.1684 0.3017
11 Little Fockers 9 3 7 0.1644 0.3611
12 Dinner for Schmucks 12 4 10 0.1432 0.3559
13 Bruno 16 4 10 0.1374 0.3809
14 Extract 5 3 7 0.1312 0.2112
15 Zombieland 22 4 9 0.1297 0.4010
16 The Proposal 28 4 11 0.1176 0.3917
17 Jennifer’s Body 8 3 7 0.1068 0.2220
18 I Love You, Man 11 3 7 0.1032 0.2475
19 Surviving Christmas 16 4 9 0.0996 0.2762
20 Kick-Ass 15 4 10 0.0875 0.2370
## Product R H IS PVR APVR
1 The Other Guys 17 4 15 1.0043 2.8454
2 40-Year-Old Virgin 25 4 15 0.5179 1.6671
3 Funny People 12 4 15 0.6016 1.4948
4 Paul 11 4 15 0.4830 1.1581
5 The Hangover 53 4 12 0.1773 0.7039
6 Eddie Murphy - Raw 10 4 13 0.2708 0.6236
7 Bridesmaids 11 4 11 0.2348 0.5631
8 Hot Tub Time Machine 14 4 11 0.1734 0.4575
9 Zombieland 22 4 9 0.1297 0.4010
10 The Proposal 28 4 11 0.1176 0.3917
11 Bruno 16 4 10 0.1374 0.3809
12 Little Fockers 9 3 7 0.1644 0.3611
13 Anger Management 61 4 10 0.0874 0.3594
14 Dinner for Schmucks 12 4 10 0.1432 0.3559
15 Get Him to the Greek 6 3 7 0.1684 0.3017
16 When Harry met Sally 58 3 7 0.0727 0.2953
17 Observe and Report 5 3 7 0.1792 0.2884
18 Toy Story 3 28 4 8 0.0863 0.2876
19 Surviving Christmas 16 4 9 0.0996 0.2762
20 I Love You, Man 11 3 7 0.1032 0.2475
a) b)
Table 2: Search result according APRV ranking with MinSup=0.00 (a), and with MinSup=0.20 (b).
## Product R H IS PVR APVR
1 The Other Guys 17 4 15 1.0043 2.8454
2 40-Year-Old Virgin 25 4 15 0.5179 1.6671
3 Funny People 12 4 15 0.6016 1.4948
4 Paul 11 4 15 0.4830 1.1581
5 The Hangover 53 4 15 0.2716 1.0784
6 Eddie Murphy - Raw 10 4 15 0.3958 0.9114
7 Bridesmaids 11 4 15 0.3750 0.8992
8 Anger Management 61 4 15 0.1735 0.7132
9 Bruno 16 4 15 0.2493 0.6913
10 The Proposal 28 4 15 0.1726 0.5752
11 Toy Story 3 28 4 15 0.1652 0.5504
12 Zombieland 22 4 13 0.1695 0.5240
13 Scott Pilgrim vs. . . . 20 4 15 0.1672 0.5008
14 Dinner for Schmucks 12 4 13 0.1884 0.4681
15 Hot Tub Time Machine 14 4 11 0.1734 0.4575
16 Aladdin 79 4 15 0.0843 0.3682
17 Little Fockers 9 3 7 0.1644 0.3611
18 Kick-Ass 15 4 13 0.1236 0.3347
19 Surviving Christmas 16 4 11 0.1178 0.3267
20 Get Him to the Greek 6 3 7 0.1684 0.3017
## Product R H IS PVR APVR
1 The Other Guys 17 4 15 1.0043 2.8454
2 40-Year-Old Virgin 25 4 12 0.2679 0.8624
3 Funny People 12 4 11 0.3446 0.8563
4 The Hangover 53 4 8 0.1309 0.5197
5 Bridesmaids 11 4 9 0.1818 0.4360
6 Zombieland 22 4 8 0.1241 0.3835
7 Little Fockers 9 3 7 0.1644 0.3611
8 Paul 11 4 9 0.1496 0.3588
9 Eddie Murphy - Raw 10 4 9 0.1542 0.3550
10 Hot Tub Time Machine 14 4 9 0.1317 0.3476
11 Bruno 16 4 8 0.1243 0.3448
12 Dinner for Schmucks 12 4 8 0.1293 0.3214
13 Get Him to the Greek 6 3 7 0.1684 0.3017
14 Observe and Report 5 3 7 0.1792 0.2884
15 Jennifer’s Body 8 3 7 0.1068 0.2220
16 National Lampoon’s. . . 83 3 5 0.0383 0.1691
17 The Proposal 28 4 6 0.0491 0.1636
18 Anger Management 61 4 6 0.0396 0.1629
19 Toy Story 3 28 4 6 0.0432 0.1438
20 I Love You, Man 11 3 6 0.0578 0.1385
a) b)
(column H), the number of mined itemsets (column
IS), and the relevance values PRV and APRV.
In particular, Table 1.a lists the best 20 movies in
reverse order of PRV. In contrast, Table 1.b lists the
best 20 product in reverse order of APRV, that are
not exactly the same movies. Movie in position 11
in Table 1.a is now in position 12: in effect, it has
a small number of reviews, w.r.t. the other movies,
and it looses a position in the ranked list. For this rea-
son, we decide to adopt the APRV as standard ranking
measure in the next experiments.
In order to evaluate the effect of the minimum sup-
port threshold value s
c
, we performed other two ex-
periments with s
c
= 0.0, whose results are reported in
Table 2.a, and s
c
= 0.2, whose results are reported in
Table 2.b.
In particular, setting s
c
= 0.0 means that every
itemset is considered, even with support lower than
single term minimum support threshold s
t
. The effect
is that a large number of lowly representative itemsets
are considered, and affect the results. Looking at Ta-
ble 2.a, movie Little Fockers drops to position 17 from
position 12. But looking at reviews, this movie is a
good sample for the query, thus this setting is counter-
TowardaProductSearchEnginebasedonUserReviews
227
effective.
Considering the setting with s
c
= 0.2 (Table 2.b),
the higher minimum support threshold obtains, as an
effect, that only strongly representative itemsets actu-
ally contribute to the value of APRV. Consequently,
movie Little Fockers now occupies position 7, thus
significantly improving its position.
Of course, these are preliminary results and a
deeper study must be performed, evaluating precision
and recall of the different settings.
6 CONCLUSIONS
In this paper we presented a novel retrieval model for
a product search engine based on user-reviews taken
from the internet.
The two basic ideas that reside under the search
engine are (1) itemset computation at query time
from frequent terms in user-reviews and (2) a ranking
model that permits to weight these itemsets. The work
is at an early stage, but based on our experiments we
can say that the whole idea seems to be quite promis-
ing.
Future Works. Our retrieval model is very context-
dependent so we have to develop automatic criteria
to enrich a generic stop-words list with those terms
that are too much context frequent. A too much con-
text frequent term has no relevant semantic value, but
nonetheless it can affect ranking in a distorted way.
For examples, in the case we showed in this paper the
terms movie and movies are considered stop-words.
Details apart, the next step is to improve our rank-
ing model including the concept of term-closeness.
At the moment while matching a query in a user-
review we do not consider term position inside the
review. We do believe that matching closer terms can
lead closer to the real meaning of the query. So we
want to develop an index based on term-closeness to
affect our ranking model.
A further step is to hook our search engine to a
dictionary or an ontology, like Wordnet, in order to
better characterize words.
REFERENCES
Agrawal, R., Imielinski, T., and Swami, A. (1993).
Database mining: A performance perspective. IEEE
Transactions on Knowledge and Data Engineering,
5(6):914–925.
Agrawal, R. and Srikant, R. (1994). Fast algorithms for
mining association rules in large databases. In Pro-
ceedings of the 20th VLDB Conference, Santiago,
Chile.
Chaovalit, P. and Zhou, L. (2005). Movie review mining:
a comparison between supervised and unsupervised
classification approaches. In HICSS. IEEE Computer
Society.
Dave, K., Lawrence, S., and Pennock, D. M. (2003). Min-
ing the peanut gallery: opinion extraction and seman-
tic classification of product reviews. In WWW, pages
519–528.
Hu, M. and Liu, B. (2004a). Mining and summarizing cus-
tomer reviews. In Kim, W., Kohavi, R., Gehrke, J.,
and DuMouchel, W., editors, KDD, pages 168–177.
ACM.
Hu, M. and Liu, B. (2004b). Mining opinion features in
customer reviews. In McGuinness, D. L. and Fergu-
son, G., editors, AAAI, pages 755–760. AAAI Press /
The MIT Press.
Kim, W., Ryu, J., Kim, K. I., and Kim, U.-M. (2009). A
method for opinion mining of product reviews using
association rules. In Sohn, S., Chen, L., Hwang, S.,
Cho, K., Kawata, S., Um, K., Ko, F. I. S., Kwack, K.-
D., Lee, J. H., Kou, G., Nakamura, K., Fong, A. C. M.,
and Ma, P. C. M., editors, Int. Conf. Interaction Sci-
ences, volume 403 of ACM International Conference
Proceeding Series, pages 270–274. ACM.
Lee, D., Jeong, O.-R., and goo Lee, S. (2008). Opinion
mining of customer feedback data on the web. In Kim,
W. and Choi, H.-J., editors, ICUIMC, pages 230–235.
ACM.
Liu, B., Hu, M., and Cheng, J. (2005). Opinion observer:
analyzing and comparing opinions on the web. In El-
lis, A. and Hagino, T., editors, WWW, pages 342–351.
ACM.
Liu, D. T. and Xu, X. W. (2001). A review of web-based
product data manageement systems. 44(1):251–262.
Ly, D. K., Sugiyama, K., Lin, Z., and Kan, M.-Y. (2011).
Product review summarization based on facet identifi-
cation and sentence clustering. CoRR, abs/1110.1428.
Sandvig, J. J., Mobasher, B., and Burke, R. D. (2007). Ro-
bustness of collaborative recommendation based on
association rule mining. In Konstan, J. A., Riedl, J.,
and Smyth, B., editors, RecSys, pages 105–112. ACM.
Zhuang, L., Jing, F., and Zhu, X. (2006). Movie review
mining and summarization. In Yu, P. S., Tsotras, V. J.,
Fox, E. A., and Liu, B., editors, CIKM, pages 43–50.
ACM.
DATA2012-InternationalConferenceonDataTechnologiesandApplications
228
