Incorporating Guest Preferences into Collaborative Filtering for Hotel

Recommendation

Fumiyo Fukumoto

, Hiroki Sugiyama

, Yoshimi Suzuki

and Suguru Matsuyoshi

Interdisciplinary Graduate School of Medicine and Engineering, Univ. of Yamanashi, Kofu, Japan

Faculty of Engineering, Univ. of Yamanashi, Kofu, Japan

Keywords:

Collaborative Filtering, Latent Dirichlet Allocation, Recommendation System, Transition Probability.

Abstract:

Collaborative ﬁltering (CF) has been widely used as a ﬁltering technique because it is not necessary to apply

more complicated content analysis. However, it is difﬁcult to take users’ preferences/criteria related to the

aspects of a product/hotel into account. This paper presents a method of hotel recommendation that incorpo-

rates different aspects of a product/hotel to improve quality of the score. We used the results of aspect-based

sentiment analysis for guest preferences. The empirical evaluation using Rakuten Japanese travel data showed

that aspect-based sentiment analysis improves overall performance. Moreover, we found that it is effective for

ﬁnding hotels that have never been stayed at but share the same neighborhoods.

1 INTRODUCTION

Collaborative ﬁltering (CF) identiﬁes the potential

preference of a consumer/guest for a new prod-

uct/hotel by using only the information collected from

other consumers/guests with similar products/hotels

in the database. It is a simple technique as it is not

necessary to apply more complicated content analy-

sis compared to the content-based ﬁltering framework

(Balabanovic and Shoham, 1997). CF has been very

successful in both research and practical systems as it

has been widely studied (Huang et al., 2004; Yildirim

and Krishnamoorthy, 2008; Liu and Yang, 2008; Li

et al., 2009; Lathia et al., 2010; Zhao et al., 2013),

and many practical systems such as Amazon for book

recommendation and Expedia for hotel recommenda-

tion have been developed.

Item-based collaborativeﬁltering is one of the ma-

jor recommendation algorithms (Sarwar et al., 2001;

Zhao et al., 2013) because of its simplicity. The algo-

rithms assume that the consumers/guests are likely to

prefer product/hotel that are similar to what they have

bought/stayed before. Unfortunately, most of them

only consider star ratings and leave consumers/guests

textual reviews. Several authors focused on the prob-

lem, and attempted to improve recommendation re-

sults by using the techniques on text analysis such

as sentiment analysis, opinion mining, or informa-

tion extraction (Cane et al., 2006; Niklas et al., 2009;

Raghavan et al., 2012). However, major approaches

aim at ﬁnding the positive/negative opinions for the

product/hotel, and do not take users preferences re-

lated to the aspects of a product/hotel into account.

For instance, one guest is interested in a nice restau-

rant for selecting hotels for her/his vacation, while an-

other guest, e.g., a businessman prefers to the hotel

which is close to the station. In this case, the aspect

of the former is different from that of the latter.

This paper presents a collaborative ﬁltering

method for hotel recommendationincorporating guest

preferences. We rank hotels according to scores. The

score is obtained by using the analysis of different

aspects of guest preferences. The method utilizes a

large amount of guest reviews which make it possi-

ble to solve the item-based ﬁltering problem of data

sparseness, i.e., some items were not assigned a label

of users preferences. We used the results of aspect-

based sentiment analysis to recommend hotels be-

cause whether or not the hotel can be recommended

depends on the guest preferences related to the as-

pects of a hotel. For instance, if one guest stays at

hotels for her/his vacation, a room with nice views

may be an important factor to select hotels, whereas

another guest who stays at hotels for business, hotels

with close to the station may be selected.

We parsed all reviews by using syntactic analyzer,

and extracted dependency triples which represent the

relationship between aspect and its preference. For

each aspect of a hotel, we identiﬁed the guest opin-

ion, e.g., the aspect, service is good or not, based

Fukumoto F., Sugiyama H., Suzuki Y. and Matsuyoshi S..

Incorporating Guest Preferences into Collaborative Filtering for Hotel Recommendation.

DOI: 10.5220/0005034000220030

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 22-30

ISBN: 978-989-758-048-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

on the dependency triples in the guest reviews. The

positive/negative opinion on some aspect is used to

calculate transitive association between hotels. Fi-

nally, we scored hotels by Markov Random Walk

(MRW) model, i.e., we used MRW based recommen-

dation technique to explore transitive associations be-

tween the hotels. Random Walk based recommenda-

tion overcomes the item-based CF problem that the

inability to explore transitive associations between the

hotels that have never been stayed but share the same

neighborhoods (Li et al., 2009).

2 RELATED WORK

CF mainly consists of two procedures, prediction and

recommendation (Sarwar et al., 2001). Here, predic-

tion refers to a numerical value expressing the pre-

dicted likeliness of item for user, and recommenda-

tion is a list of items that the user will like the most.

As the volume of online reviews has drastically in-

creased, sentiment analysis, opinion mining, and in-

formation extraction for the process of prediction are

a practical problem attracting more and more atten-

tion. Several efforts have been made to utilize these

techniques to recommend products (Niklas et al.,

2009; Faridani, 2011). Cane et. al have attempted

to elicit user preferences expressed in textual reviews,

and map such preferences onto some rating scales that

can be understood by existing CF algorithms (Cane

et al., 2006). They identiﬁed sentiment orientations of

opinions by using a relative-frequency-based method

that estimates the strength of a word with respect to a

certain sentiment class as the relative frequency of its

occurrence in the class. The results using movie re-

views from the Internet Movie Database (IMDb) for

the MovieLens 100k dataset showed the effectiveness

of the method, while the sentiment analysis they used

is limited, i.e., they used only adjectives or verbs.

Niklas et. al have attempted to improve the accu-

racy of movie recommendations by using the results

of opinion extraction from free-text reviews (Niklas

et al., 2009). They presented three approaches: (i)

manual clustering, (ii) semi-automatic clustering by

Explicit Semantic Analysis (ESA), and (iii) fully

automatic clustering by Latent Dirichlet Allocation

(LDA) (Blei et al., 2003) to extract movie aspects

as opinion targets, and used them as features for the

collaborative ﬁltering. The results using 100 random

users from the IMDb showed that the LDA-based

movie aspect extraction yields the best results. Our

work is similar to Niklas et. al method in the use

of LDA. The difference is that our approach applied

LDA to the dependency triples while Niklas applied

LDA to single words. Raghavan et. al have attempted

to improve the performance of collaborative ﬁlter-

ing in recommender systems by incorporating quality

scores to ratings (Raghavan et al., 2012). The quality

scores are used to decide the importance given to the

individual rating. Their method using quality scores

consists of two steps. In the ﬁrst step, the quality

scores of ratings using the review and user data set

are estimated. The second step involves rating by us-

ing the quality scores as weights. They adapted the

probabilistic matrix factorization (PMF) framework.

The PMF aims at inferring latent factors of users and

items from the available ratings. The experimental

evaluation on two product categories of a benchmark

data set, i.e., Book and Audio CDs from Amazon.com

showed the efﬁcacy of the method.

In the context of recommendation, several au-

thors have attempted to rank items by using graph-

based ranking algorithms (Yin et al., 2010; L.Li et al.,

2014). Wijaya et. al have attempted to rank items di-

rectly from the text of their reviews (Wijaya and Bres-

san, 2008). They constructed a sentiment graph by

using simple contextual relationships such as colloca-

tion, negative collocation and coordination by pivot

words such as conjunctions and adverbs. They ap-

plied PageRank algorithm to the graph to rank items.

They reported that the results using 50 movies ran-

domly selected from box ofﬁce list of Nov. 2007 to

Feb. 2008 showed effectiveness of their method while

combination of positive and negative orientation did

not work well in a linear fashion.

Li et. al proposed a basket-sensitive random walk

model for personalized recommendation in the gro-

cery shopping domain (Li et al., 2009). The method

extends the basic random walk model by calculating

the product similarities through a weighted bi-partite

network which allows the current shopping behaviors

to inﬂuence the product ranking. Empirical results us-

ing three real-world data sets, LeShop, TaFeng and

an anonymous Belgium retailer showed that a perfor-

mance improvement of the method over other existing

collaborative ﬁltering models, the cosine, conditional

probability and the bi-partite network based similar-

ities. However, the transition probability from one

product node to another product node is computed

based on a user’s purchase frequency of a product

with regardless of the users’ positive or negative opin-

ions concerning to the product.

There are three novel aspects in our method.

Firstly, we propose a method to incorporate differ-

ent aspect of a hotel into users preferences/criteria

to improve quality of recommendation. Secondly,

from a ranking perspective, the MRW model we used

is calculated based on the polarities of reviews. Fi-

IncorporatingGuestPreferencesintoCollaborativeFilteringforHotelRecommendation

1. Aspect analysis

2. Pos/Neg opinion detection

3. Pos/Neg review

identification

4. Scoring hotels by

MRW model

Reviews

Aspects

䞉䞉䞉䞉

SVM

Training reviews

Test reviews

Location

pos review

Location

neg review

1. 䞉䞉䞉䞉

2. 䞉䞉䞉䞉

3. 䞉䞉䞉䞉

Overall

Pos/Neg

Room

Pos/Neg

Location

Pos/Neg

….

Figure 1: Overview of the method.

nally, from the opinion mining perspective, we pro-

pose overcoming with the unknown polarized words

by utilizing LDA.

3 SYSTEM DESIGN

Figure 1 illustrates an overview of the method. It

consists of four steps: (1) Aspect analysis, (2) Posi-

tive/negative opinion detection based on aspect anal-

ysis, (3) Positive/negative review identiﬁcation, and

(4) Scoring hotels by MRW model.

3.1 Aspect Analysis

The ﬁrst step to recommend hotels based on guest

preferences is to extract aspects for each hotel from

a guest review corpus. All reviews were parsed by the

syntactic analyzer CaboCha (Kudo and Matsumoto,

2003), and all the dependency triples (rel, x, y) are

extracted. Here, x refers to a noun/compound noun

word related to the aspect. y shows verb or adjec-

tive word related to the preference for the aspect. rel

denotes a grammatical relationship between x and y.

We classiﬁed rel into 9 types of Japanese particle,

“ga(ha)”, “wo”, “ni”, “he”, “to”, “de”, “yori”, “kara”

and “made”. For instance, from the sentence “Cy-

ousyoku (breakfast) ga totemo (very) yokatta (good).”

(The breakfast was very good.), we can obtain the de-

pendency triplet, (ga, cyousyoku, yokatta). The triplet

represents positive opinion, “yokatta”(good) concern-

ing to the aspect, “Cyousyoku”(meal).

3.2 Positive/Negative Opinion Detection

We identiﬁed positive/negative opinion based on the

aspects of a hotel. We classiﬁed aspects into seven

types: “Location”, “Room”, “Meal”, “Spa”, “Ser-

vice”, “Amenity”, and “Overall”. These types are de-

ﬁned by Rakuten travel

. For each aspect, we identi-

ﬁed positive/negative opinion by using Japanese sen-

timent polarity dictionary (Kobayashi et al., 2005),

i.e., we regarded the extracted dependency triplet as

positive/negativeopinion if y in the triplet (rel, x, y) is

classiﬁed into positive/negative classes in the dictio-

nary. However, the dictionary makes it nearly impos-

sible to cover all of the words in the review corpus.

For unknown verb or adjective words that were

extracted from the review corpus, but did not ap-

pear in any of the dictionary classes, we classiﬁed

them into positive or negative class by using a topic

model. Topic models such as probabilistic latent se-

mantic indexing (Hofmann, 1999) and Latent Dirich-

let Allocation (LDA) (Blei et al., 2003) are based

on the idea that documents are mixtures of topics,

where each topic is captured by a distribution over

words. The topic probabilities provide an explicit

low-dimensional representation of a document. They

have been successfully used in many domains such

as text modeling and collaborative ﬁltering (Li et al.,

2013). We used LDA and classiﬁed unknown words

into positive/negative classes. LDA presented by

(Blei et al., 2003) models each document as a mixture

of topics, and generates a discrete probability distri-

bution over words for each topic. The generative pro-

cess for LDA can be described as follows:

1. For each topic k = 1, ···, K, generate φ

, multino-

mial distribution of words speciﬁc to the topic k

from a Dirichlet distribution with parameter β;

2. For each document d = 1, ···, D, generate θ

multinomial distribution of topics speciﬁc to the

document d from a Dirichlet distribution with pa-

rameter α;

3. For each word n = 1, ···, N

in document d;

(a) Generate a topic z

of the n

word in the doc-

ument d from the multinomial distribution θ

(b) Generate a word w

, the word associated with

the n

word in document d from multinomial

zdn

Like much previous work on LDA, we used Gibbs

sampling to estimate φ and θ. The sampling probabil-

ity for topic z

in document d is given by:

P(z

| z

,W) =

\i, j

+ β)(n

\i, j

+ α)

\i, j

+Wβ)(n

\i,·

+ Tα)

. (1)

refers to a topic set Z, not including the current

assignment z

. n

\i, j

is the count of word v in topic j

that does not include the current assignment z

, and

http://rit.rakuten.co.jp/rdr/index.html

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

Topic_id1

Topic_id2

Topic_id3

(ga, heya, yoi)

room was nice

(ha, heya, good-da)

room was nice

(ga, heya, insyouteki)

room was impressive

(ha, heya, kakubetu-da)

room was very nice

(ha, heya, semai)

room was small

(ga, heya, niou)

room smells of a cigarette

positive

negative

…..

………….

……….

………..

cluster

Figure 2: Clusters obtained by LDA.

\i, j

indicates a summation over that dimension. W

refers to a set of documents, and T denotes the total

number of unique topics. After a sufﬁcient number

of sampling iterations, the approximated posterior can

be used to estimate φ and θ by examining the counts

of word assignments to topics and topic occurrences

in documents. The approximated probability of topic

k in the document d,

, and the assignments word w

to topic k,

are given by:

+ α

+ αK

. (2)

+ β

+ βV

. (3)

For each aspect, we manually collected reviews

and created a review set. We applied LDA to each

set of reviews consisted of triples. We need to esti-

mate the number of topics k for the result obtained

by LDA. Figure 2 illustrates the result obtained by

LDA. The aspect type is “room”. The triplet marked

with box includes unknown adjective words. We can

see from Figure 2 that the result can be regarded as a

clustering result: each element of the clusters is pos-

itive/negative opinion according to the sentiment po-

larity dictionary, or unknown words. We estimated

the number of topics (clusters) k by using Entropy

measure given by:

E = −

logk

∑

P(A

)logP(A

).(4)

k refers to the number of clusters. P(A

) is a prob-

ability that the elements of the cluster C

assigned to

the correct class A

. N denotes the total number of

elements and N

shows the total number of elements

assigned to the clusterC

. The value of E ranges from

0 to 1, and the smaller value of E indicates better re-

sult. We chose the parameter k whose value of E is

smallest. For each cluster, if the number of positive

opinion is larger than those of negative ones, we re-

garded a triplet including unknownword in the cluster

as positive and vice versa. For example, “yoi”(nice)

in the Topic

id1 cluster shown in Figure 2 is regarded

to a positive as the number of positive and negative

were one and zero, respectively.

3.3 Positive/Negative Review

Identiﬁcation

We used the result of positive/negative opinion detec-

tion to classify guest reviews into positive or nega-

tive related to the aspect. Like much previous work

on sentiment analysis based on supervised machine

learning techniques (Turney, 2002) or corpus-based

statistics, we used Support Vector Machine (SVMs)

to annotate automatically (Joachims, 1998). For

each aspect, we collected positive/negative opinion

(triples) from the results of LDA

. Each review in the

test data is represented as a vector where each dimen-

sion of a vector is positive/negative triplet appeared

in the review, and the value of each dimension is a

frequency count of the triplet. For each aspect, the

classiﬁcation of each review can be regarded as a two-

class problem: positive or negative.

3.4 Scoring Hotels by MRW Model

The ﬁnal procedure for recommendation is to rank

each hotel. We used a ranking algorithm, the MRW

model that has been successfully used in Web-link

analysis, social networks (Xue et al., 2005), and rec-

ommendation (Li et al., 2009; Yin et al., 2010; L.Li

et al., 2014). We applied the algorithm to rank hotels.

Given a set of hotels H, Gr = (H, E) is a graph re-

ﬂecting the relationships between hotels in the set. H

is the set of nodes, and each node h

in H refers to the

hotel. E is a set of edges, which is a subset of H ×

H. Each edge e

in E is associated with an afﬁnity

weight f(i → j) between hotels h

and h

(i 6= j). The

weight of each edge is a value of transition probability

P(h

| h

) between h

and h

, and deﬁned by:

P(h

| h

) =

|Gr|

∑

k=1

c(g

)

(

∑

c(g

,·))

c(g

)

(

∑

c(·,h

))

. (5)

Eq. (5) shows the preference voting for target hotel h

from all the guests in Gr who stayed at h

. We note

that we classiﬁed reviews into positive/negative. We

used the results to improve the quality of score. More

We used the clusters that the number of positive and

negative words is not equal.

IncorporatingGuestPreferencesintoCollaborativeFilteringforHotelRecommendation

precisely, we used positive review counts to calcu-

late transition probability, i.e., c(g

) and c(g

)

in Eq. (5) refer to the lodging count that the guest g

reviewed the hotel h

) as positive. P(h

| h

) in Eq.

(5) is the marginal probability distribution over all the

guests. The transition probability obtained by Eq. (5)

shows a weight assigned to the edge between hotels

and h

We used the row-normalized matrix U

)

|H|×|H|

to describe Gr with each entry corre-

sponding to the transition probability, where U

p(h

| h

). To make U a stochastic matrix, the rows

with all zero elements are replaced by a smoothing

vector with all elements set to

|H|

. The matrix form

of the recommendation score Score(h

) can be formu-

lated in a recursive form as in the MRW model:

λ =

µU

λ+

(1−µ)

|H|

~e, where

λ = [Score(h

)]

|H|×1

is a vector

of saliency scores for the hotels. ~e is a column vector

with all elements equal to 1. µ is a damping factor.

We set µ to 0.85, as in the PageRank (Brin and Page,

1998). The ﬁnal transition matrix is given by:

M = µU

(1− µ)

| H |

~e~e

. (6)

Each score is obtained by the principal eigenvector of

the new transition matrix M. We applied the algo-

rithm to the graph. The higher score based on tran-

sition probability the hotel has, the more suitable the

hotel is recommended. For each aspect, we chose the

topmost k hotels according to rank score. For each se-

lected hotel, if the negative review is not included in

the hotel reviews, we regarded the hotel as a recom-

mendation hotel.

4 EXPERIMENTS

4.1 Data

We used Rakuten travel data

. It consists of 11,468

hotels, 348,564 reviews submitted from 157,729

guests. We used plda

to assign positive/negative tag

to the aspects. For each aspect, we estimated the num-

ber of topics (clusters) by searching in steps of 100

from 200 to 1,000. Table 1 shows the minimum en-

tropy value and the number of topics for each aspect.

As shown in Table 1, the number of topics ranges

from 500 to 700. For each of the seven aspects, we

used these number of topics in the experiments. We

http://rit.rakuten.co.jp/rdr/index.html

http://code.google.com/p/plda

Table 1: The minimum entropy value and the # of topics.

Aspect Entropy Topics

Location 0.209 700

Room 0.460 600

Meal

0.194 700

Spa 0.232 500

Service

0.226 700

Amenity 0.413 600

Overall 0.202 700

used linear kernel of SVM-Light (Joachims, 1998)

and set all parameters to their default values. All re-

views were parsed by the syntactic analyzer CaboCha

(Kudo and Matsumoto, 2003), and 633,634 depen-

dency triples are extracted. We used them in the ex-

periments.

We had an experiment to classify reviews into pos-

itive or negative. For each aspect, we chose the top-

most 300 hotels whose number of reviews are large.

We manually annotated these reviews. The evalua-

tion is made by two humans. The classiﬁcation is

determined to be correct if two human judges agree.

We obtained 400 reviews consisting 200 positive and

200 negative reviews. 400 reviews are trained by

using SVMs for each aspect, and classiﬁers are ob-

tained. We randomly selected another 100 test re-

views from the topmost 300 hotels, and used them as

test data. Each of the test data was classiﬁed into pos-

itive or negative by SVMs classiﬁers. The process is

repeated ﬁve times. As a result, the macro-averaged

F-score concerning to positive across seven aspects

was 0.922, and the F-score for negative was 0.720.

For each aspect, we added the reviews classiﬁed by

SVMs to the original 400 training reviews, and used

them as a training data to classify test reviews.

We created the data which is used to test our rec-

ommendation method. More precisely, we used the

topmost 100 guests staying at a large number of dif-

ferent hotels as recommendation. For each of the 100

guests, we sorted hotels in chronological order. We

used these with the latest ﬁve hotels as test data. To

score hotels by MRW model, we used guest data stay-

ing at more than three times. The data is shown in Ta-

ble 2. “Hotels” and “Different hotels” in Table 2 refer

to the total number of hotels, and the number of dif-

ferent hotels that the guests stayed at more than three

times, respectively. “Guests” shows the total number

of guests who stayed at one of the “Different hotels”.

“Reviews” shows the number of reviews with these

hotels.

For evaluation measure used in recommendation,

we used MAP (Mean-Averaged Precision) (Yates and

Neto, 1999). For a given set of guests G = {g

, · ·· ,

}, and H = {h

, · ·· , h

} be a set of hotels that

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

Table 2: Data used in the experiments.

Hotels 30,358

Different hotels

6,387

Guests 23,042

Reviews

116,033

Table 3: Recommendation results.

Method MAP

Trans. pro. without review 0.257

Content Words 0.304

Without reviews by SVMs 0.356

Without neg review ﬁltering 0.378

Aspect-based SA 0.392

should be recommended for a guest g

, the MAP of G

is given by:

MAP(G) =

| G |

|G|

∑

j=1

∑

k=1

Precision(R

).(7)

in Eq. (7) refers to the set of ranked retrieval

results from the top result until we get hotel h

Precision indicates a ratio of correct recommendation

hotels by the system divided by the total number of

recommendation hotels.

4.2 Recommendation Results

We compared the results obtained by our method,

aspect-based sentiment analysis (ASA) with the fol-

lowing four approaches to examine how the results of

each method affect the overall performance.

1. Transition probabilities without review (TPWoR)

The probabilityP(h

| h

) used in the method is the

preference voting for the target hotel h

from all

the guests in a set G who stayed at h

, regardless

of positive or negative review of G.

2. Content Words (CW)

The difference between content words method

and our method, ASA is that the former applies

LDA to the content words.

3. Without reviews classiﬁed by SVMs (WoR)

SVMs used in this method classiﬁes test data by

using only the original 400 training reviews.

4. Without negative review ﬁltering (WoNRF)

The method selected the topmost k hotels accord-

ing to the MRW model, and the method dose not

use negative reviews as a ﬁltering.

Table 3 shows averaged MAP across seven aspects.

As we can see from Table 3 that aspect-based senti-

ment analysis was the best among four baselines, and

.259

.267

.253

.243

.262

.264

.249

.299

.285

.311

.301

.289

.314

.329

.344

.338

.367

.349

.351

.363

.381

.367

.359

.387

.381

.369

.378

.408

.391

.373

.403

.392

.382

.391

.414

0.1

0.2

0.3

0.4

0.5

Location

Room

Meal

Spa

Service

Amenity

Overall

MAP

Transition probabilities

Content words

Without reviews classified by SVMs

Without negative review filtering

Aspect-based sentiment analysis

Figure 3: The results against each aspect.

MAP score attained at 0.392. The result obtained by

transition probability without review was worse than

any other results. This shows that the use of guest

review information is effective for recommendation.

Table 3 shows that the result obtained by content

words method was worse than the result obtained by

aspect-based sentiment analysis, and even worse than

the results without reviews classiﬁed by SVMs (WoR)

and without negative review ﬁltering (WoNRF). Fur-

thermore, we can see from Table 3 that negative re-

view ﬁltering was a small contribution, i.e., the im-

provement was 0.014 as the result without negative

review ﬁltering was 0.378 and aspect-based SA was

0.392. One reason is that the accuracy of negative re-

view identiﬁcation. The macro-averagedF-score con-

cerning to negative across seven aspects was 0.720,

while the F-score for positive was 0.922. Negative re-

view ﬁltering depends on the performance of negative

review identiﬁcation. Therefore, it will be necessary

to examine features other than word triples to improve

negative review identiﬁcation.

Table 4 shows sample clusters regarded as posi-

tive for three aspects, “location”, “room”, and “meal”

obtained by LDA. Each cluster shows the top 5 triples

and content words. We observed that the extracted

triples show positive opinion for each aspect. This in-

dicates that aspect extraction contributes to improve

overall performance. In contrast, some words such as

yoi (be good) and manzoku (satisfy) in content word

based clusters appear across aspects. Similarly, some

words such as ricchi (location) and cyousyoku (break-

fast) which appeared in negative cluster are an ob-

stacle to identify positive/negative reviews in SVMs

classiﬁcation.

It is very important to compare the results of our

method with four baselines against each aspect. Fig-

ure 3 shows MAP against each aspect. The results ob-

tained by aspect-based sentiment analysis were statis-

IncorporatingGuestPreferencesintoCollaborativeFilteringforHotelRecommendation

Table 4: Top 5 triples and content words.

Aspects Content Words

Location Room Meal Location Room Meal

(ni, eki, chikai) (ga, heya, yoi) (ga, shokuji, yoi) ricchi heya syokuji

be near to the station room was nice breakfast was nice location room meal

(ha, hotel, chikai) (ha, heya, hiroi) (ha, shokuji, yoi) eki hiroi yoi

the hotel is close the room is wide meal was nice station be wide be good

(ni, hotel, chikai) (ga, heya, kirei) (ha, restaurant, good) yoi kirei cyousyoku

be near to the hotel

A room is clean a restaurant is good be good be clean breakfast

(ni, parking, chikai) (de , sugoseru, heya) (ha, restaurant, yoi) mise manzoku oishii

be near to the can spend restaurant is nice store satisfy be delicious

parking in the room

(ga, konbini, aru) (ha, heya, jyuubun) (ha, buffet, yoi) subarashii yoi manzoku

be near to the

a room is Buffet is delicious be great be good satisfy

convenience store enough good

Table 5: Recommendation list for user ID 2037.

R TPWoR CW WoR WoNRF ASA

1 2349 2203 2614 3022 449

2 604 2349 554 604 30142

12869 30142 30142 449 18848

4 90 604 604 30142 531

5 666 12869 3022 18848 769

2149 39502 531 531 2223

7 38126 449 449 769 15204

449 31209 18848 2223 20428

tically signiﬁcant compared to other methods except

for the aspects “spa” and “overall in “without negative

review ﬁltering” method.

Table 5 shows a ranked list of the hotels for

one guest (guest ID: 2037) obtained by using each

method. The aspect is “meal”, and each number

shows hotel ID. Bold font in Table 5 refers to the

correct hotel, i.e., the latest ﬁve hotels that the guest

stayed at. As can be seen clearly from Table 5, the

result obtained by our method includes all of the ﬁve

correct hotels within the topmost eight hotels, while

without negative review ﬁltering (WoNRF) was four.

TPWoR, CW, and WoR did not work well as the num-

ber of correct hotel was no more than three, and these

were ranked seventh and eighth.

It is interesting to note that some recommended

hotels are very similar to the correct hotels, while

most of the eight hotels did not exactly match these

correct hotels except for the result obtained by aspect-

based sentiment analysis method. If these hotels were

similar to the correct hotels, the method is effective

for ﬁnding transitive associations between the hotels

that have never been stayed but share the same neigh-

borhoods. Therefore, we examined how these hotels

are similar to the correct hotels. To this end, we calcu-

lated distance between correct hotels and other hotels

Table 6: Distance between correct hotel and another hotel.

Method Dis

Trans. pro. without review 3.067

Content Words

2.859

Without reviews by SVMs 2.721

Without neg review ﬁltering

2.532

Aspect-based SA 2.396

within the rank for each method by using seven pref-

erences. The preferences have star rating, i.e., each

has been scored from 1 to 5, where 1(bad) is low-

est, and 5(good) is the best score. We represented

each ranked hotel as a vector where each dimension

of a vector is these seven preferences and the value

of each dimension is its score value. The distance be-

tween correct hotel and other hotels within the rank

for each method X is deﬁned as:

Dis(X) =

| G |

|G|

∑

i=1

argmin

j,k

d(R

,C h

). (8)

| G | refers to the number of guests. R

refers to a

vector of the j-th ranked hotels except for the correct

hotels. Similarly, C

stands for a vector represen-

tation of the k-th correct hotel. d refers to Euclidean

distance. Eq. (8) shows that for each guest, we ob-

tained the minimum value of Euclidean distance be-

tween R

and C h

. We calculated the averaged

summation of the 100 guests. The results are shown

in Table 6.

The value of “Dis” in Table 6 shows that the

smaller value indicates a better result. We can see

from Table 6 that the hotels except for the correct ho-

tels obtained by our method are more similar to the

correct hotels than those obtained by four baselines.

The results show that our method is effective for ﬁnd-

ing hotels that have never been stayed at but share the

same neighborhoods.

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

5 CONCLUSIONS

We proposed a method for recommending hotels by

incorporating different aspects of a hotel to improve

quality of score. We used the results of aspect-based

sentiment analysis for guest preferences. We parsed

all reviews by the syntactic analyzer, and extracted

dependency triples. For each aspect, we identiﬁed

the guest opinion to positive or negative, by using

dependency triples in the guest review. We calcu-

lated transitive association between hotels based on

the positive/negative opinion. Finally, we scored ho-

tels by Markov Random Walk model. The compar-

ative results using Rakuten travel data showed that

aspect analysis of guest preferences improves overall

performance, and it is effective for ﬁnding hotels that

have never been stayed at but share the same neigh-

borhoods.

There are a number of directions for future work.

In the aspect-based sentiment analysis for guest pref-

erences, we should be able to obtain further advan-

tages in efﬁcacy by overcoming the lack of sufﬁcient

reviews in data sets by incorporating transfer learn-

ing approaches (Blitzer et al., 2007; Dai et al., 2007).

We used Rakuten Japanese travel data in the experi-

ments, while the method is applicable to other textual

reviews. To evaluate the robustness of the method,

experimental evaluation by using other data such as

grocery stores: LeShop

and movie data: movieLens

can be explored in future. Finally, comparison to

other recommendation methods, e.g., matrix factor-

ization methods (MF) (Koren et al., 2009) and combi-

nation of MF and the topic modeling (Wang and Blei,

2011) will also be considered in the future.

ACKNOWLEDGEMENTS

The authors would like to thank the referees for their

valuable comments on the earlier version of this pa-

per. We also thank Rakuten Institute of Technology

for providing Japanese travel data. This work was

supposed by the Grant-in-aid for the Japan Society

for the Promotion of Science (No. 25330255).

REFERENCES

Balabanovic, M. and Shoham, Y. (1997). Fab Content-

based Collaborative Recommendation. In Communi-

cations of the ACM, 40:66–72.

www.beshop.ch

http://www.grouplens.org/node/73

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

Dirichlet Allocation. Machine Learning, 3:993–1022.

Blitzer, J., Dredze, M., and Pereira, F. (2007). Biogra-

phies, Bollywood, Boom-boxes and Blenders: Do-

main Adaptation for Sentiment Classiﬁcation. In

Proc. of the 45th Annual Meeting of the Association

for Computational Linguistics, pages 187–295.

Brin, S. and Page, L. (1998). The Anatomy of a Large-

scale Hypertextual Web Search Engine. Computer

Networks, 30(1-7):107–117.

Cane, W. L., Stephen, C. C., and Fu-lai, C. (2006). Integrat-

ing Collaborative Filtering and Sentiment Analysis. In

Proc. of the ECAI 2006 Workshop on Recommender

Systems, pages 62–66.

Dai, W., Yang, Q., Xue, G., and Yu, Y. (2007). Boosting for

Transfer Learning. In Proc. of the 24th International

Conference on Machine Learning, pages 193–200.

Faridani, S. (2011). Using Canonical Correlation Analysis

for Generalized Sentiment Analysis Product Recom-

mendation and Search. In Proc. of 5th ACM Confer-

ence on Recommender Systems, pages 23–27.

Hofmann, T. (1999). Probabilistic Latent Semantic Index-

ing. In Proc. of the 22nd Annual International ACM

SIGIR Conference on Research and Development in

Information Retrieval, pages 50–57.

Huang, Z., Chen, H., and Zeng, D. (2004). Applying Asso-

ciative Retrieval Techniques to Alleviate the Sparsity

Problem in Collaborative Filtering. ACM Transactions

on Information Systems, 22(1):116–142.

Joachims, T. (1998). SVM Light Support Vector Machine.

In Dept. of Computer Science Cornell University.

Kobayashi, N., Inui, K., Matsumoto, Y., Tateishi, K., and

Fukushima, S. (2005). Collecting Evaluative Expres-

sions for Opinion Extraction. Journal of Natural Lan-

guage Processing, 12(3):203–222.

Koren, Y., Bell, R. M., and Volinsky, C. (2009). Matrix

Factorization Techniques for Recommender Systems.

In IEEE Computer, 42(8):30–37.

Kudo, T. and Matsumoto, Y. (2003). Fast Method for

Kernel-based Text Analysis. In Proc. of the 41st An-

nual Meeting of the Association for Computational

Linguistics, pages 24–31.

Lathia, N., Hailes, S., Capra, L., and Amatriain, X. (2010).

Temporal Diversity in Recommender Systems. In

Proc. of the 33rd ACM SIGIR Conference on Re-

search and Development in Information Retrieval,

pages 210–217.

Li, M., Dias, B., and Jarman, I. (2009). Grocery Shopping

Recommendations based on Basket-Sensitive Ran-

dom Walk. In Proc. of the 15th ACM SIGKDD Con-

ference on Knowledge Discovery and Data Mining,

pages 1215–1223.

Li, Y., Yang, M., and Zhang, Z. (2013). Scientiﬁc Articles

Recommendation. In Proc. of the ACM International

Conference on Information and Knowledge Manage-

ment CIKM 2013, pages 1147–1156.

Liu, N. N. and Yang, Q. (2008). A Ranking-Oriented Ap-

proach to Collaborative Filtering. In Proc. of the

31st Annual International ACM SIGIR Conference on

Research and Development in Information Retrieval,

pages 83–90.

IncorporatingGuestPreferencesintoCollaborativeFilteringforHotelRecommendation

L.Li, Zheng, L., Fan, Y., and Li, T. (2014). Modeling

and Broadening Temporal User Interest in Personal-

ized News Recommendation. In Expert system with

Applications, 41(7):3163–3177.

Niklas, J., Stefan, H. W., c. M. Mark, and Iryna, G. (2009).

Beyond the Stars: Exploiting Free-Text User Reviews

to Improve the Accuracy of Movie Recommendations.

In Proc. of the 1st International CIKM workshop on

Topic-Sentiment Analysis for Mass Opinion, pages

57–64.

Raghavan, S., Gunasekar, S., and Ghosh, J. (2012). Re-

view Quality Aware Collaborative Filtering. In Pro.c

of the 6th ACM Conference on Recommender Systems,

pages 123–130.

Sarwar, B., Karypis, G., Konstan, J., and Reidl, J. (2001).

Automatic Multimedia Cross-Model Correlation Dis-

covery. In Proc. of the 10th ACM SIGKDD Con-

ference on Knowledge Discovery and Data Mining,

pages 653–658.

Turney, P. D. (2002). Thumbs Up or Thumbs Down? Se-

mantic Orientation Applied to Un-supervised Classi-

ﬁcation of Reviews. In Proc. of the 40th Annual Meet-

ing of the Association for Computational Linguistics,

pages 417–424.

Wang, C. and Blei, D. M. (2011). Collaborative Topic Mod-

eling for Recommending Scientiﬁc Articles. In Proc.

of the 17th ACM SIGKDD Conference on Knowledge

Discovery and Data Mining, pages 448–456.

Wijaya, D. T. and Bressan, S. (2008). A Random Walk on

the Red Carpet: Rating Movies with User Reviews

and PageRank. In Proc. of the ACM International

Conference on Information and Knowledge Manage-

ment CIKM 2008, pages 951–960.

Xue, G. R., Yang, Q., Zeng, H. J., Yu, Y., and Chen, Z.

(2005). Exploiting the Hierarchical Structure for Link

Analysis. In Proc. of the 28th ACM SIGIR Confer-

ence on Research and Development in Information

Retrieval, pages 186–193.

Yates, B. and Neto, R. (1999). Modern Information Re-

trieval. Addison Wesley.

Yildirim, H. and Krishnamoorthy, M. S. (2008). A Random

Walk Method for Alleviating the Sparsity Problem in

Collaborative Filtering. In Proc. of the 3rd ACM Con-

ference on Recommender Systems, pages 131–138.

Yin, Z., Gupta, M., Weninger, T., and Han, J. (2010). A

Uniﬁed Framework for Link Recommendation Using

Random Walks. In Proc. of the Advances in Social

Networks Analysis and Mining, pages 152–159.

Zhao, X., Zhang, W., and Wang, J. (2013). Interactive Col-

laborative Filtering. In Proc. of the 22nd ACM Con-

ference on Information and Knowledge Management,

pages 1411–1420.

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval