=Paper=
{{Paper
|id=Vol-2068/wii8
|storemode=property
|title=An E-Commerce Recommender System using Complaint Data and Review Data
|pdfUrl=https://ceur-ws.org/Vol-2068/wii8.pdf
|volume=Vol-2068
|authors=Toshinori Hayashi,Yuanyuan Wang,Yukiko Kawai,Kazutoshi Sumiya
|dblpUrl=https://dblp.org/rec/conf/iui/HayashiWKS18a
}}
==An E-Commerce Recommender System using Complaint Data and Review Data==
https://ceur-ws.org/Vol-2068/wii8.pdf
An E-Commerce Recommender System using Complaint
Data and Review Data
Toshinori Hayashi Yuanyuan Wang Yukiko Kawai
Kwansei Gakuin University Yamaguchi University Kyoto Sangyo University
den82687@kwansei.ac.jp y.wang@yamaguchi-u.ac.jp kawai@cc.kyoto-su.ac.jp
Kazutoshi Sumiya
Kwansei Gakuin University
sumiya@kwansei.ac.jp
ABSTRACT
In recent years, the use of e-commerce recommender systems
has become more widespread, with key applications including
tracking user purchase histories, considering value estimates
and product review comments, and recommending higher-
rated related items. However, traditional recommender sys-
tems are not consistent when recommending alternative items
based on user input. Although users choose options on an exist-
ing product or service, (e.g., screen size and quality), it is still
difficult to satisfy users’ requirements. To solve this problem,
we propose a novel item recommender system that analyzes
two kinds of data: complaint data from the Fuman Kaitori
Center and reviewer comments on e-commerce. First, the
system generates the negative vectors of user-checked items
from complaint data and positive vectors of related item data
by subtracting lower-rated reviews from higher-rated reviews. Figure 1. Example of Recommendation
Next, the system calculates the similarities between these two
vectors and determines which reviews can resolve complaints even consumers find reliable information, they still face diffi-
related to user-checked items. Thus, the proposed system can culty finding items that meet their needs. Thus, we propose
provide suitable substitutes for user checked items. In this pa- a novel recommender system that extracts the drawbacks of
per, we describe our proposed recommendation method based products [4] and reviews. Consumer’s needs tend to come
on complaint data and review data, and verify its efficacy using from complaints; therefore, in our proposed system facilitates
qualitative evaluation. users finding suitable items based on their complaints. Finally,
ACM Classification Keywords this paper is focused on complaints related to products (and
H.1.2 User/Machine Systems: Human information processing not services). Hence, users can find suitable alternative items
(item Y1 , item Y2 , and item Y3 ) that will address complaints
Author Keywords regarding their original purchase (item X) (see Figure1).
E-Commerce; Recommender System; Complaint; Review; RELATED WORK
INTRODUCTION E-commerce recommenders have been extensively studied.
As the prevalence of e-commerce has increased, many rec- Yandi et al. [7] proposed a recommender system that used
ommender systems have been proposed by researchers [1, 2, coupons. In addition, different methods of recommending sub-
5, 6, 9]. These systems help users to find items and promote stitutes have been studied and are well-documented. McAuley
their purchase. In addition, a user can view other consumers’ et al. [3] considered the relationship between substitutes and
product reviews to acquire information about a given item be- complements based on items with reviews and their cost. They
fore purchasing. However, some reviews contain both positive proposed methods for clarifying the relationship between sub-
and negative information. In these cases, users often cannot stitutes and complements for a given topic. Zheng et al. [10]
recognize genuine complaints about a given item. Moreover, also analyzed the relationship between substitutes and comple-
ments by applying economic principles. In a similar vein, our
Permission to make digital or hard copies of all or part of this work for per-
sonal or classroom use is granted without fee provided that copies are not
proposed method recommends items based on complaint data
made or distributed for profit or commercial advantage and that copies bear and review data.
this notice and the full citation on the first page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with Finally, review bias has also been extensively studied. Zhang
credit is permitted. To copy otherwise, or republish, to post on servers or to et al. [8] calculated review bias using 2 factors: user prefer-
redistribute to lists, requires prior specific permission and/or a fee. Request ences and prejudices caused by reading other reviews. Since
permissions from Permissions@acm.org.
©2018. Copyright for the individual papers remains with the authors. we propose a recommender system based on review data, we
Copying permitted for private and academic purposes. will analyze the reliability of said data.
WII’18, March 11, 2018, Tokyo, Japan
Table 1. Categories of Complaints
Main # of Sub Main # of Sub
Life 14 Vehicle 14
Fashion 10 Hobby 13
Food 9 Restaurant 9
Medical 5 Outdoor 6
Consumer Electronics 6 Industry 14
Health 5 Sightseeing 3
Public environment 6 Education 10
International Culture 5 Politics 4
Human relationships 8 Jobs 5
Pets 6 Other 1
DATASET OF COMPLAINTS Figure 2. System Overview
One of the datasets used with our proposed system is com-
plaint data provided by Insight Tech Inc from the Fuman
Kaitori Center website, which provides a platform for con-
sumer complaints. Here, users post complaints with their own
account on the website, receiving points each time they post.
Furthermore, they can exchange these points for coupons,
which can be used on purchases. Consumers post complaints
regarding a wide range of products, services, and subjects.
Each complaint contains the following metadata: Posted
user ID, Category, and Complaints. Our proposed system
uses Complaints, Product_name, Category, and Sub_category.
Since Product_name is not required, the system uses complaint
data with Product_name. Figure 3. System Interface
Table 1 shows complaint categories, and each category has VF denotes the vectors comprising negative feature words from
several subcategories. In total, these amount to 20 main cat- complaints. VR denotes the vectors with positive feature words
egories and 153 subcategories. Users choose a category and from reviews. In the next section we further explain this vector
subcategory when they post complaints. generation.
PROPOSED SYSTEM In basic terms, this system recommends other items which
System Overview complaints for an item X by finding an item Y which has same
The purpose of this system is to recommend alternatives to feature words and has been positively evaluated. This system
items with which users are dissatisfied. Figure 2 shows the uses similarity values to rank item Ys for each complaint.
overview of our proposed system. For item X, which is an Extraction of Negative Feature Words of Complaints
item that the user peruses, this system finds item Y which Our proposed system uses a complaint dataset. 50% of com-
solves one of the problems with item X. plaint data ia labeled “product_name” which denotes the object
To accomplish this, the system first extracts the item’s corre- of a complaint. This system only uses complaint data which
sponding complaint data and generates vectors from negative has “product_name” in its metadata.
feature words. At the same time, the system extracts the re- First, the system extracts nouns from complaints by each item.
view data of related items (item Ys) from e-commerce sites to Following this, it calculates the weight of each feature word
generate positive feature vectors (we use related items from using the term frequency–inverse document frequency (TF-
the same categories). However, many reviews contain both IDF) method as follows:
positive and negative words. To remove the negative feature ni, j |D|
words from the generated vectors, the system generates 2 vec- t fi, j · id fi = · log (2)
tors from high-evaluation and low-evaluation reviews. Next, ∑k nk, j d fi
by subtracting low-evaluation vectors from high-evaluation d denotes the document that is integrated by all complaints for
vectors, the vectors with positive feature words are calculated. one item
Finally, the system finds a suitable alternative (item Y) by
calculating similarity between negative feature vectors and Extraction of Positive Feature Words of Item Review
positive feature vectors. It calculates the vectors using cosine This system analyzes items checked by consumers on e-
similarity as follows: commerce sites, and extracts reviews of related items (item Y)
∑ wn · (wpi − wni ) to identify positive and negative words related to both. For re-
Sim(vF , vR ) = q (1) lated items, this system generates feature vectors by extracting
|V | p
∑i=1 (wn )2 · ∑(wpi − wni )2 positive words from review data.
Table 2. Top 15 negative feature Words of item A with evaluation. Table 4. Top 10 negative feature words from complaints.
Complaint Evaluation Score Review Evaluation Score Items Top 10 feature words
Stereo No-N initial No-N item B scratch, cover, Sheet, processing,
music No-N friend No-N
XPERIA Negative download No-N
book, Built in, Type, Grip sensor,
battery Negative Car navigation No-N player, flash.
car No-N self No-N item C Exchange, Photo, Recently, input,
sound No-N Sim No-N character, capacity, battery, Radio wave,
memory Negative Beauty No-N Repair, SoftBank.
movie Negative Initially No-N
USB Negative Packaging Negative
item D Call out, Going, number, phone
sebum No-N necessary No-N fingertip, purpose, Partner, limit
LCD Negative Terminal Negative worry, History, browser, Incoming.
package Negative ROM No-N
brightness Negative A moment No-N Table 5. Top 10 positive feature words from reviews.
Skype Negative Upper-part No-N
group Negative level No-N Items Top 10 feature words
item 1 compact, heat, volume, position,
Table 3. Average precision of item A to item D. button, sound, Conpact, Replacement,
Average precision Complaint Review Authentication, Book
item A 0.49 0.16 item 2 Radio wave, going out, reaction,
item B 0.46 0.41 Real racing, Preparation, frequency,
item C 0.53 0.65 Pursuit, defect, heat, China
item D 0.36 0.27 item 3 charging, Body, Caution, scam, sim,
Purchase, A case, Trouble, Upper-left, regret
item 4 Professional, Notice, Completion,
First, this system extracts two types of item reviews (high- Reconstruction, charging, Excitement,
evaluation reviews and low-evaluation) for each item, and Contact, Spoofing, Malignant, Free of charge
generates two types of vectors based on these reviews. Evalu- item 5 Activation lock, OCN, SIM, Cheap, Safety,
ation values are distinguished by the number of stars on five Lock, Profile, Relief, Sim, Connection.
grading criteria. Reviews of 4 or 5 stars are evaluated as high, item 6 action, attachment, Abundance,
and reviews of 1 or 2 stars are evaluated as low. image quality, camera, Attached,
To generate vectors, this system uses the same formula as (2). Travel, Success, Snorkeling, Record
D denotes twice the number of same categorized items on an item 7 Windows, keyboard, board, PC,
e-commerce site, as the system generates two vectors per item. Key, computer, 64bit, cost, USB, Allowance
Thus, the feature vectors Vr p and Vr n are generated.
vrp = (wp1 , wp2 , · · · , wpi , · · · , wpm ) (3)
vrn = (wn1 , wn2 , · · · , wni , · · · , wnm ) (4) shown along with the number and nature of complaints posted
by other users.(center, Figure 3). Finally, when the user clicks
Finally, this system subtracts the weight of low-evaluation a type of complaint, the system shows an item that solves the
vectors Vr n from the weight of high-evaluation vectors Vr p on problem expressed in that complaint (right, Figure 3).
each item. This calculation generates vectors containing only
positive words for each item. Feature words with negative EVALUATION
values imply negative subjects. This system does not require Our evaluations were performed using two data types: com-
negative words for generated vectors, so it removes feature plaint data (provided from Insight Tech Inc.) and review data
words with negative values. Thus, the generated vectors have from Amazon. Each item had approximately 50 complaint
only positive feature words, and each value implies the level posts. In the Amazon data, each item had approximately 60
of positivity for these feature word. For example, the feature high-evaluation and low-evaluation reviews. We evaluated 4
words with values at or near 1 are considered very positive: items with 200 complaint posts, and 7 items with 420 reviews.
vR = (vrp − vrn ) =
Comparison of Complaints and Negative Review
(wp1 − wn1 , · · · , wpi − wni , · · · , wpm − wnm ) (5)
We assumed that complaint data shows more negative features
Interface of items than review data. Users referred to opened reviews on
This system finds items with positive reviews relating to nega- e-commerce sites during their shopping; however, the contents
tive points of a user-checked item. Based on common feature of complaint data were more straightforward than those of
words of the negative vectors of item Xs and positive vectors review data (review data is closed and users tend to only post
of item Ys, this system recommends alternative items for each when complaining). To verify those differences, we evaluated
consumer. These are ranked based on the calculated similari- via a five-subject questionnaire.
ties between X and Y items. Figure 3 shows the interface of
our proposed system. We extracted feature words from both complaint and review
data for the same items from the same category (phones).
The system displays items as general e-commerce websites Following this, we subtracted high-evaluation reviews from
(left, Figure 3). When users click on an item, its details are low-evaluation reviews to remove positive words.
Table 6. Similarity between Negative Complaints and Positive Reviews. Table 8. Average of Recommendation Evaluated Manually.
Similarity item A item B item C item D Values item 1 item 2 item 3 item 4 item 5
item 1 0.60 0.16 0.46 0.05 Complaint1 0.00 0.00 0.33 0.33 0.33
item 2 0.40 0.17 0.46 0.20 Complaint2 0.50 0.83 0.33 0.67 1.00
item 3 0.18 0.20 0.28 0.10 Complaint3 0.17 0.17 0.00 0.00 0.00
item 4 0.43 0.27 0.60 0.06 Complaint4 0.00 0.83 0.67 0.83 1.00
item 5 0.33 0.25 0.63 0.12 Complaint5 0.33 0.83 0.83 0.33 1.00
item 6 0.03 0.06 0.23 0.02
item 7 0.21 0.17 0.53 0.04
Table 6 shows the values of the similarities between negative
Table 7. Recommendation on Proposed System of Each Complaint. vectors from complaints and positive vectors from reviews
used to rank items. High values of similarity could be seen
Values item 1 item 2 item 3 item 4 item 5
after normalizing for items A and C. However, the values for
Complaint1 0.00 0.00 0.50 0.25 1.00
items B and D are lower because their vectors have many
Complaint2 1.00 0.60 0.8 0.40 0.20
dimensions. We can also see that our proposed methods is
Complaint3 0.50 0.00 0.25 0.00 1.00
effective for recommending other categorized items (such as
Complaint4 0.00 0.50 0.25 0.75 1.00
cameras and PCs).
Complaint5 0.00 1.00 0.75 0.25 0.50
Table 7 shows the results of system recommendations by com-
plaint for item A. For example, the system recommended 3
Table 2 shows the negative feature words of item A from both items for complaint 1 based on feature words. The values
complaints and reviews. We also calculated for items B, C, were calculated using the order of ranking score and similarity
and D, and verified the difference between the feature words values between items.
from complaints and reviews via five-subject questionnaire.
• Complaint1: Complaint about battery
Table 2 also shows the results of this evaluation for item A. • Complaint2: Complaint about Internet connection
“Negative” means they recognized as negative objects of the • Complaint3: Complaint about charger
item. “No-N” means they did recognized as non-negative
objects. We adopted majority rule to calculate precision for • Complaint4: Complaint about photo and memory
their different answers. Subjects evaluated each feature words • Complaint5: Complaint about reaction of screen
after checking complaints and reviews of each item. To evaluate our proposed system, we produced answer data for
Table 3 shows the average precision from items A–D. The recommendations via questionnaire. 6 subjects rated whether
value of precision for extracting negative words from com- items 1–5 could solve the problem of certain complaints. Sub-
plaints is higher than that from reviews on all items except jects scored 1.0 if the recommended item could solve the
item C. Some words extracted from reviews on item C were problem of each complaint, and 0.0 if it could not. Table 8
too general. We could see more negative feature words ex- shows the averages of the rating results.
tracted from complaints than from reviews. In future work, we The correlation coefficient value between the results of pro-
must verify this discrepancy for many items to remove general posed system and the manual answer data was 0.45. We can
terms on TF-IDF methods. see that our proposed system performs with similar accuracy
Comparison of System and Results Manually to manual recommendation. In future work, we plan to add
We evaluated this system compared to the results of manually more factors to raise the correlation coefficient value.
selecting alternative items. We first calculated the number of
CONCLUSION
recommended items using our proposed method, the complaint
data of 4 items, and the review data of 7 items. The categories In this paper, we proposed a recommender system that uses
for these items are described below. complaint and review data to recommend alternative purchase
items on e-commerce websites. We extracted feature words
item A to item D :items from complaints categorized phone from complaints with low-evaluation reviews, and verified
item 1 to item 5 :items from reviews categorized phone the efficacy of our proposed system with 200 complaints and
420 reviews. Our results showed that complaint data was an
item 6 :items from reviews categorized Camera effective information source for our purposes, and that our
item 7 :items from reviews categorized PC proposed system performed well when compared with manual
recommendation methods.
Table 4 shows the top 10 feature words we extracted for items
B–D using complaint data. The feature words extracted for In future work, we plan to further validate the accuracy of
item A are shown in table 2. Most of these feature words complaint data by analyzing consumer complaints for many
imply the objects of complaints. more items. Furthermore, we will consider new extraction
methods to enhance the precision of the proposed system.
Table 5 shows the top 10 feature words extracted for items 1-7
using review data. It may prove beneficial to only extract posi- ACKNOWLEDGEMENTS
tive words using our method; however, we will also examine In this paper, we used an FKC Data Set provided for research purposes by the
extracting additional parts of speech in future work. National Institute of Informatics in cooperation with Insight Tech Inc.
REFERENCES 6. B. Smith and G. Linden. 2017. Two Decades of Recommender Systems
1. L. N. Anh-Thu, H-H Nguyen, and N. Thai-Nghe. 2016. A Context-aware at Amazon. com. IEEE Internet Computing 21, 3 (2017), 12–18.
implicit feedback approach for online shopping recommender systems. 7. Y. Xia, G. Di Fabbrizio, S. Vaibhav, and A. Datta. 2017. A
In Asian Conference on Intelligent Information and Database Systems. Content-based Recommender System for E-commerce Offers and
Springer, 584–593. Coupons. (2017).
2. U. Leimstoll and H. Stormer. 2007. Collaborative recommender systems
for online shops. AMCIS 2007 Proceedings (2007), 156. 8. Xiaoying Zhang, Junzhou Zhao, and John C.S. Lui. 2017. Modeling the
Assimilation-Contrast Effects in Online Product Rating Systems:
3. J. McAuley, R. Pandey, and J. Leskovec. 2015. Inferring Networks of Debiasing and Recommendations. In Proceedings of the Eleventh ACM
Substitutable and Complementary Products. In Proc. of the 21th ACM Conference on Recommender Systems (RecSys ’17). ACM, New York,
SIGKDD International Conference on Knowledge Discovery and Data NY, USA, 98–106. DOI:
Mining (KDD ’15). 785–794. http://dx.doi.org/10.1145/3109859.3109885
4. K. Mitsuzawa, M. Tauchi, M. Domoulin, M. Nakashima, and T.
Mizumoto. 2016. FKC Corpus: a Japanese Corpus from New Opinion 9. Qi Zhao. 2016. E-commerce Product Recommendation by Personalized
Survey Service. In Proc. of the Novel Incentives for Collecting Data and Promotion and Total Surplus Maximization. In Proc. of the 9th ACM
Annotation from People: types, implementation, tasking requirements, International Conference on Web Search and Data Mining. 709–709.
workflow and results. 11–18. 10. J. Zheng, X. Wu, J. Niu, and A. Bolivar. 2009. Substitutes or
5. A. O Omondi and A. W Mbugua. 2017. An Application of association Complements: Another Step Forward in Recommendations. In Proc. of
rule learning in recommender systems for e-Commerce and its effect on the 10th ACM Conference on Electronic Commerce (EC ’09). 139–146.
marketing. (2017).