Professional Documents
Culture Documents
3 - Computers & Industrial Engineering - For Strategy Yihau Shen
3 - Computers & Industrial Engineering - For Strategy Yihau Shen
A R T I C L E I N F O A B S T R A C T
Keywords: In the era of Industry 4.0, the rapid development of information technology in the last decade has provided new
Quality function deployment (QFD) challenges for product improvement by enabling users to give their feedback and sentiments in real time. In this
Online reviews paper, combining with genetic algorithm back propagation neutral network, fuzzy inference method, and
Text mining
entropy-based synthesis evaluation method, we raise a full-process product improvement solution driven by
Voice of customer (VoC)
GA-BP neural network
online reviews, from the initial online review collection to the final engineering characteristic prioritization. The
proposed novel integrated quality function deployment-based approach adheres to the customer-oriented design
principle, allowing manufacturers to strengthen the launched product based on the spontaneously-articulated
voice of customer, rather than the traditional expertise. In this way, an off-the-shelf product improvement
strategy is available for enterprises, and its special advantages like fast adaptation and real-time responsiveness,
would significantly reduce management costs, shorten response time to market dynamics, and enhance customer
satisfaction. In addition, a case study in smartphone industry is conducted for illustration, and the results clearly
demonstrate the effectiveness and practicability of the treatment.
* Corresponding author.
E-mail address: zhou_jian@shu.edu.cn (J. Zhou).
https://doi.org/10.1016/j.cie.2022.108233
Received 27 November 2021; Received in revised form 24 March 2022; Accepted 5 May 2022
Available online 11 May 2022
0360-8352/© 2022 Elsevier Ltd. All rights reserved.
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
without any time or space constraints, which enables them to include a two categories: selection of indicators of helpfulness and model of evalua
large amount of heterogeneous information, such as factual statements, tion of helpfulness. For the former, most end users (i.e., scholars and/or
sentiments, quantifiable satisfaction feedback, and even constructive practitioners) choose the indicators from the perspective of the cus
suggestions. In addition to offering multidimensional information, such tomers, such as metadata features (like emotional review information
online reviews have the merits of high availability and strong timeliness. (Felbermayr and Nanopoulos, 2016; Krishnamoorthy, 2015) and review
Online review data has a distinctive advantage in conveying rating information (Heng et al., 2018; Jones et al., 2004; Cao et al.,
customer demands. Nevertheless, in terms of uncovering and summa 2011)) and reviewer features (like reviewer characteristics (Pan and
rizing valuable information provided by a very large number of cus Zhang, 2011; Zhang and Lin, 2018), reviewer’s disclosure of relevance-
tomers, it poses an enormous new challenge to practitioners. VoC online descriptive information (Forman et al., 2008; Korfifiatis et al., 2008)).
review data is not only big in volume, but also extremely unstructured in Among these, it is worth mentioning that the online helpfulness voting
terms of data format, due to the freedom it provides the users to post ratio is regarded as the golden rule in measuring reviews’ usefulness (Liu
comments in any way they want. Machines and computers are capable of et al., 2013). However, to the best of our knowledge, a very limited
dealing with organized statistics, but are not intelligent enough to un number of studies have been conducted so far from the point of view of
derstand and quantify the feelings, sentiments, tones, irony, and context product engineers, who are responsible not only for collecting and
of human language. In order to comprehend textual content in large analyzing, but also for applying the reviews in the product improvement
quantities, text mining, as an interdisciplinary technique combining practice. In this literature (e.g., Liu et al., 2013; Qi et al., 2016), lin
statistics, machine learning, and computational linguistics, has been guistic features, features based on information quality, and features
developed to discover inner patterns and extract high-quality informa using information theory are the most popular features.
tion from plain-text data. Furthermore, we are concerned that most of Formulating a model for evaluating helpfulness, regression and
the literature utilizing text mining concentrates on one or two specific machine-learning methods are considered the two most popular ap
procedures of QFD, rather than complete QFD optimization. See Singh proaches. On the one hand, existing literature performs regression
et al. (2020) for quantifying consumer opinions, Trappey et al. (2018) analysis on selected indicators, such as ordinal logistic regression (Cao
and Ozdagoglu et al. (2018) for key terms and CR extraction, and Jin et al., 2011), multiple regression (Ghose and Ipeirotis, 2011), Tobit
et al. (2016) and Ireland and Liu (2018) for CR analysis. However, to our regression (Liu and Park, 2015), and partial least squares regression (Lee
knowledge, the utilization of text mining in the complete QFD optimi et al., 2018b). On the other hand, the common practice of the machine-
zation procedure is something that has, to date, received little attention learning methods is to collect a training reviews dataset through manual
in the literature (e.g., Jin et al. (2014)). annotation, and then achieve automated identification based on the
In truth, for online reviews, such a fresh source of VoC with clustering results. For example, see Lee et al. (2018a) for a ten-fold cross-
completely different carrier and linguistic features, practitioners may be validation, Lee and Choeh (2014) for the back propagation neural
disorientated about how to process and exploit this type of information, network, and Hu and Chen (2016) for the model tree approach.
and therefore, offering a thorough solution of product improvement is of To summarize, for the selection of good indicators of helpfulness,
great practical significance and extensive applicability for manufac both customers and product engineers have indispensable roles in the
turers. Motivated by this issue, our paper proposes a full-process QFD- review usefulness assessment. However, the majority of existing studies
based solution for facilitating data-driven product improvement, which pay considerably more attention to the customers’ perspective, and
may be summarized as follows. First, our treatment comprehensively there still exists an important gap on the helpfulness of online reviews,
selects and defines twenty-eight indicators of online review helpfulness, which has not been completely perceived or measured by product en
from both the customers’ and the product engineers’ perspective, and gineers. Besides, for the model of the evaluation of helpfulness, most
constructs an online review screening model of usefulness using an in research prefers machine-learning methods, which may involve some
formation entropy-based synthetic evaluation method. Second, this manual estimation.
research conducts CR identification with the support of the text-mining
methodology, and assigns weights to CRs using the fuzzy inference 2.2. CR identification and its prioritization
method. Finally, we map the CRs to ECs in a nonlinear way by imple
menting a genetic algorithm back propagation (GA-BP) neural network The identification of CRs is a vital step in QFD research, as it is the
model in the QFD relationship matrix. On this foundation, correspond starting point of the House of Quality. In previous studies, traditional
ing product improvement plans are formulated in the light of generated methods such as questionnaires (Haber et al., 2011; Amritesh and
EC priorities. Our proposed framework offers practitioners a detailed Chatterjee, 2018), panel discussions (Tontini, 2007; Vanany et al.,
product improvement guidance with the merits of quick response and 2019), interviews (Medeiros et al., 2018; Ferrari et al., 2016) and
ease of use, by which enterprises can respond to the constantly growing literature inspection (Shokouhyar et al., 2017; Piedras et al., 2006) are
and dynamic market in a real-time manner, and thereby, achieve higher generally used. Further, several studies combine these traditional tech
and more sustainable customer satisfaction. niques to comprehensively acquire customer demands (Li et al., 2018;
The rest of this research is organized as follows. In Section 2, related Dou et al., 2019; Borgianni, 2018; Yang, 2011; Fang et al., 2022).
literature regarding the evaluation of online review helpfulness, CR Nevertheless, in the big data era, online reviews have become an
identification and prioritization, and relationship matrix determination appealing source of information and business understanding. Consid
is respectively reviewed. The proposed novel theoretical framework for ering the importance of strong timeliness, high availability and spon
data-driven product improvement is illustrated in Section 3. Further taneous expression, ample research has made efforts to mine CRs from
more, in Section 4, the Xiaomi 8 mobile phone is chosen as a case to online reviews (e.g., Jin et al., 2016; Xu et al., 2011; Yin et al., 2013; Li
demonstrate the framework, and finally, some concluding remarks are and Wu, 2010; Singh et al., 2020; Li et al., 2019; Ireland and Liu, 2018;
provided in Section 6. Trappey et al., 2018; Ozdagoglu et al., 2018; Jin et al., 2015; Jin et al.,
2014).
2. Related Work In addition, understanding customers’ perceptions of CRs and how to
prioritize the CRs is also of great importance to decision makers. In
2.1. Online review helpfulness evaluation practice, it is very common for customers to evaluate every CR directly
using a five-point scoring scale (Nahm et al., 2013; Griffin and Hauser,
Considering the unevenly distributed quality of online reviews, first 1993). Also, some group decision-making techniques (Ho et al., 1999;
it is necessary to evaluate the helpfulness of the online reviews before Zhang and Chu, 2009), and multi-criteria decision-making methods like
making full use of them. So far, research in this area mainly considers AHP and ANP (Lu et al., 1994; Bhattacharya et al., 2005; Ucler, 2017),
2
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Fig. 1. The proposed novel theoretical framework for data-driven product improvement based on online reviews.
are widely employed in weighing CRs (Zhou et al., 2022). Following the et al., 2020; Wu et al., 2020; Xie et al., 2020). As opinion varies
proposal of the (Kano, 1984) model, several studies implement the Kano significantly on how to model the nonlinear correlation for the rela
categories (must-be, one-dimensional or attractive attribute) into CR tionship matrix in the House of Quality, further research is needed.
prioritization (Batarfi et al., 2017; Potra et al., 2017). Furthermore, to
fully utilize the imprecise information implied in judgments and lin 3. Proposed Framework
guistic descriptions, fuzzy theory has frequently been used to assign CR
importance (Gungor et al., 2011; Fang et al., 2020; Wang, 2013). In In this section, a novel theoretical framework for data-driven product
summary, looking back at these treatments, the majority involve a large improvement based on online reviews is proposed, and the whole pro
number of subjective judgments and evaluations, which may largely cess is graphically summarized in Fig. 1, which includes four phases.
depend on previous experience and knowledge. Firstly, we collect online reviews for the targeted product through web
crawlers, pre-treat them using a four-step process, and establish a
specialized set of dictionaries, such as a sentiment dictionary. For the
2.3. Relationship matrix determination
second phase, we select and define twenty-eight indicators under five
key features of online reviews for the evaluation of helpfulness, and on
One important use of QFD is to transform VoC into designer technical
this foundation, we design a helpful reviews screening model by utiliz
language, which is implemented through the relationship matrix of the
ing the information entropy method. Further, CRs are extracted from a
House of Quality. Questions such as “what should the relationship be
product characteristics dictionary through dimension reduction. In
tween CRs and ECs be?” and “How should the relationship be
order to determine the weights of the CRs, we formulate two indicators,
modelled?” have attracted great attention both in academia and in
namely attention and satisfaction, and the fuzzy inference method is
dustry. For the sake of simplicity, one mainstream approach is to
chosen to identify the inner connection between the CR weights and the
consider the relationship in a linear way, and thus propose corre
two above-mentioned indicators. Finally, we adopt the GA-BP neural
sponding solutions, such as linear regression (Celik and Ustasuleyman,
network to approximate the correlation between CRs and ECs due to its
2019; Liu et al., 2018) and linear programming (Becker and Gaivor
strong nonlinearity. With the optimized EC priorities outputted by a
onski, 2018; Kamvysi et al., 2014; Ko and Chen, 2014; Wang and Chin,
well-trained network, product engineers will be able to propose corre
2011). However, this linear assumption is so restrictive that it cannot
sponding product improvement plans.
properly fit the actual design and manufacturing. In order to expand the
scope of applications, an increasing number of studies have preferred to
adopt the nonlinear mode to simulate the relationship between CRs and 3.1. Online review collection and structuration
ECs, and finally to obtain the priority of ECs through various ap
proaches, such as the adaptive genetic algorithm (He et al., 2017), 0–1 As shown in Fig. 1, this framework begins with the collection of the
programming (Geng and Dong, 2017), and other nonlinear program customer review data. Online reviews are obtained by two major sour
ming methods (Liu et al., 2014; Chen and Chen, 2006). Moreover, multi- ces: eCommerce websites and the third party reviews sites, with
criteria decision-making methods are popular alternatives for deter different features. The former provides sale service, in which reviews are
mining EC prioritization (Wang et al., 2019; Ping et al., 2019; Ocampo in large quantities and with unevenly distributed information
3
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 1 Table 3
Four steps in data pre-treatment. Indicators of features based on information quality.
No. Steps Definition Common practice Indicator Indicator Description
alias
1 Data cleaning Generally, the majority of Screening unnecessary
the reviews collected from fields, filling in missing Information IQ-NSS Number of subjective sentences in the
websites are in non- values, transforming accuracy review text.
standardized language, in inconsistent formats, and IQ-NOS Number of objective sentences in the review
which data noise such as correcting wrong values text.
punctuation, emotions, through regular expression Information IQ-TIM Number of total elapsed days since the
parsed tags, hypertext or a manual process are timeliness review was posted.
markup language (HTML) commonly used Information IQ-NRP Number of products referred to in the review
codes, JavaScript (JS) approaches in data comparability text.
codes, and comments is cleaning. Information IQ-NPF Number of product features mentioned in
widely present. To coverage the review text.
strengthen the availability Information IQ-NSPF Number of sentences referring to product
of reviews, this data noise relevance features in the review text.
needs to be roughly IQ-RPFR Number of product features/Number of
cleared. sentences referring to product features.
2 Segmentation Segmentation means to Different from English IQ-RPFS Number of features/Number of sentences.
break sentences into pieces word segmentation, IQ-RRS Number of sentences referring to product
containing a word or Chinese word segmentation features/Number of sentences.
phrase, which is the divides successive Chinese
minimum independent character strings into
unit of meaningful several single words or locators (URL), parse the domain name system (DNS) in the identified
linguistic components. The phrases, according to the
URL, and finally download required information in big quantities. In this
precision of segmentation users’ expression patterns,
directly influences the and Chinese semantics and
regard, Python is recommended as the programming language, due to
accuracy of the semantic grammatical structure. high performance, easiness of operating, and its open source nature.
analysis of text. In most cases, raw data obtained through a web crawler is largely
3 POS tagging As the basis of text mining, At present, BosonNLP, unstructured, which poses significant challenges for subsequent anal
POS tagging consists of Language Technology
ysis. In order to improve the efficiency and accuracy of text mining, data
labelling every word or Plantform Cloud, NLPIR,
phrase with a tag for its Jieba, and SCWS are pre-treatment is necessary, including four steps: data cleaning, seg
part of speech, namely popular Chinese mentation, part of speech (POS) tagging and stop words removal. The
adjective, verb, noun, etc. segmentation and POS definition and common practice of the four steps are presented in
According to the POS tags, tagging systems/engines Table 1. Furthermore, considering sentiment analysis is a significant
we are able to identify that have been widely
what the targeted word is applied in the existing
procedure in follow-up sections, some relevant dictionaries need to be
trying to describe: a research. established for preparation, including sentiment words, adverbs of de
feature, opinion, or the gree and negative words.
degree of something.
4 Stop words In review text, some There are several well-
removal frequently occurring but recognized Chinese stop
3.2. Online review helpfulness evaluation
meaningless words like word dictionaries that can
auxiliary words and modal aid with stop word Since review data is provided in very big quantities, and its quality
words are known as stop removal, such as ‘Stop varies a lot, it is challenging for the product engineers to directly obtain
words. Removing stop words list of HIT’, etc.
useful information from it. Therefore, before further work is carried out,
words can largely improve
the efficiency of text it is necessary to establish an effective mechanism to evaluate the
mining. helpfulness of online reviews. In this subsection, the evaluation of re
view helpfulness includes two parts: indicator selection and quantifi
Abbreviations: POS: Part of speech; NLPIR: Natural language processing and
information retrieval sharing platform; SCWS: Simple Chinese word segmenta cation, and the construction of a review helpfulness screening model.
tion; HIT: Harbin Institute of Technology.
3.2.1. Indicator selection and quantification
As mentioned in Section 2.1, although the golden rule of evaluating
Table 2 review helpfulness from the perspective of customers has been deter
Indicators of linguistic features. mined, there still exists a gap that online review helpfulness is not
Indicator Indicator Description completely perceived and evaluated by product engineers, who is the
alias actual person to comprehend and utilize the reviews. Therefore, we
Number of words L-NW Number of words in the review text. choose and define twenty-eight indicators under five key features for
Number of sentences L-NS Number of sentences in the review text. evaluating the helpfulness of online reviews for both customers (i.e.,
Average length of L-ALS Average number of words per sentence in reviewer features and metadata features) and product engineers (i.e.,
sentences the review text. linguistic features, features based on information quality, and features
Number of adjectives L-NADJ Number of adjectives in the review text.
Number of adverbs L-NADV Number of adverbs in the review text.
using information theory).
Linguistic features: Linguistic features refer to the text features of
online reviews, and are like descriptive statistics of text information. We
availability. The latter one, as the platform for individuals volunteered select five indicators from Liu et al. (2013) to quantify the linguistic
to share products information and user experiences, contains less but features of online reviews, and detailed information such as indicator
higher quality review data. End-users may make their own decisions on alias, description and reference are presented in Table 2. Among them,
the selection of the data sources accommodating the actual demand. the first three indicators, number of words, number of sentences, and
After determining the data sources, web crawler, a kind of program or average length of sentences, portray information abundance and
script for systematically identifying sites and extracting valid informa comprehensiveness (Kuan et al., 2015), while number of adjectives and
tion, is employed to collect review data from a big pile of websites. The number of adverbs convey reviewers’ attitudes and emotional intensity,
operating principle of a web crawler is to visit a list of uniform resource respectively.
4
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 5 in which Sij indicates the sentiment about product feature fj occur
Indicators of reviewer features. ring in reviewi .
• Sentiment divergence (IT-DS)
Indicator Indicator Description
alias
Online reviews that contain both preferences and complaints stand
a higher chance of being rated as helpful, and the presence of
Number of reviews R-NR Number of reviews by the reviewer.
different sentiments in such reviews is called sentiment divergence.
Number of friends R-FRI Number of friends of the reviewer.
Number of followers R-FAN Number of followers of the We use the sum of the self-information, SI(fj , S), regarding both
reviewer. positive and negative sentiment, about product feature fj to quantify
Number of being R-FOCU Number of being focused of the the sentiment divergence, DS(fj ),
focused reviewer.
( ) ( )
∑
DS fj = SI fj , S . (4)
Features based on information quality: Liu et al. (2013) report pos,neg
5
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 6 For the benefit type of indicator, e.g. L-NW, the normalization
Indicators of metadata features. equation is given by
Indicator Indicator Description xij − minxij
alias aij = i
. (9)
maxxij − minxij
Whether pros is M-WPF Whether pros is completed or i i
completed not.
Whether cons is M-WCF Whether cons is completed or For the cost type of indicator, e.g. IQ-TIM, the normalization
completed not.
equation is provided by
Number of helpful votes M-NHV Number of helpful votes.
Number of replies M-NR Number of replies. maxxij − xij
Number of stars M-NS Number of stars. aij = i
. (10)
maxxij − minxij
i i
information about one or more aspects of data, and is used to describe When maxxij = minxij (the denominator equals 0), aij = 1.
i i
some properties of reviews. According to Liu et al. (2013), as a typical
Step 3. Entropy and entropy weight calculation
metadata feature, the voting ratio of helpfulness is assumed to be the
According to the definition, the information entropy of the jth
golden criterion in evaluating the helpfulness of a review from the
indicator is as follows:
perspective of the customer, because it is one of the most intuitive ap
proaches for expressing customer opinions.1 Thus, we include five in ∑
m
dicators in this category (see Table 6), including the descriptions of the Hj = − k fij lnfij , (11)
i=1
review text (provided by the reviewer), e.g. pros and cons, and the
evaluation level of the selected review (provided by other customers), where
such as number of helpful votes, replies, and stars.
aij
fij = ∑
m , (12)
3.2.2. Helpful review screening model based on information entropy method aij
In fact, the evaluation of the helpfulness of online reviews can be
i=1
6
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
characteristics candidates. Then, we conduct a word frequency analysis fields, while adverbs of degree M and negative words N are
to locate those candidates with a high frequency, and gather them optional fields.
together as the product characteristics dictionary. The captured product Step 2. Calculate emotional values: If M sentences can be obtained from
characteristic words are quite complicated and of a high dimensionality, a review, then the positive sentiment value, Spos
pi r , and the nega
so, finally, we extract the CRs by performing dimension reduction on the tive sentiment value, Sneg
pi r , of feature pi in review r can be
product characteristics based on the similarity of the semantics. calculated by
attention and satisfaction, with the aid of the text-mining technique. Sneg
pi r = Eneg
pi rs ⋅Mpi rs ⋅Npi rs , (18)
Furthermore, the fuzzy inference method is used to assign weights to the
s=1
(15)
pi ∈qi
Aq i = ∑ n ,
∑
N ∑
N
i=1
Fpi Spos
pi = Spos
pi r , Sneg
pi = Sneg
pi r . (20)
r=1 r=1
∑
N
Fpi = kpi r , (16) where Spos
pi and Spi
neg
are the positive and negative sentiment
r=1 values of the product feature pi in all selected online reviews.
where kPi r is the number of times product characteristic pi occurs in the Fuzzy inference method: As mentioned in Section 2.2, although a
review text r, and Fpi is the number of times the product characteristic pi lot of scholars have made efforts to identify the underlying mechanism
occurs in the texts of all the reviews. by which the factors influence the CR weights, it has been found to be
Determining “satisfaction” based on sentiment analysis: As the difficult to exactly and clearly identify the connection. Therefore,
name indicates, satisfaction reflects the sentiment inclination of the considering its easy-to-understand structure, as well as its interpretable
customers, where satisfaction and favor imply positive feelings, and and intuitive rule-based nature (Jassbi et al., 2006), the Mamdani
dissatisfaction and criticism imply negative feelings. For a certain CR, (1977)-type fuzzy inference method is adopted in this paper to deter
the satisfaction is represented by the ratio of the corresponding product mine the internal correlation between the CR weight and the two factors,
characteristics’ emotional values to the total product characteristics’ attention and satisfaction. The basic principle is to achieve a mapping
emotional values. based on a set of fuzzy reasoning rules, with crisp inputs and outputs.
Sentiment is a two-dimensional concept, including both the polarity The detailed procedures are as follows. Firstly, the membership func
and the degree of an attitude. Compared to attention, the calculation of tions of input (attention and satisfaction) and output (CR weight) values
satisfaction is rather complicated, with four dictionaries utilized: product are fuzzified as triangular fuzzy numbers, according to their corre
characteristics, sentiment, adverbs of degree and negative words. In the sponding linguistic descriptions. Next, we design a questionnaire for
calculation of satisfaction, taking each sentence as a unit, we firstly product engineers and related experts, to collect the recognized logical
check whether product characteristics words exist in the sentence by “conditions-result” combinations, and then determine the fuzzy infer
matching the segmentation results with the product characteristics ence rules. In this research, the basic form of a compositional fuzzy rule
dictionary. If at least one product characteristics word exists, then we go can be represented as
on to match the sentence to the sentiment dictionary, adverbs of degree
dictionary, and negative words dictionary, to calculate the emotional Rl : if attention is Al , and satisfaction is Bl , then CR weight is Cl ,
values. If no product characteristics word is found in the sentence, then
the sentence is discarded, and we proceed to check the next sentence, where Rl represents the lth fuzzy rule, l = 1, 2, …, s, attention and satis
until all online reviews have been processed. The specific process is as faction are input linguistic variables, CR weight is the output linguistic
follows: variable, Al , Bl and Cl are fuzzy sets defined on the universes of discourse
X, Y and Z, respectively. Thirdly, with a set of defined rules, Mamdani
Step 1. Identify the “product features-sentiment words” pairs: By (1977)’s approach is employed to perform fuzzy inference, which can be
matching the segmented sentence with the four dictionaries, we concluded to be a fuzzy max–min operation. Let μAl (x), μBl (y), and μCl (z)
can obtain a quadruple <Feature words F, Sentiment words E, be the membership functions of Al , Bl , and Cl , and then the calculation
Adverbs of degree M, Negative words N> for each sentence, in can be written as follows:
which product features F and sentiment words E are mandatory
7
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
U = {CR1 ,CR2 ,…,CRi ,…,CRn ,EC1 ,EC2 ,…,ECj ,…,ECm }, where the first n
terms are the weights of the CRs and the latter m terms are the corre
sponding priorities of the ECs, in the range [0, 10]. Subsequently, to
ensure the data of all indicators are of the same magnitude, the map
minmax function is used for normalization, which is calculated by
(ymax − ymin ) ∗ (x − xmin )
y= + ymin , (24)
xmax − xmin
where x and y are the input and output data, xmax and xmin are the
maximum value and minimum value of the input data, and ymax and ymin
are the maximum value and minimum value of the output data.
8
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
more than 190 million monthly active users (HKE, 2018), and the
Table 7
accumulation of online review data for them provides a solid foundation
Details of collected online reviews.
for applying our framework. In this regard, we choose Xiaomi 8, one of
Data Number of Details the star products in its smartphone series, for the case study discussed in
platform online reviews
this section to test and verify the effectiveness of the treatment.
ZOL.com 384 text content, time of posting reviews, number of
likes and reviews, score (0–10), reviewer
information (number of posts, friends, fans, 4.1. Online review collection and structuration
followers)
JD.com 500 text content, time of posting reviews, number of
likes and reviews
To collect online review data for Xiaomi 8, both of the major sources,
Taobao. 500 text content, time of posting reviews i.e., eCommerce websites and third-party review sites, are considered. In
com terms of eCommerce websites, three leading platforms in China - JD.
Suning.com 500 text content, time of posting reviews, number of com, Taobao.com and Suning.com - are chosen for their extensive
likes and reviews
coverage of online reviews. As one of the most valuable and professional
third-party websites for IT products, ZOL.com is also selected as a data
corresponding EC priorities by inputting the CR weights into this trained source. Then, we develop a Python program to crawl the webs and
network. Based on the obtained EC priorities, the product improvement collect online review data between 31/05/20182 and 01/05/2019. Part
team is able to propose a comprehensive product improvement plan, of the Python code is presented in Appendix G, and a summary of the
taking simultaneously some other factors into account, including but not details of the collected online reviews is shown in Table 7.
limited to manufacturers’ constraint, enterprise strategy, etc. We pre-treat the crawled data as follows: Firstly, we de-noise the
review text though regular expressions and manual operations, examine
4. Case Study the data consistency, and deal with the invalid and missing values. Next,
since online reviews are typically highly colloquial, containing casual
Smartphone technology is especially sensitive to customers’ re wording and being full of incorrectly written words, the accurate mode
quirements, so developing a practicable and effective product in Jieba, which better fits those qualities, is adopted to conduct seg
improvement paradigm is vital for smartphone manufacturers. As the mentation and POS tagging. Finally, we organize and aggregate four
world’s fourth-largest smartphone manufacturer (HKE, 2018), Xiaomi well-accepted stop words lists: the Chinese stop words list, the stop
has persisted in user-core designs since its foundation, which exactly fits words list of the Harbin Institute of Technology, the Baidu stop words
our research framework for turning updated VoC into product
improvement practice. Moreover, the Xiaomi series of smartphones have
2
Xiaomi 8 was launched on May 31st, 2018.
9
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 8
The weights of the indicators on JD.com.
Indicator Weight Indicator Weight Indicator Weight Indicator Weight
Table 9
The weights of the indicators on Suning.com.
Indicator Weight Indicator Weight Indicator Weight Indicator Weight
Table 10
The weights of the indicators on Taobao.com.
Indicator Weight Indicator Weight Indicator Weight Indicator Weight
list, and the stop words list of Sichuan University, into a refined stop 4.2. Online reviews helpfulness evaluation
words list. After matching the segmented reviews with the refined stop
words list, we are able to identify and remove the stop words. To obtain a helpful online reviews database for the subsequent
Furthermore, we build a comprehensive and improved sentiment analysis, we select several evaluation indicators for the helpfulness of
dictionary including 7176 positive sentiment words and 12062 negative online reviews from both the customers’ and the engineers’ perspective,
sentiment words, by integrating two well-recognized sentiment vocab and then build a helpfulness screening model based on the information
ulary lists: HowNet and the National Taiwan University Sentiment entropy method.
Dictionary (NTUSD). We also identify an adverb of degree dictionary We try to quantify the twenty-eight review indicators of helpfulness
(219 words) and negative words dictionary (31 words) based on (mentioned in Section 3.2.1) for each platform, and the collected data
HowNet. for the different platforms varies slightly. All twenty-eight indicators can
Table 11
The weights of the indicators on ZOL.com.
Indications Weight Indicator Weight Indicator Weight Indicator Weight
Table 12
The performance of the indicators before and after applying the helpfulness evaluation model (taking ZOL.com as an example).
Indicator L-NW L-NS L-ALS L-NADJ L-NADV IQ-NSS IQ-NOS
10
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 13 evaluation method is used to calculate the score of each online review.
Fifteen extracted CRs and corresponding product features. Listing the reviews in descending order, for each platform, we choose
CRs Product characteristics the top 200 reviews, giving 800 in total. In order to illustrate the validity
of our model, taking ZOL.com as an example, we present the indicator
Clear screen Screen, Desktop, Touch screen, Capacitive screen, Display,
Main screen, Curved surface, Resolution, Luminance, performance of the filtered and unfiltered reviews in Table 12. We
Display, Blue light, Interface, Color, Picture quality, observe that nearly all the indicators have been greatly improved. For
Clarity, Picture, Bangs, Internal screen, Split screen example, the average number of words (L-NW) in the reviews has
Durable battery Battery, Standby, Charger, Charging, Fast charge, Power, increased from 169 to 291, and the average number of product features
Power consumption, Duration, Endurance, Heat, Fever,
mA, Power supply, Power saving, Hot
mentioned (IQ-NPF) has risen from 6 to 9, showing that the filtered
Clear camera Camera, Flash, Pixel, Selfie, Soft light, Background, reviews provide more detailed descriptions and abundant information
Clarity, Color, Photography, Lens, Picture sense, Artifact, for the engineers. Besides, the average number of helpful votes (M-NHV)
Picture quality, Beauty, Effect, Shooting, Dual camera, of each review has increased from 73 to 133, and the average number of
Picture, Imaging, Photo
stars (M-NS) has grown from 5.46 to 8.35, indicating that the filtered
Large memory Memory, Capacity, RAM, ROM, Operating speed,
capacity Response speed, Delay, Storage, Space, SD online reviews receive higher popularity scores and recognition from
Complete accessories Accessories, Shell, Film, Glass film, Protective film, Cover, customers. In other words, the filtered reviews offer insights that help
Mobile phone cover, Earphone, Data cable, Power adapter capture CRs. In addition, the information timeliness (IQ-TIM) has been
Smooth system System, Operation, Smooth, IOS, Android, Black screen, reduced from 272 days to 259 days, so that the reviews are more time-
Bug, Compatible, Flashing root, Frozen, Startup,
Stuttering, Stuck machine, Not stuck, Jailbreak
sensitive and representative. Overall, after the reviews have been pro
Complete functions Applications, Programs, APP, Functions, Software, cessed by the helpfulness screening model, and only the most helpful
Navigation, Webpages, Browser, Maps, Entertainment, selected, the performance of the chosen reviews has been significantly
Play games, Mobile games, Arena of Valor, Function, Call, improved compared to the performance of the entire sample, building a
MMS, SMS, Alarm clock, Address book, Information, GPS
robust data foundation for follow-up procedures.
Audiovisual Multimedia, Radio, Sound, Voice, Ringtones, Earphones,
entertainment Sound quality, Music, Video, Audio, Movies, Volume,
Player, TV Shows, Speakers, Listening to music, Boom, 4.3. CRs and the determination of their importance
Noise, Receiver, Call, Sound effect, Alarm
Beautiful appearance Appearance, Shape, Body, Volume, Gap, Craftsmanship,
4.3.1. CR identification
Color, Lines, Design, Personality, Shell, Size, Model,
Beautiful, Scratches, Thickness, Frame, Back cover, Back As demonstrated in Section 3.3.1, we build the product character
shell, Size, Pretty istics dictionary and extract CRs through dimension reduction, for which
Comfortable feel Feel, Weight, Texture, Touch, Material, Workmanship, the “Synonymy Thesaurus of the Harbin Institute of Technology” is
Material
selected as the supportive material, and some manual operations are
Satisfied service After-sales, Service, Logistics, Customer service, Attitude,
Word of mouth, SF, Seller, Maintenance, Warranty, included. The fifteen extracted CRs and their corresponding product
Packaging characteristics are displayed in Table 13. The fifteen CRs are “Clear
Reasonable price Price, Cost-effective, Price reduction, Discount, Pricing, screen”, “Durable battery”, “Clear camera”, “Large memory capacity”,
List price, Money, Quotation, Cheap “Complete accessories”, “Smooth system”, “Complete functions”, “Au
Fast running speed Processor, CPU, Snapdragon, Quad-core, Single-core, GPU
diovisual entertainment”, “Beautiful appearance”, “Comfortable feel”,
Good network signal Network, Signal, Internet, WLAN, 4G, Bluetooth, Dual-
band, WIFI, Infrared, Wireless, Network speed “Satisfied service”, “Reasonable price”, “Fast running speed”, “Good
Multi-functional Fingerprint, Unlock, Face recognition, Voice recognition network signal”, and “Multi-functional unlock”. It is noteworthy that
unlock there might be a slight semantic difference between the extracted CRs
and their corresponding product characteristics. For example, a series of
product characteristics like “Battery”, “Standby”, “Charger”, “Fast
be found in the third-party website, ZOL.com, while only twenty-one,
charge”, “Duration” and “Endurance” etc., describe more than one
twenty-one, and nineteen indicators are quantified in JD.com, Taobao.
aspect of a battery, including its durability and speed of charging.
com, and Suning.com, respectively. In these eCommerce websites, we
However, we tended to retain the duration as its representative feature
failed to obtain some of the reviewer features and metadata features
(that is, the CR “Durable battery”) after considering their occurrences in
because these kinds of websites have specific privacy policies to protect
online reviews. It can be seen from Table 13 that not only the basic
the reviewers’ information.
requirements are extracted (e.g. “Clear screen” and “Durable battery”),
As introduced in Section 3.2.2, the information entropy method is
but also some updated and even attractive requirements are identified
performed as follows: First, we construct four evaluation matrices as
(e.g.“Multi-functional unlock”), the latter of which would be difficult to
given by Eq. (8), and then these matrices are normalized using Eqs. (9)
identify through traditional methods like questionnaires.
and (10) to obtain the non-dimensional forms. Finally, according to Eqs.
(11)–(14), the entropy weights of the indicators on the four platforms
4.3.2. CR weights determination based on fuzzy inference method
are calculated, and the results are shown in Tables 8–11, respectively. As
In order to obtain the CR weights, the customers’ attention to and
shown in the bold part, although the reviewer and metadata features
satisfaction with product needs have to be determined. We quantify
cannot be found in every platform, it is obvious that these kinds of
customers’ attention paid to the Xiaomi 8 through word frequency
features have a remarkably significant influence on reviews’ helpfulness
analysis (Eqs. (15) and (16)), and the results are shown in Table 14. It is
evaluations.
obvious that “Clear screen” is the CR with the highest attention score of
After the entropy weights have been obtained, the synthesis
19.37, far exceeding the second highest one. Not surprisingly, CRs
Table 14
Attention scores of CRs.
CR Clear screen Durable battery Clear camera Large memory capacity Complete accessories
Attention 19.37 10.42 8.32 9.72 1.65
CR Smooth system Complete functions Audiovisual entertainment Beautiful appearance Comfortable feel
Attention 9.85 10.55 4.12 7.58 2.27
CR Satisfied service Reasonable price Fast running speed Good network signal Multi-functional unlock
Attention 0.66 6.72 2.51 1.19 5.07
11
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 15
Satisfaction scores of CRs.
CR Clear screen Durable battery Clear camera Large memory capacity Complete accessories
Satisfaction 0.2727 0.1735 0.4395 0.3125 − 0.1034
CR Smooth system Complete functions Audiovisual entertainment Beautiful appearance Comfortable feel
Satisfaction 0.4438 0.3919 0.2990 0.4044 0.2840
CR Satisfied service Reasonable price Fast running speed Good network signal Multi-functional unlock
Satisfaction 0.3913 0.6468 0.5281 0.2857 0.3297
receiving the most attention are those must-be attributes like screen,
Table 16
battery, system, capacity, etc. It is worth mentioning that some
Sixteen fuzzy inference rules obtained.
emerging CRs, such as “multi-functional unlock”, are also of high
1 IF Attention IS L Satisfaction IS L THEN CR IS NI concern to users. Moreover, customers are inclined to attach more
2 IF Attention IS L Satisfaction IS M THEN CR IS NI
importance to the product itself than to services.
3 IF Attention IS L Satisfaction IS H THEN CR IS ENI
4 IF Attention IS L Satisfaction IS VH THEN CR IS ENI For the CR satisfaction quantification, we follow the three-step
5 IF Attention IS M Satisfaction IS L THEN CR IS I procedures described in Section 3.3.2, and present the obtained CR
6 IF Attention IS M Satisfaction IS M THEN CR IS I satisfaction scores in Table 15. In general, except for “Complete acces
7 IF Attention IS M Satisfaction IS H THEN CR IS NI
sories”, all the satisfaction scores of the CRs are greater than 0, with an
8 IF Attention IS M Satisfaction IS VH THEN CR IS NI
9 IF Attention IS H Satisfaction IS L THEN CR IS VI
average value of 0.3716, showing that customers are basically content
10 IF Attention IS H Satisfaction IS M THEN CR IS I with the Xiaomi 8. Specifically, “Reasonable price”, as shown in bold in
11 IF Attention IS H Satisfaction IS H THEN CR IS I the table, has the highest satisfaction score of 0.6468, which indicates
12 IF Attention IS H Satisfaction IS VH THEN CR IS I that the price of this phone is regarded as reasonable and acceptable by
13 IF Attention IS VH Satisfaction IS L THEN CR IS EI
consumers. In practice, the Xiaomi series of phones indeed enjoy a
14 IF Attention IS VH Satisfaction IS M THEN CR IS VI
15 IF Attention IS VH Satisfaction IS H THEN CR IS I reputation of being cost-effective, and this is also a critical competitive
16 IF Attention IS VH Satisfaction IS VH THEN CR IS I advantage of the Xiaomi company. However, the only negative score for
satisfaction, which belongs to “Complete accessories”, should be taken
VL: very low; L: low; M: medium; H: high; VH: very high.
EUI: extremely unimportant; UI: unimportant; I: important; VI: very important; seriously. Otherwise, this dissatisfaction could directly influence cus
EI: extremely important. tomers’ appetite for buying the phone.
After the attention and satisfaction results have been obtained, the
fuzzy inference method is performed to acquire the CR weights. To apply
the fuzzy inference method, a set of inference rules should first be
determined. In this case, we collect data from two groups of people:
customers and engineers, using interviewing questionnaires (see Ap
pendix A). In terms of consumers, we adopt a convenient sampling
approach approaching one hundred ten potential interviewees who
intend to purchase a new smartphone within the next year from the area
Fig. 5. The membership function of the input variables: CRs’ attention and satisfaction.
12
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 17
The weights of the CRs.
CR Clear screen Durable battery Clear camera Large memory capacity Complete accessories
Weight 6.2292 4.4806 3.0748 3.5250 3.1901
CR Smooth system Complete functions Audiovisual entertainment Beautiful appearance Comfortable feel
Weight 3.5629 3.8073 3.3326 2.7809 3.1593
CR Satisfied service Reasonable price Fast running speed Good network signal Multi-functional unlock
Weight 1.5510 2.4980 2.1036 2.7785 3.1786
Table 18
Seventeen ECs of Xiaomi 8.
Screen CPU ROM Battery Network
Camera Design Size Weight Body material
Common functions Operational systems Graphics processor Random access memory Fingerprint unlock
Video and audio support Accessories
Fig. 7. The error evolution curve and training results of the GA-BP neural network.
3
In our case, ten engineers are included in interviewing questionnaires. As
mentioned above, all the ten engineers have the experience of using Xiaomi
smartphones, and some of them are even long-time users. Hence, they can be
assumed as a special group of customers, who treat CRs with both considerable
expertise in product development, and the original intention from actual
adopters, and this is the reason why the product engineers are also included.
13
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Fig. 5(b)), respectively. This circumstance meets the inference rules 6, 7, appropriate attention to forward-looking attributes, such as the camera
10, and 11 in Table 16. Processed with fuzzy max–min operation and the and unlocking mode, in order to maintain a differentiated competitive
centroid method, the crisp weight value 4.4806 is produced according to advantage.
Eqs. (21)–(23). The visualized fuzzy inference process is illustrated in In fact, Xiaomi 9, the successor to Xiaomi 8, was publicly released on
Appendix D. February 20, 2019. It comes as no surprise that the new characteristics of
Similarly, the weights of all fifteen CRs are obtained by performing Xiaomi 9 are broadly in line with our proposed improvement plan,
the fuzzy inference method, and the results are shown in Table 17. In especially for the screen and battery. Xiaomi 9 not only has a vastly
this table, “Clear screen” (bold in Table 17) is the most essential CR with increased screen-to-body ratio, up from 83.83% to 90.7% due to open
the highest weight of 6.2292, followed by “Durable battery” with ings design, but also supports 27 W fast charging, which dramatically
4.4806. In fact, not a few reviews have mentioned that Xiaomi 8’s shortens the battery charging time. In addition to the two major en
screen-to-body ratio and battery are barely satisfactory, and indeed, hancements, there have been a number of improvements to other ECs,
inferior performance of such “must-be” attributes will stimulate an ur such as the rear fingerprint being replaced by under-screen fingerprint,
gent desire among customers for those attributes to be improved. etc. Since being launched in February 2019, Xiaomi 9 has received a
Moreover, “Satisfied service” has the lowest importance due to its warm response from both the market and customers. ZOL.com, as one of
extremely low attention score, indicating that the targeted customers are the most professional interactive forums for IT products in China, offers
more concerned about the product itself, rather than related services. a comprehensive score for the major IT products, which is drawn from a
The weights of the other CRs are in the same level with small variations. large number of customers as well as IT enthusiasts. The comprehensive
score given to Xiaomi 9 was up to 9.00, far exceeding Xiaomi 8’s score of
4.4. Product improvement based on a GA-BP neural network-based QFD 6.4, and indicating that the new characteristics of Xiaomi 9 fitted cus
approach tomers’ wishes exactly. The popularity of Xiaomi 9 also proves the
practicability and validity of our proposed framework.
Next, we determine seventeen ECs of Xiaomi 8 on the basis of phone
specifications and expertise (see Table 18). Furthermore, we collect 5. Discussion
training data for a GA-BP neural network, which consists of two parts:
CR weights and EC priorities. To be precise, we distribute another 5.1. Theoretical implication
questionnaire to the one hundred ten customers (the same group of re
spondents we interviewed in Section 4.3.2), in which each of them is First, a full-process product improvement solution, which in
invited to rate the importance of the fifteen CRs in a ten-point scale corporates the informative value of customer big data and the instru
(from 1 to 10). However, seven respondents were unable to finish the mental value of quality function deployment, is proposed in this paper.
questionnaire properly, and we further removed forty-three samples In the previous literature, online review data has already been utilized in
with repetitive answers. As a result, sixty sets of CR weights are the field of product development, however, most of the literature
collected. Further, we request the ten specially-invited engineers to concentrated on one or two specific procedures of QFD (Singh et al.,
provide corresponding EC priorities in accordance with each set of CR 2020; Ozdagoglu et al., 2018; Ireland and Liu, 2018), and practitioners
weights. Hence, sixty sets of CRs importance with corresponding ECs’ may be disorientated about how to exploit this new form of VoC. This
priorities are prepared as the training data set, and being pre-processed study built a bridge between product improvement and the big data of
using Eq. (24), subsequently. VoC, specifically, online review data provides an abundant and reliable
After specifying the ECs and collecting the training data, we deter supply of customer demand, while QFD, as a theoretical mature tool,
mine some basic parameters of the GA and the structure of the BPNN. offers general guidance of product improvement. In the era of data
The specific settings are shown in Appendices E and F, respectively. We science, this off-the-shelf product improvement strategy with special
use Matlab to realize the designed GA-BP neural network, and the core advantages like real-time responsiveness and fast adaption may gain
part of the code is presented in Appendix G. The GA error evolution some insights for the emerging data-driven product improvement.
curve is shown in Fig. 7(a), and it clearly depicts that the error evolution Further, previous product improvement methods that extract useful
curve reaches a stable state with an error value of 5.84, after about fifty information from online reviews such as (Jin et al., 2016; Ozdagoglu
iterations. The acquired optimal initial weights and thresholds from the et al., 2018; Jin et al., 2014), were built under crisp and deterministic
GA optimization are used for BPNN training, and the training results are environments. However, in the practical product improvement, most of
shown in Fig. 7(b), in which the error of the GA-BP neural network the involved parameters, like customer satisfaction, linguistic descrip
gradually decreases to the expected level of 0.0001 after only nine it tion, expertise, etc., always potentially encompass uncertainty, which is
erations. So far, our designed GA-BP neural network is well trained to an also why fuzzy set theory has been comprehensively utilized in product
acceptable level for mapping the nonlinear relationship. development during the past decades. Our treatment making full use of
With the CR weights obtained in Section 4.3 inputted, the trained the advantages of fuzzy inference system, provides an appropriate
GA-BP neural network outputs the EC priorities of Xiaomi 8, as displayed approach to model customers’ perception (attention and satisfaction, in
in Table 19 along with their ranks. Obviously, among the seventeen ECs our case) on a certain product with the help of triangular fuzzy numbers,
of Xiaomi 8, “Screen” ranks top with an importance score of 6.8224, and as long as human’s judgment, attitude or verbal description are
closely followed by “Battery” with 4.5025, indicating that product en included, our approach may act as an available solution to give a proper
gineers should assign more resources to the improvement of these two measurement on them for further exploration. To summarize, the
features. In addition, “Common functions” should be carefully consid method we proposed has taken the vagueness existing in product
ered. Combining the EC priorities and the corresponding online reviews, development into consideration, which suits for a wider range of real-
we find that customers’ dissatisfaction is mainly concentrated on the low world applications, especially for the “customer-oriented” design
screen-to-body ratio and disappointing battery life. Accordingly, we philosophy.
propose an improvement plan for Xiaomi 8, which contains three di
mensions. First, the screen-to-body ratio of Xiaomi 8 needs to be 5.2. Practical implication
increased through the removal of the so-called “fringe” and “big chin”,
which would enlarge the available area of the phone. Secondly, the While our case was concentrated on the smartphone industry, it
battery life of Xiaomi 9 should be prolonged, and innovative charging should be noted that our proposed treatment is also applicable to some
approaches need to be developed as far as possible. Finally, in addition other industrial contexts, as long as the target product has accumulated
to the attributes that may cause dissatisfaction, Xiaomi should pay adequate high quality online reviews, because the quality and quantity
14
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
of selected online reviews are essential data basis of the final results. make full use of online reviews, an important form of user-generated
What is more, our framework may performed better for those fast iter contents, this paper proposes a full-process QFD-based solution for
ative products, such as fast moving consumer goods and consumer data-driven product improvement, which is of great practical signifi
electronics, because this kind of products is characterized as high fre cance and extensive applicability for manufacturers. First, for the online
quency of purchase and wide coverage of target customers, and as a review helpfulness assessment, we consider both the customers’ and the
result, the variation of customers demand and attitude would be product engineers’ perspectives, and include information from online
instantly reflected in the review data. reviews to evaluate the model construction, which has not received
As illustrated above, the review data crawling and processing are appropriate attention in the existing literature. Second, we propose two
critical components in the suggested framework, and careful consider indicators which are developed with the aid of text mining, attention
ation should be given to these procedures in practice. The first thing is to and satisfaction, and the fuzzy inference method is applied to prioritize
cover as a greater diversity of review data type as possible when CRs. Finally, a GA-BP neural network model is designed and adopted,
determining the review source, including but not limited to eCommerce which is appropriate for the relationship matrix of QFD due to its strong
platforms, third-party review sites and social media, etc., because the nonlinearity. With the proposed treatment, manufacturers are capable
obtained results would be more reliable if more multi-dimensional of making quick response to market dynamics in a real-time manner
customer information is collected, despite the increase in workloads with minimum adaptation cost.
and difficulties in data processing. Secondly, in our case of Xiaomi 8, we This research is not free of limitations, and also offer channels for
employed two kinds of sentiment words: positive sentiment words and future research. One of the critical limitations is the data acquisition
negative sentiment words, by integrating two accepted sentiment vo issue. On the one hand, some personalized data and detailed information
cabulary lists (HowNet and the National Taiwan University Sentiment could not be accessed due to privacy protection. One the other hand, the
Dictionary). In fact, the numbers of the two dictionaries differ slightly sources of this research are finite, only encompassing eCommerce
(7176 for positive and 12062 for negative), and the obtained sentiment websites and third-party review sites. In the future, other platforms
strength might be partially biased by such unbalanced dictionary. might prove indispensable for VoC gathering, such as social media and
Hence, when practising the suggested framework, it is recommended exclusive communities. In addition to the data acquisition dilemma, this
that a balanced lexicon with roughly equivalent amount of sentiment research presumably fails to offer a thorough guidance of handling
words with different polarities should be established. manufacturers’ constraints due to limited accessibility. Therefore,
Finally, as mentioned in Section 4.3.1, for a CR, we only preserve another possible direction of future research could include manufac
those with the greatest occurrence due to some practical restrictions, turers’ constraints and other aspects of QFD into this full-process solu
whereas this choice does not mean that the other product characteristics tion, further facilitating manufacturers’ rapid responsiveness and
with relatively low frequency might not be useful somehow. When it adjustments to market requirements and needs.
goes to the last step of our approach, i.e., formulating an overall product
improvement plan, it is strongly recommended that the practitioners Declaration of Competing Interest
should trace back to the table of product characteristics, and give a
second thought on those “discarded” dimensions of CRs. Such a review The authors declare that they have no known competing financial
might provide useful insights and offer inspirations for product engi interests or personal relationships that could have appeared to influence
neers, offering a shortcut of creating attractive features, and Xiaomi 9’s the work reported in this paper.
initiative on fast-charging is just such a case.
Acknowledgements
6. Conclusion
The authors would like to acknowledge the gracious support of this
As a critical trend in the transformation of economic structure in work by the National Natural Science Foundation of China (Grant No.
Industry 4.0, the utilization of digital technologies provides significant 71872110) and the National Foreign Experts Program of China (Grant
supports for the emerging product improvement paradigm. In order to No. G2021013045L). Any remaining errors are ours.
Table 20
Table 20
The questionnaire structure of determining fuzzy inference rules.
Attention (Level) Satisfaction (Level) CRs importance (ENI) CRs importance (NI) CRs importance (I) CRs importance (VI) CRs importance (EI)
L L
L M
L H
L VH
M L
M M
M H
M VH
H L
H M
H H
H VH
VH L
VH M
VH H
VH VH
15
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Remarks: Please check your ideal weight (√) according to the attention and satisfaction level.
Appendix B. Details of the engineers invited in the questionnaire
Table 21
Personal details of the ten invited product engineers agreed to participate in our survey.
Engineer Age Seniority Education Title Occupation
Table 21
Table 22.
Table 22
The number of interviewees of different levels of attention and satisfaction.
Attention (Level) Satisfaction (Level) CRs importance (ENI) CRs importance (NI) CRs importance (I) CRs importance (VI) CRs importance (EI)
L L 11 80 5 0 0
L M 8 78 8 1 1
L H 75 10 5 4 2
L VH 72 12 6 5 1
M L 2 6 80 5 3
M M 1 9 75 10 0
M H 4 82 8 2 0
M VH 2 88 4 2 0
H L 1 3 6 78 8
H M 2 6 72 14 2
H H 1 1 3 86 5
H VH 0 0 14 70 12
VH L 0 1 1 4 90
VH M 0 1 3 86 6
VH H 0 8 80 8 0
VH VH 0 2 88 4 2
16
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Fig. 8.
Table 23.
Table 23
GA parameters settings.
Parameters Settings
17
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Table 24.
Table 24
BP neural network structure settings.
Structure Settings
References Heng, Y., Gao, Z., Jiang, Y., & Chen, X. (2018). Exploring hidden factors behind online
food shopping from Amazon reviews: A topic mining approach. Journal of Retailing
and Consumer Services, 42, 161–168.
Amritesh, M. S. C., & Chatterjee, J. (2018). Quality framework for credence-based
HKE (2018). Xiaomi Corporation (1810) Global Offering. Available online:https://www.
informational services: applying Kano’s method. Total Quality Management &
hkenews.hk (accessed on 25 December 2020).
Business Excellence, 29(1–2), 116–147.
Ho, E. S. S. A., Lai, Y. J., & Chang, S. I. (1999). An integrated group decision-making
Basuroy, S., & Ravid, S. A. (2013). How critical are critical reviews? The box office effects
approach to quality function deployment. IIE Transactions, 31(6), 553–567.
of film critics, Star Power, and Budgets. Journal of Marketing, 67(4), 103–117.
Hu, N., Zhang, J., & Pavlou, P. A. (2009). Overcoming the J-shaped distribution of
Batarfi, B., Guergachi, A., & Wahab, M. I. M. (2017). The life cycle of a feature modelling
product reviews. Communications of the ACM, 52(10), 144–147.
the transitions between feature states. International Journal of Quality and Reliability
Hu, Y., & Chen, K. (2016). Predicting hotel review helpfulness: The impact of review
Management, 34(8), 1229–1251.
visibility, and interaction between hotel stars and review ratings. International
Becker, D. M., & Gaivoronski, A. A. (2018). Optimisation approach to target costing
Journal of Information Management, 36(6), 929–944.
under uncertainty with application to ICT-service. International Journal of Production
Ireland, R., & Liu, A. (2018). Application of data analytics for product design: Sentiment
Research, 56(5), 1904–1917.
analysis of online product reviews. CIRP Journal of Manufacturing Science and
Bhattacharya, A., Sarkar, B., & Kumar, S. M. (2005). Integrating AHP with QFD for robot
Technology, 23, 128–144.
selection under requirement perspective. International Journal of Production Research,
Jassbi, J.J., Serra, P.J.A., Ribeiro, R.A., and Donati, A. (2006). A comparison of Mandani
43(17), 3671–3685.
and Sugeno inference systems for a space fault detection application. 2006 World
Borgianni, Y. (2018). Verifying dynamic Kano’s model to support new product/service
Automation Congress, pages 1 – 8.
development. Journal of Industrial Engineering and Management, 13(4), 569–587.
Jin, J., Ji, P., & Liu, Y. (2014). Prioritising engineering characteristics based on customer
Cao, Q., Duan, W., & Gan, Q. (2011). Exploring determinations of voting for the
online reviews for quality function deployment. Journal of Engineering Design, 25
‘helpfulness‘ of online user reviews: A text mining approach. Decision Support
(7–9), 303–324.
Systems, 50(2), 511–521.
Jin, J., Ji, P., Liu, Y., & Lim, S. J. (2015). Translating online customer opinions into
Celik, P., & Ustasuleyman, T. (2019). Integrated QFD, fuzzy linear regression and ZOGP:
engineering characteristics in QFD: A probabilistic language analysis approach.
An application of E-Store design. International Journal of Business Analytics, 6(4),
Engineering Applications of Artificial Intelligence, 45, 115–127.
61–73.
Jin, J., Liu, Y., Ji, P., & Liu, H. (2016). Understanding big consumer opinion data for
Chen, Y., & Chen, L. (2006). A non-linear possibilistic regression approach to model
market-driven product design. International Journal of Production Research, 54(10),
functional relationships in product planning. The International Journal of Advanced
3019–3024.
Manufacturing Technology, 28, 1175–1181.
Jones, Q., Ravid, G., & Rafaeli, S. (2004). Information overload and the message
Dou, R., Lee, W., & Nan, G. (2019). An integrated approach for dynamic customer
dynamics of online interaction spaces: A theoretical model and empirical
requirement identification for product development. Enterprise Information Systems,
exploration. Information Systems Research, 15(2), 194–210.
13(4), 448–466.
Kamvysi, K., Gotzamani, K., Andronikidis, A., & Georgiou, A. C. (2014). Capturing and
Fang, X., Shen, Y., Zhou, J., Pantelous, A.A., and Zhao, M. (2020). Qfd-based product
prioritizing students’ requirements for course design by embedding Fuzzy-AHP and
design for multisegment markets: a fuzzy chance-constrained programming
linear programming in QFD. European Journal of Operational Research, 237(3),
approach. IEEE Transactions on Engineering Management (online).
1083–1094.
Fang, X., Zhou, J., Zhao, H., & Chen, Y. (2022). A biclustering-based heterogeneous
Kano, N. (1984). Attractive quality and must-be quality. Hinshitsu (Quality, The Journal of
customer requirement determination method from customer participation in product
Japanese Society for Quality Control), 14, 39–48.
development. Annals of Operations Research, 309(2), 817–835.
Ko, W., & Chen, L. (2014). An approach of new product planning using quality function
Felbermayr, A., & Nanopoulos, A. (2016). The role of emotions for the perceived
deployment and fuzzy linear programming model. International Journal of Production
usefulness in online customer reviews. Journal of Interactive Marketing, 36(3), 60–76.
Research, 52(6), 1728–1743.
Ferrari, A., Spoletini, P., Silva, C., & Gnesi, S. (2016). Ambiguity and tacit knowledge in
Korfifiatis, N., Rodríguez, D., and Sicilia, M.A. (2008). The impact of readability on the
requirements elicitation interviews. Requirements Engineering, 21, 333–355.
usefulness of online product reviews: A case study on an online bookstore. World
Forman, C., Ghose, A., & Wiesenfeld, B. (2008). Examining the relationship between
Summit on Knowledge Society. Springer, Berlin, Heidelberg, 70(5):423–432.
reviews and sales: The role of reviewer identity disclosure in electronic markets.
Krishnamoorthy, S. (2015). Linguistic features for review helpfulness prediction. Expert
Information Systems Research, 19(3), 291–313.
Systems with Applications, 42(7), 3751–3759.
Geng, X., & Dong, X. (2017). Configuration optimization of product functional
Kuan, K. Y., Hui, K. L. B., Lai, H. Y., & Prasarnphanich, P. (2015). What makes a review
requirements based on cloud model and information axiom. Computer Integrated
voted? An empirical investigation of review voting in online review systems. Journal
Manufacturing Systems, 24(1), 154–163.
of the Association for Information Systems, 16(1), 48–71.
Ghose, A., & Ipeirotis, P. G. (2011). Estimating the helpfulness and economic impact of
Lee, P., Hu, Y., & Lu, K. (2018a). Assessing the helpfulness of online hotel reviews: A
product reviews: Mining text and reviewer characteristics. IEEE Transactions on
classification-based approach. Telematics and Informatics, 35(2), 436–445.
Knowledge and Data Engineering, 23(10), 1498–1512.
Lee, S., & Choeh, J. Y. (2014). Predicting the helpfulness of online reviews using
Griffin, A., & Hauser, J. R. (1993). The voice of the customer. Marketing Science, 12(1),
multilayer perceptron neural networks. Expert Systems with Applications, 41(6),
1–27.
3041–3046.
Gungor, Z., Delice, E. K., & Kesen, S. E. (2011). New product design using FDMS and
Lee, S. G., Trimi, S., & Yang, C. G. (2018b). Perceived usefulness factors of online
FANP under fuzzy environment. Applied Soft Computing, 11(4), 3347–3356.
reviews: a study of Amazon.com. Journal of Computer Information Systems, 58(4),
Haber, N., Fargnoli, M., & Sakao, T. (2011). Integrating QFD for product-service systems
344–352.
with the Kano model and fuzzy AHP. Total Quality Management & Business Excellence,
Li, N., & Wu, D. (2010). Using text mining and sentiment analysis for online forums
31(9–10), 929–954.
hotspot detection and forecast. Decision Support Systems, 48(2), 354–368.
He, L., Wu, Z., Xu, Z., Zheng, M., & Ming, X. (2017). Quantification and integration of an
Li, S., Tang, D., & Wang, Q. (2019). Rating engineering characteristics in open design
improved Kano model into QFD based on multi-population adaptive genetic
using a probabilistic language method based on fuzzy QFD. Computers & Industrial
algorithm. Computers & Industrial Engineering, 114, 183–194.
Engineering, 135, 348–358.
18
Y. Shen et al. Computers & Industrial Engineering 169 (2022) 108233
Li, Y., Du, Y., & Chin, K. (2018). Determining the importance ratings of customer Singh, A., Jenamani, M., & Thakkar, J. (2020). Do online consumer reviews help to
requirements in quality function deployment based on interval linguistic evaluate the performance of automobile manufacturers. Journal of Enterprise
information. International Journal of Production Research, 56(14), 4692–4708. Information Management, 33(5), 1153–1198.
Liu, Y., Han, Y., Zhou, J., Chen, Y., & Zhong, S. (2018). Establishing the relationship Tontini, G. (2007). Integrating the Kano model and QFD for designing new products.
matrix in QFD based on fuzzy regression models with optimized h values. Soft Total Quality Management, 18(6), 599–612.
Computing, 22, 5603–5615. Trappey, A. J. C., Trappey, C. V., Fan, C. Y., & Lee, I. J. Y. (2018). Consumer driven
Liu, Y., Jin, J., Ji, P., & Harding, J. A. (2013). Identifying helpful online reviews: A product technology function deployment using social media and patent mining.
product designer’s perspective. Computer-Aided Design, 45(2), 180–194. Advanced Engineering Informatics, 36, 120–129.
Liu, Y. Y., Zhou, J., & Chen, Y. (2014). Using fuzzy non-linear regression to identify the Ucler, C. (2017). Brainstorming the cryoplane layout by using the iterative AHP-QFD-
degree of compensation among customer requirements in QFD. Neurocomputing, 142, AHP approach. Aviation, 21(2), 55–63.
115–124. Vanany, I., Maarif, G. A., & Soon, J. M. (2019). Application of multi-based Quality
Liu, Z., & Park, S. (2015). What makes a useful online review? Implication for travel Function Deployment (QFD) model to improve halal meat industry. Journal of Islamic
product websites. Tourism Management, 47, 140–151. Marketing, 10(1), 97–124.
Lu, H. L., Madu, C. N., Kuei, C. H., & Winokur, D. (1994). Integrating QFD, AHP and Wang, C. (2013). Incorporating customer satisfaction into the decision-making process of
benchmarking in stragetic marketing. Journal of Business and Industrial Marketing, 9 product configuration: a fuzzy Kano perspective. International Journal of Production
(1), 41–50. Research, 51(22), 6651–6662.
Mamdani, E. H. (1977). Application of fuzzy logic to approximate reasoning using Wang, X., Fang, H., & Song, W. (2019). Technical attribute prioritisation in QFD based on
linguistic synthesis. IEEE Transactions on Computers, 26(12), 1182–1191. cloud model and grey relational analysis. International Journal of Production Research,
Medeiros, J., Vasconcelos, A., Silva, C., & Goulao, M. (2018). Quality of software 58(19), 5751–5768.
requirements specification in agile projects: A cross-case analysis of six companies. Wang, Y., & Chin, K. (2011). A linear goal programming approach of determining the
Journal of Systems and Software, 142, 171–194. relative importance weights of customer requirements in quality function
Nahm, Y. E., Ishikawa, H., & Inoue, M. (2013). New rating methods to prioritize deployment. Information Sciences, 181(24), 5523–5533.
customer requirements in QFD with incomplete customer preferences. International Wu, S., You, X., Liu, H., & Wang, L. (2020). Improving quality function deployment
Journal of Advanced Manufacturing Technology, 65(9–12), 1587–1604. analysis with the cloud MULTIMOORA method. Sustainable Production and
Ngo-Ye, T. L., & Sinha, A. P. (2014). The influence of reviewer engagement Consumption, 27, 1600–1621.
characteristics on online review helpfulness: a text regression model. Decision Xie, J., Qin, Q., and Jiang, M. (2020). Multiobjective decision-making for technical
Support Systems, 61(4), 47–58. characteristics selection in a house of quality. Mathematical Problems in
Ocampo, L. A., Labrador, J. J. T., Jumaoas, A., & Rama, A. M. O. (2020). Integrated Engineering, 2020.
multiphase sustainable product design with a hybrid quality function deployment – Xu, K., Liao, S., & Li, J. (2011). Mining comparative opinions from customer reviews for
multi-attribute decision-making (QFD- MADM) framework. Sustainable Production competitive intelligence. Decision Support Systemst, 50(4), 743–754.
and Consumption, 24, 62–78. Yang, C. (2011). Identification of customer delight for quality attributes and its
Ozdagoglu, G., Kapucugil-Ikiz, A., & Celik, A. F. (2018). Topic modelling-based decision applications. Total Quality Management, 22(1), 83–98.
framework for analysing digital voice of the customer. Total Quality Management & Yin, P., Wang, H., & Guo, K. (2013). Feature–opinion pair identification of product
Business Excellence, 29(13–14), 1545–1526. reviews in Chinese: A domain ontology modeling method. New Review of Hypermedia
Pan, Y., & Zhang, J. (2011). Born unequal: A study of the helpfulness of user-generated and Multimedia, 19(1), 3–24.
product reviews. Journal of Retailing, 87(4), 598–612. Zhang, Y., & Lin, Z. (2018). Predicting the helpfulness of online product reviews: A
Piedras, H., Yacout, S., & Savard, G. (2006). Concurrent optimization of customer multilingual approach. Electronic Commerce Research and Applications, 27, 1–10.
requirements and the design of a new product. International Journal of Production Zhang, Z., & Chu, X. (2009). Fuzzy group decision-making for multi-format and multi-
Research, 44(20), 4401–4416. granularity linguistic judgments in quality function deployment. Expert Systems with
Ping, Y., Liu, R., Lin, W., & Liu, H. (2019). A new integrated approach for engineering Applications, 36(5), 9150–9158.
characteristic prioritization in quality function deployment. Advanced Engineering Zhou, J., Shen, Y., Pantelous, A.A., and Liu, Y. (2022). Quality function deployment: A
Informatics, 45. bibliometric-based overview. IEEE Transactions on Engineering Management
Potra, S. A., Izvercian, M., Pugna, A. P., & Dahlgaard, J. J. (2017). The HWWP, a refined (online).
IVA-Kano model for designing new delightful products or services. Total Quality Zhou, J., Zhai, L., & Pantelous, A. A. (2020). Market segmentation using high-
Management & Business Excellence, 28(1–2), 104–117. dimensional sparse consumers data. Expert Systems with Applications, 145, 113136.
Qi, J., Zhang, Z., Jeon, S., & Zhou, Y. (2016). Mining customer requirements from online Zhu, F., & Zhang, X. (2013). Impact of online consumer reviews on sales: the moderating
reviews: A product improvement perspective. Information and Management, 53(8), role of product and consumer characteristics. Journal of Marketing: A quarterly
951–963. publication of the American marketing association, 74(2), 133–148.
Shannon, C. E., & Weaver, W. (1948). The mathematical theory of communication. Bell Zhu, L., Yin, G., & He, W. (2014). Is this opinion leader’s review useful? Peripheral cues
System Technical Journal, 27(3), 379–423. for online review helpfulness. Journal of Electronic Commerce Research, 15(4),
Shokouhyar, S., Safari, S., & Mohsenian, F. (2017). The HWWP, a refined IVA-Kano 267–280.
model for designing new delightful products or services. Total Quality Management &
Business Excellence, 28(1–2), 104–117.
19