Identification of key customer requirements based on online reviews

Abstract

Customer requirements are the essential driving force for successful product development. They can be grouped into several categories, including basic requirements, indifferent requirements, reverse requirements, expected requirements, and attractive requirements. Among these, the latter two are crucial for improving customer satisfaction and can be classified as key requirements. However, the literature on identifying key requirements suffers from issues related to subjective interference and the lack of a specific quantitative calculation process. Thus, this study proposes a model for identifying critical customer requirements. First, use Python to run the web crawler for extracting online customer reviews. Second, extract product engineering characteristics using the relevant text mining technology and latent Dirichlet allocation topic clustering algorithm. Third, we combine sentiment analysis and other factors that influence customer satisfaction with the product engineering characteristics to conduct the conjoint analysis and calculate utility values for the product engineering characteristics. Finally, integrate Kano model to formulate the requirements hierarchy rules, determine the final key requirements index, and identify the key customer requirements. And a case study implemented the key customer requirements identification problem for a smartphone to demonstrate the feasibility and effectiveness of the proposed methodology.

Keywords

Key customer requirements engineering characteristic extraction online reviews Kano model product update

1 Introduction

As international competitive pressure increases, product homogeneity intensifies, innovation costs rises, and customer requirements become more diverse. More companies are focusing the limited resources on determining how to better capture key customer requirements (KCRs) and achieve product innovation. The Kano model groups customer requirements into five categories: basic requirements, indifferent requirements, reverse requirements, expected requirements, and attractive requirements [1]. When the basic requirements are met, customer satisfaction is not significantly improved; instead, if they are not met, the customer is dissatisfied. Indifferent requirements have no effect on user satisfaction regardless of whether they are met, and customers are dissatisfied when reverse requirements are met. Unlike the other requirements, the expected and attractive requirements have natural advantages, especially the attractive requirements, as they can greatly improve user satisfaction when they are satisfied [1], these two requirements are strongly competitive for companies committed to improve product design quality and customer satisfaction [2]. Thus, identifying requirements that can increase customer satisfaction, particularly expected and attractive requirements, is essential for improving product design quality and customer satisfaction, and called this kind of requirements as KCRs.

The current research on customer requirements mining is divided into two categories based on whether they create a hierarchy of requirements [3, 4]. Related studies that do not consider the requirements hierarchy are often based on market investigations, and requirements mining is carried out using mass requirements data with clear concentration trends. However, in the face of market competition, analysing requirements that have achieved a satisfactory level is not valid, it needs to please customers through unexpected and attractive functions [5]. In recent years, some studies have focused on determining customers’ potential requirements [6]. The prominent features of such potential requirements are that they are hidden and unconscious, but they are not always the key requirements that can raise customer satisfaction to promote product or service improvements. To stand out amongst fierce market competition, companies not only identify potential requirements but also identify KCRs that are more sensitive to satisfaction. The requirements hierarchy is usually based on the Kano model, and some studies emphasize the significance of expected and attractive requirements. Among them, a previous study proposes a user demand mining model [7]. This model analyses the emotional tendencies of review text, but the process for quantitative analysis is oversimplified, as it focuses on the effect of emotional factors on customer requirements, and other factors are rarely involved.

Thus, this study integrates the above two categories to mine customer potential requirements from the perspective of customer satisfaction using online review data and creates a requirements hierarchy based on the Kano model. A modified model of mining requirements is established by integrating various factors to quantify the customer satisfaction or dissatisfaction utility for engineering characteristics, and KCRs can be identified based on expected and attractive requirements data.

Thus, a method is needed to comprehensively analyse and quantify the utility values of multiple factors that embody customer requirements in online data, and we use conjoint analysis (CA) to determine the importance of these requirements by calculating the horizontal utility value of each factor that has been extensively used to identify customer preferences and requirements in recent years [8]. CA is therefore applied to quantify various factors reflecting customer requirements and determine the overall utility values of the factors in this study.

At the same time, an index is required to quantify the importance of customer requirements. Online reviews can effectively reflect customer preferences and express customer satisfaction or dissatisfaction. Customer requirements are implied when customers comment on product engineering characteristics. The customer satisfactory or dissatisfactory utility values of product engineering characteristics can be seen as quantifying the importance of customer requirements and can be used to calculate the key requirements index (KRI). Finally, the KCRs are identified according to the quantitative results.

The rest of this article is organized as follows. Section 2 reviews the literature and Section 3 explains the framework and methodology. Section 4 illustrates the case studies and provides a discussion. Finally, Section 5 draws conclusions and suggests avenues for further research.

2 Literature review

2.1 Customer requirements analysis

Requirements metrics are the foundation of strategic marketing development and product positioning [9], and responding to customer requirements immediately is a critical issue in recommendation system [3 , 11]. In [3], Dadouchi et al. applied the recommendation system to consider actual inventory levels to predict the most accurate customer requirements. In [10], Song et al. developed preference indicators to obtain preference categories by analyzing eye tracking and social behaviour. In [11], Lee et al. analysed and predicted consumers’ opinions of products or services through sentiment analysis and text analysis. Additionally, customer requirements are closely related to customer satisfaction. In [12], Mkpojiogu et al. revealed the potential relationship between the customer satisfaction index of the Kano model and the importance of customer demand. However, the limitation of Kano model is that it is a qualitative study and lacks a quantitative process. Thus, customer requirements are closely related to customer satisfaction, and the importance of these requirements can be determined from the perspective of quantitative customer satisfaction.

Traditional methods for identifying customer requirements include interviews and questionnaires. These methods are subjective and lack the ability to integrally capture customer requirements [13]. The Internet provides a crucial resource for acquiring requirements as a bridge between customers and enterprises. Customers express their opinions on products and personal requirements through the Internet. Enterprises use the Internet to understand customer requirements to update products and provide services. With the rapid development of the Internet, e-commerce generates rich transaction data which allow enterprises to explore customer behaviour, habits, preferences, and characteristics and can help enterprises understand customer requirements, which have tremendous commercial value [14]. These data include product purchases, characteristics of hot products, online reviews, and so on.

2.2 Online reviews and text mining

Online reviews are an important data source for investigating customer requirements and preferences [15]. Customers’ real requirements and preferences are often hidden in online reviews in the form of natural language [16], they can serve as a vital source of information for companies by helping them precisely understand customer requirements. Text data mining techniques are widely employed as the number and dimensions of reviews grow exponentially and the complexity of reviews continues to increase [28].

Text data mining technology is a variant of data mining technology related to text sources which is characterized by transforming unstructured text document data into a formal form to extract valuable information [17]. In [18], Sun et al. summarized the following primary process of text mining: information retrieval, information extraction, knowledge discovery and knowledge application. In [27], text topics are extracted for text classification and application recommendation based on the clustering model of latent association rules. Customers’ personal opinions are often expressed in the form of words. The core of text mining analysis involves transforming text documents into forms that facilitate analysis and proper calculation [17]. Supervised and unsupervised methods can be used to handle unstructured text data from online reviews [6]. For supervised methods, problems of laborious manual marking and the costs of training models need to be addressed. Unsupervised methods are relatively less cumbersome, but the accuracy needs to be improved.

2.3 Sentiment analysis

Online reviews are often considered to be vehicles for expressing customer sentiment, and, thus, sentiment analysis of reviews is essential [19]. In [20], Liu et al. described a process of searching for topics and product opinions from reviews called opinion mining. Sentiment analysis can include sentiment polarity and sentiment degree analysis, most studies focus on sentiment polarity analysis [6], but few quantitative studies focus on sentiment degree analysis. In addition, the factors that reflect customer requirements are not just emotional factors, such as mobile phone, the customer hierarchy, number of sub-reviews, review times, and number of supporters, can reflect customer requirements to some extent.

Motivated by this discussion, this study contributes to the literature in the following ways:

It defines KCRs calculates utility values of customers’ satisfactory or dissatisfactory and KRI for product engineering characteristics based on online review data;

It combines sentiment analysis and a comprehensive consideration of various factors, especially considering the new word lexicon of the products mentioned in the reviews as one of the vital factors, and it analyses customer satisfaction and dissatisfaction with the product engineering characteristics to identify KCRs based on Kano model.

3 Methodology

3.1 Framework

This study builds a key requirements identification model to calculate the utility values of customers’ satisfaction (SU), dissatisfactory utility values (DSU) and KRI for engineering characteristics when a certain product engineering characteristic is available based on online reviews to identify KCRs. The framework of the full text analysis is presented in Fig. 1. The analysis is implemented in two phases. Phase 1 is the acquisition of customer requirements, including the acquisition and pre-processing of online review data and the extraction of product engineering characteristics. Phase 2 involves the sentiment analysis and the quantification of the utility values of other factors that affect customer satisfaction. These utility values are combined with the product engineering characteristics to determine the KRI and identify KCRs.

Fig. 1

A framework of the methodology.

3.2 Data acquisition and pre-processing

Python is utilized to write regular expressions so as to run a web crawler and effectively obtain online review text. The crawling content includes the text of reviews, user levels, sub-reviews, the number of supporters, and the review time. Reviews can be represented as a set P = {P₁, P₂, …, P_i, …, P_m}, where P_i denotes the i-th review.

Text pre-processing a critical factor affecting the accuracy and efficiency of data mining. This study uses the Natural Language Processing & Information Retrieval Sharing Platform-Institute of Computing Technology, Chinese Lexical Analysis System (NLPIR-ICTCLAS), a Chinese lexical analysis system for text processing. Text processing includes text denoising, sentence segmentation, text segmentation, part-of-speech tagging, and de-stopping words.

3.3 Extraction of product engineering characteristics

In this section, the emotional lexicon, degree lexicon, negative vocabulary, and new word lexicon are constructed using the results of the word segmentation obtained in the last section, and then the latent Dirichlet allocation (LDA) topic model is used to extract the product engineering characteristics.

Customers usually express their requirements for products’ engineering characteristics when they review them. Thus, the KCRs can be identified by extracting engineering characteristics to analyse each customer’s satisfaction or dissatisfaction with the product engineering characteristics.

The flow of product design is generally from customer requirements, to functional requirements, to engineering characteristics [21]. So, the functional requirements are determined firstly to better analyze and distinguish the product’s engineering characteristics, and the functional requirements of the product can be extracted from customer requirements. To illustrate the relationship between functional requirements and engineering characteristics, an example of a smartphone is shown in Fig. 2.

Fig. 2

Relationship of functional requirements and engineering characteristics.

If a product contains n functional requirements, then the functional requirements can be represented as a set S = {S₁, S₂, …, S_j, …, S_n}, where S_j denotes the j-th functional requirement. If each functional requirement contains θ_j engineering characteristics, θ_j = 1, 2, …, l, then the engineering characteristics can be represented as a set F = {S_1,1, S_1,2, … , S_{1, θ1}, … , S_j,1, S_j,2, … S_{j,k_j}, …, S_{j,θ_j}, …, S_{n,θ_n}}, where S_{j,k_j} denotes the k_j-th engineering characteristic of the j-th functional requirement, k_j = 1, 2, …, θ_j. Among all the parameters, we suppose that i = 1,2, ..., m; j = 1,2, ..., n.

Then, this study uses the LDA topic clustering model to extract the engineering characteristics. The LDA model was proposed by [22] to determine the implied subject of a text by automatically constructing and processing text characteristics, which is an internal semantic knowledge extraction model that is widely applied to text mining tasks, such as retrieval, summarization, and clustering in recent years [23]. Compared to previous text analysis methods, LDA can perform many text analysis steps without manual intervention and is more suitable for processing unstructured online comment text [24]. It is an unsupervised machine learning technique, also known as a three-layer Bayesian probability model, consisting of a three-tiered structure of words, topics, and documents. Each document appearing in the model has hidden topics and also includes all characteristics that are reduced to hidden topics [25]. Thus, the LDA method not only identifies the topic of a text but also selects characteristics. This study applies this automatic training technology to determine the product characteristics, and the vocabulary-theme-document three-layer structure is regarded as the three-layer framework of engineering characteristics-theme-reviews to extract engineering characteristics.

3.4 Calculation of the satisfactory or dissatisfactory utility values of engineering characteristics

3.4.1 Construction of the thesaurus

The thesaurus is constructed with segmented text data, the degree lexicon is collected using all degree adverbs in the text, and the negative lexicon is constructed in the same way. These two lexicons are used to calculate the utility value of sentiment for engineering characteristics; select nouns, adjectives, and verbs are compared with the emotional words in the HowNet sentiment dictionary one by one to judge emotional polarity. Finally, some special characteristic terms to describe specific products are summarized and called the new word lexicon, and the function of discovering new words in ICTCLAS is applied to construct the new word lexicon.

3.4.2 Sentiment analysis

Sentiment analysis includes emotional polarity and emotional intensity analyses. Customers often express satisfaction or dissatisfaction with a product using subjective personal emotions. The product engineering characteristics of non-inductive products are not usually mentioned, so this study divides reviews into positive and negative categories. For the emotional intensity analysis, the influence of emotional polarity, degree adverbs, and negative words are analysed. The process of quantifying emotional intensity is presented as follows:

The utility value of emotional polarity is defined as SE_{j,k_j}. Positive emotions are assigned a value of 1, and negative emotions are assigned a value of – 1. Then, the constructed sentiment lexicon is compared with the positive and negative sentiment words in the HowNet dictionary downloaded from the network’s official website by adopting an ergodic function to distinguish positive and negative sentiment words in the sentiment lexicon. These words are assigned values of +1 and – 1, respectively as given by Equation (1), to achieve an emotional polarity classification of the reviews. ${SE}_{{j, k}_{j}} = {\begin{matrix} 1, positive emotional polarity \\ - 1, negative emotional polarity \end{matrix}$ (1)

The intensity of emotions is affected by degree adverbs. For instance, in the review ‘The camera design is a bit ugly, too expensive’, ‘too’ has a greater effect than ‘a bit’ has the emotional intensity. Thus, degree adverbs are quantified to indicate emotional intensity and are defined as λ_{j,k_j}. Based on the work of [7], the degree word lexicon is defined as λ and is divided into 6 levels, as shown in Fig. 3.

Fig. 3

Quantitative value of degree words.

The effect of negative words on emotional polarity is defined as μ_i as presented in Equation (2): $μ_{i} = {\begin{matrix} - 1, The number of negative words is odd \\ 1, The number of negative words is even \end{matrix}$ (2)

Whether the reviews involve an engineering characteristic is defined as δ_{ij,k_j} as presented in Equation (3): $δ_{{ij, k}_{j}} = {\begin{matrix} 1, \exists S_{{j, k}_{j}} \in P_{i} \\ 0, \forall S_{{j, k}_{j}} \notin P_{i} \end{matrix}$ (3)

The customer’s utility value of sentiment for engineering characteristics can be quantified based on the above parameters.

When SE_{j,k_j} * μ_i = 1, the utility value of positive sentiment for engineering characteristic S_{j,k_j} is defined as $E_{{j, k}_{j}}^{+}$ , as shown in Equation (4): $E_{{j, k}_{j}}^{+} = \sum_{i = 1}^{m} λ_{{j, k}_{j}} δ_{{ij, k}_{j}}$ (4)

When SE_{j,k_j} * μ_i = - 1, the utility value of negative sentiment for engineering characteristic S_{j,k_j} is defined as $E_{{j, k}_{j}}^{-}$ , as shown in Equation (5): $E_{{j, k}_{j}}^{-} = \sum_{i = 1}^{m} - λ_{{j, k}_{j}} δ_{{ij, k}_{j}}$ (5)

3.4.3 Calculation of the utility values of other factors Customer level

The customer’s credit rating is reflected by the customer’s level of online reviews, and a higher credit rating has a higher reference value for a review. Thus, the utility value of the customer level in each review is defined as F_i, where f denotes the customer level. The equation is given by Equation (6): $F_{i} = \frac{\max (f) + 1 - f}{\sum f}$ (6)

Number of supports

The more supports a review has, the greater utility value of the customers’ opinions on the engineering characteristics are. Supporters can be defined as Q_i, where q_i denotes the number of supporters of P_i and q_max denotes the maximum number of supporters, and 1≤Q_i≤2 as shown in Equation (7): $Q_{i} = 1 + \frac{{\ln (q}_{i} + 1)}{{\ln (q}_{\max} + 1)}$ (7)

Number of sub-reviews

Similarly, the number of sub-reviews can also explain the problem reflected in number of supporters. The number of sub-reviews is represented by R_i, where r_i denotes the number of sub-reviews of P_i, and 1≤R_i<2 as shown in Equation (8): $R_{i} = 1 + \frac{r_{i}}{\sum r_{i}}$ (8)

Review time

The release and sale of products occurs over a life cycle, and the customer requirements hierarchy also goes through a life cycle. Thus, reviews in the current period reflect the current level of customer requirements more than reviews from three months ago do. Thus, a parameter that indicates the extent of the impact on customer requirements owing to the review time can be defined as T_i, where t_i denotes the review time, the value of T_i is given by Equation (9): $T_{i} = 1 + \frac{t_{i}}{\sum t_{i}}$ (9)

New word lexicon

When new products are released, some new engineering characteristics are introduced to attract customers, in addition to the general function. These words may not appear frequently, but they reflect the uniqueness of the product. The preference for strong expressions is more evident and more reflective of customer requirements. For example, in mobile phone reviews, ‘photographing’ is not a new word, but ‘980 processor’ is a new word, a term unique to mobile phones. Thus, the weight of this new word being included in reviews can be defined as N_{j,k_j}, which takes a larger value of 2 as shown in Equation (10) when the engineering property S_{j,k_j} contains a new word, and N represents the new word lexicon: $N_{{j, k}_{j}} = {\begin{matrix} 2, \exists S_{{j, k}_{j}} \in N \\ 1, \forall S_{{j, k}_{j}} \notin N \end{matrix}$ (10)

In summary, the other factors proposed in this section are 5 attributes of online reviews and weighted to obtain a total utility value of U_{ij,k_j} as shown in Equation (11): $U_{{ij, k}_{j}} {= Fi + Qi + Ri + Ti + N}_{{j, k}_{j}}$ (11)

3.5 Identification model of KCRs

In this section, CA is combined to calculate the values of SU, DSU, and KRI of each product engineering characteristic. The calculation process for the utility value of the product engineering characteristics needs to combine multiple factors. Thus, this study adopts the part-worth model of CA to combine the various indicators determined in Section 3.4. The corresponding values are calculated by Equations (12 –14): ${SU}_{{j, k}_{j}} = \sum_{i = 1}^{m} E_{{j, k}_{j}}^{+} U_{{ij, k}_{j}}$ (12) ${DSU}_{{j, k}_{j}} = \sum_{i = 1}^{m} E_{{j, k}_{j}}^{-} U_{{ij, k}_{j}}$ (13)

The specific process of identifying KCRs is shown in Fig. 4. As shown in Fig. 4, the case of SU_{j,k_j} = 0 and DSU_{j,k_j} < 0 indicates a sort of reverse requirement if the product has been deployed with the engineering characteristics. If not, it expresses a customer’s expectation. This kind of expected requirement can be treated as a kind of KCRs, and the corresponding engineering characteristic can be appended when making product updates.

Fig. 4

Identification of KCRs.

When SU_{j,k_j} > 0 and DSU_{j,k_j} = 0, it follows that the customer is relatively satisfied with the engineering characteristics, and the reflected customer requirements may be expected requirements or attractive requirements. If the utility value of sentiment is greater than or equal to 1.4 and contains a new word N_{j,k_j} = 2), the customer requirement reflected by the engineering characteristics is regarded as a kind of KCRs.

The case of SU_{j,k_j} = 0 and DSU_{j,k_j} = 0 indicates that the engineering characteristic reflects an indifferent requirement.

Finally, when SU_{j,k_j} ≠ 0 and DSU_{j,k_j} ≠ 0, the two values are integrated to obtain the KRI_{jk
_j}, as follows: ${KRI}_{{j, k}_{j}} = | \frac{{SU}_{{j, k}_{j}}}{{DSU}_{{j, k}_{j}}} |$ (14)

When KRI_{j,k_j}≤1, the requirement is considered a biased reverse requirement or a basic requirement, and when KRI_{j,k_j}>1, SU_{j,k_j}> DSU_{j,k_j}, indicating that customers are more satisfied with the corresponding engineering characteristics. The top 50% of all values of KRI_{j,k_j} > 1 are attractive requirements, and the corresponding engineering characteristics reflected expected requirements. If the utility value of sentiment is greater than or equal to 1.4 and contains a new word (N_{j,k_j} = 2), the customer requirements reflected by the engineering characteristics are regarded as KCRs.

The steps of the model for identifying KCRs are summarized as follows:

– Determine the utility value of sentiment for engineering characteristics $E_{{j, k}_{j}}^{+}$ and $E_{{j, k}_{j}}^{-}$ according to emotional intensity value λ_{jk
_j} and Equations (1 –3), $E_{{j, k}_{j}}^{+}$ is calculated according to Equation (4), $E_{{j, k}_{j}}^{-}$ is calculated according to Equation (5);

The utility values of the other factors (F_i, Q_i, R_i, T_i, N_{j,k_j}) are calculated according to Equations (6 –10), and the total utility value U_{ij,k_j} is calculated according to Equation (11);

The satisfactory utility value SU_{j,k_j} and the dissatisfactory utility value DSU_{j,k_j} are calculated according to Equations (12-13);

When SU_{j,k_j} ≠ 0 and DSU_{j,k_j} ≠ 0, KRI_{j,k_j} is calculated according to Equation (14);

Based on the quantitative results, KCRs are identified.

4 Case study

4.1 Review acquisition and processing

Reviews from the Zhongguancun online website (http://mobile.zol.com.cn/) are selected as the data source for this case study, 834 online reviews are crawled, 800 remained after text denoising, and 8,230 sub-reviews were identified after segmentation. Some of the results are shown in Table 1. In addition, the ICTCLAS system was adopted to perform word segmentation, part-of-speech tagging, and word de-stopping.

Table 1
Result of segmentation of partial reviews

Number of reviews Customer level Review content Review time Number of supporters Number of sub-reviews

P1 Level 1 Use the Kirin 980 processor with 7 nm process 8 curved 3D glass body Looks very good The price is as high as ever Portrait shot is very great 17-Oct-18 1104 38

... ... ... ... ... ...

Number of reviews	Customer level	Review content	Review time	Number of supporters	Number of sub-reviews
P1	Level 1	Use the Kirin 980 processor with 7 nm process 8 curved 3D glass body Looks very good The price is as high as ever Portrait shot is very great	17-Oct-18	1104	38
...	...	...	...	...	...

4.2 Extraction of engineering characteristics from smartphone online reviews

After multiple parameter debugging of the LDA topic model, the best results obtained are that the number of topics is 10, the number of words is 15, and the number of iterations is 200, as is shown Fig. 5. Thus, the functional requirements identified are hardware, photography, camera, speed, battery, screen, appearance, operation, price, and others, for a total of ten. They can be written as S ={ S₁, S₂, S₃, S₄, S₅, S₆, S₇, S₈, S₉, S₁₀ }, and the product engineering characteristics included in each functional requirement are shown in Fig. 6.

Fig. 5

Results of LDA topic clustering model.

Fig. 6

Classification of functional requirements and engineering characteristics.

4.3 Identification of KCRs

The process of calculating the utility values of the engineering characteristics according to the proposed method in Section 3 is mainly divided into two steps: calculating the utility value of sentiment for engineering characteristic $E_{{j, k}_{j}}^{+}$ ( $E_{{j, k}_{j}}^{-}$ ) and calculating the utility values of other factors U_{ij,k_j}.

4.3.1 Calculation of the utility value of sentiment Thesaurus construction

The degree adverbs are filtered from the reviews after part-of-speech tagging. And the degree lexicon that indicating emotional intensity is presented in Fig. 3, the lexicon of negative words no, not only, not just, never, little, few, neither, hardly, seldom, and the lexicon of new words N and the lexicon of sentiment is applied in the classification of emotional polarity in Section 4.3.2 and it is not listed here for space reasons.

Calculation of the utility value of sentiment

Firstly, the utility values of emotional polarity of the reviews are determined by Equation (1), then count parameters λ_{j,k_j}, μ_i and δ_{ij,k_j}, and calculate the utility value of sentiment for engineering characteristics according to the Equations (4-5). The statistical results are shown in Table 2.

Table 2
Utility values of sentiment and other factors

Number of reviews Number of engineering characteristics $E_{{jk}_{j}}^{+} (E_{{jk}_{j}}^{-})$ F_i Q_i R_i N_jk T_i

P₁ S_1,1 1 1/3 1.068 1.168 2 1.0476

S_8,1 1 2

S_8,2 1.6 1

S_9,1 – 1.4 1

P₂ S_2,1 1.6 1/3 1.056 1.035 1 1.0952

S_2,2 1 1

S_3,1 1.4 1

S_2,3 1 1

S_10,1 1.4 2

S_3,2 #– 1 1

... ... ... ... ... ... ... ...

P₂₀₀ S_1,2 1 1/5 1 1 1 1.0952

Number of reviews	Number of engineering characteristics	$E_{{jk}_{j}}^{+} (E_{{jk}_{j}}^{-})$	F_i	Q_i	R_i	N_jk	T_i
P₁	S_1,1	1	1/3	1.068	1.168	2	1.0476
	S_8,1	1				2
	S_8,2	1.6				1
	S_9,1	– 1.4				1
P₂	S_2,1	1.6	1/3	1.056	1.035	1	1.0952
	S_2,2	1				1
	S_3,1	1.4				1
	S_2,3	1				1
	S_10,1	1.4				2
	S_3,2	#– 1				1
...	...	...	...	...	...	...	...
P₂₀₀	S_1,2	1	1/5	1	1	1	1.0952

4.3.2 Calculation of the utility values of other factors

In online reviews, the customer rating ranges from 1 to 5, so the values of Fi are 1/15, 2/15, 1/5, 4/15, and 1/3; the total number of reviewers is 16,350. The total number of sub-reviews is 2170. Q_i and R_i can be obtained by Equations (7-8), and the value of N_{j,k_j} can be obtained according to the new thesaurus N established in Section 4.3.1. The publication period of the reviews is from November 2018 to April 2019. Thus, the values of T_i are 1/21, 2/21, 1/7, 4/21, 5/21, 2/7, 1/3, and the sum of each of these and 1, as presented in Table 2.

4.3.3 Identification of KCRs

The requirement types are confirmed by calculating the values of SU_{j,k_j}, DSU_{j,k_j}, and KRI_{j,k_j} as reported as Tables 3-1, 3-2, and 3-3.

Table 3-1
Statistical result of SU_{j,k_j} = 0, DSU_{j,k_j} < 0

Characteristics DSU _{j,k_j} SU _{j,k_j}

S_3,3 – 1.9077 0

S_3,4 – 0.9319 0

S_6,5 – 1.3647 0

S_10,4 – 0.9529 0

S_10,8 – 0.9457 0

S_10,9 – 0.941 0

Characteristics	DSU _{j,k_j}	SU _{j,k_j}
S_3,3	– 1.9077	0
S_3,4	– 0.9319	0
S_6,5	– 1.3647	0
S_10,4	– 0.9529	0
S_10,8	– 0.9457	0
S_10,9	– 0.941	0

Table 3-2

Statistical result of SU_{j,k_j} > 0, DSU_{j,k_j} = 0

Characteristics	SU _{j,k_j}	DSU _{j,k_j}	Characteristics	SU _{j,k_j}	DSU _{j,k_j}	Characteristics	SU _{j,k_j}	DSU _{j,k_j}
S_1,1	31.0096	0	S_4,3	2.6116	0	S_8,3	0.9647	0
S_1,2	24.8161	0	S_4,4	1.8427	0	S_8,4	8.3327	0
S_1,3	10.85	0	S_5,2	17.1145	0	S_8,6	17.9514	0
S_1,5	1.2854	0	S_5,3	4.0525	0	S_10,2	1.9188	0
S_1,6	0.9228	0	S_5,4	12.9247	0	S_10,3	2.1715	0
S_2,1	4.03	0	S_6,1	0.9592	0	S_10,5	1.354	0
S_2,3	11.996	0	S_6,6	7.9366	0	S_10,6	2.6916	0
S_2,4	2.7132	0	S_6,7	3.7256	0	S_10,7	3.3345	0
S_2,6	4.0415	0	S_6,9	5.7292	0	S_10,10	2.6613	0
S_2,7	1.9188	0	S_7,3	10.368	0	S_10,12	1.3348	0
S_4,1	31.6961	0	S_7,4	6.6268	0	S_10,13	0.9348	0
S_4,2	2.87	0	S_8,1	10.1806	0	S_10,14	2.2831	0

Table 3-3

Statistical result of SU_{j,k_j}, DSU_{j,k_j} and KRI_{j,k_j}

Characteristics	SU _{j,k_j}	DSU _{j,k_j}	KRI _{j,k_j}	Characteristics	SU _{j,k_j}	DSU _{j,k_j}	KRI _{j,k_j}
S_1,4	4.154	– 0.9391	4.4234	S_6,10	1.8655	– 3.9976	0.4667
S_2,2	43.9174	– 0.9438	46.5325	S_7,1	0.9852	– 3.0116	0.3271
S_2,5	12.6953	– 4.7414	2.6775	S_7,2	47.193	– 4.565	10.338
S_3,1	4.9231	– 10.4458	0.4713	S_8,2	6.0717	– 6.1768	0.983
S_3,2	2.6861	– 0.7685	3.4953	S_8,5	2.7188	– 3.2075	0.8476
S_5,1	24.8495	– 6.2884	3.9516	S_8,7	2.7601	– 0.9365	2.9473
S_6,2	9.886	– 4.526	2.1843	S_9,1	11.7855	– 11.8233	0.9968
S_6,3	5.917	– 7.2796	0.8128	S_9,2	7.3466	– 0.9438	7.7841
S_6,4	5.4631	– 0.9486	5.7591	S_10,1	6.4492	– 3.6084	1.7873
S_6,8	1.8725	– 5.376	0.3483	S_10,11	0.9248	– 0.9352	0.9889

In Table 3-1, SU_{jk
_j} = 0 and DSU_{jk
_j} < 0 for S_3,3, S_3,4, S_6,5, S_10,4, S_10,8, S_10,9. The corresponding engineering characteristics are the lifting camera and the front camera, a screen with eye protection, software adaptation, and film quality. This result indicates that customers are not satisfied with the mobile phone’s lifting camera, front camera, software adaptation, and film quality, as customers expressed reverse requirements for these engineering characteristics. This phone does not have a screen with eye protection. Expressed as a expected requirement of customers, it can be considered a kind of KCRs.

In Table 3-2, among the values that satisfy SU_{jk
_j} >0 and DSU_{jk
_j} = 0, the corresponding engineering characteristics for which the mean utility value of sentiment is greater than or equal to 1.4 that contain a new word (N_{j,k_j} = 2) are wide-angle shooting, phone icons, processor, emerald green, extreme frame rate, and 5G network. This result indicates that customers are very concerned about these engineering characteristics, and the corresponding requirements are expected and even attractive requirements. These engineering characteristics are also unique functions for mobile phone products, and thus, the corresponding requirements are regarded as KCRs. For instance, the phone icon is an engineering characteristic with less attention to the mobile phone. The customer does not mind an ordinary phone icon design, but a novel and distinctive design brings greater customer satisfaction. Thus, product updates can be realized without solving hard technical requirements. The 5G network is also worthy of attention for both enterprises and customers.

In Table 3-3, similarly, among the characteristics with KRI_{jk
_j}>1, the top 50% are S_1,4, S_2,2, S_9,2, S_6,4, S_7,2, and S_5,1. The corresponding engineering characteristics with mean utility value of sentiment greater than or equal to 1.4 that contain a new word (N_{j,k_j} = 2) are wired fast-charging, shooting effects, and sound quality, indicating that fast-charging mobile phones can bring greater satisfaction to customers in this fast-paced era. At the same time, for shooting effects and sound quality, customers want better experiences, which provides a direction for product updates. Thus, the corresponding customer requirements are treated as KCRs.

5 Discussion

To better verify the validity and innovation of the proposed KCRs identification model, compared our model with models of mining customer requirements proposed in the previous studies [7, 26], and applied them to the data in this study.

In [7], it applies weighted summation of sentiment polarity value and sorts by the mean, the requirements corresponding to the sentiment polarity mean of engineering characteristics ≥2 are regarded as attractive requirements, otherwise regarded as the expected requirements.

Table 4 is the result of applying the model of requirement mining in [7]. As shown in Table 4, the sentiment polarity mean of S_1,1, S_2,1, S_2,2, S_4,1, S_7,2, and S_10,5 is ≥2, The corresponding engineering characteristics are processor, portrait shooting, shooting effect, running speed, appearance, and dual SIM, represents that the corresponding requirements are attractive requirements, and the rest belongs to expected requirements. Compared with this study, although this model emphasizes the importance of attractive requirements and expected requirements, the calculation process and results are too simplistic, and the requirements hierarchy is clearly unfair to those engineering characteristics with a large number of reviews. It cannot identify the KCRs, that are more vital to product updates. For example, portrait shooting and dual SIM are not the most novel requirements for the customer; although customers care about them, they are not critical to product updates.

Table 4
Utility value of sentiment and other factors of model in [7]

Engineering characteristics Sentiment polarity mean Engineering characteristics Sentiment polarity mean Engineering characteristics Sentiment polarity mean Engineering characteristics Sentiment polarity mean

S_1,1 2.1132 S_3,4 – 1.0000 S_6,8 0.4951 S_10,1 0.3202

S_1,2 1.0230 S_4,1 2.2284 S_6,9 1.3009 S_10,2 1.0266

S_1,3 1.2038 S_4,2 1.0457 S_6,10 – 0.3733 S_10,3 2.0089

S_1,4 0.6080 S_4,3 1.5066 S_7,1 – 0.4602 S_10,4 – 1.0089

S_1,5 1.5000 S_4,4 1.5000 S_7,2 2.0285 S_10,5 1.0266

S_1,6 1.0000 S_5,1 0.8294 S_7,3 1.2522 S_10,6 1.0243

S_2,1 2.1006 S_5,2 0.9805 S_7,4 0.6667 S_10,7 1.3333

S_2,2 1.2220 S_5,3 1.0236 S_8,1 1.1131 S_10,8 – 1.0000

S_2,3 1.2060 S_5,4 1.0844 S_8,2 0.0055 S_10,9 – 1.0000

S_2,4 1,0290 S_6,1 1.0310 S_8,3 1.0133 S_10,10 1.0000

S_2,5 0,5022 S_6,2 0.5000 S_8,4 1.3739 S_10,11 0.0022

S_2,6 1.0148 S_6,3 – 0.1217 S_8,5 – 0.1117 S_10,12 1.0266

S_2,7 1.0266 S_6,4 0.6696 S_8,6 1.2824 S_10,13 1.0266

S_3,1 0.3579 S_6,5 1.0133 S_8,7 0.5000 S_10,14 1.0221

S_3,2 0.6637 S_6,6 1.4425 S_9,1 – 0.0161

S_3,3 1.5600 S_6,7 1.0077 S_9,2 0.8822

Engineering characteristics	Sentiment polarity mean	Engineering characteristics	Sentiment polarity mean	Engineering characteristics	Sentiment polarity mean	Engineering characteristics	Sentiment polarity mean
S_1,1	2.1132	S_3,4	– 1.0000	S_6,8	0.4951	S_10,1	0.3202
S_1,2	1.0230	S_4,1	2.2284	S_6,9	1.3009	S_10,2	1.0266
S_1,3	1.2038	S_4,2	1.0457	S_6,10	– 0.3733	S_10,3	2.0089
S_1,4	0.6080	S_4,3	1.5066	S_7,1	– 0.4602	S_10,4	– 1.0089
S_1,5	1.5000	S_4,4	1.5000	S_7,2	2.0285	S_10,5	1.0266
S_1,6	1.0000	S_5,1	0.8294	S_7,3	1.2522	S_10,6	1.0243
S_2,1	2.1006	S_5,2	0.9805	S_7,4	0.6667	S_10,7	1.3333
S_2,2	1.2220	S_5,3	1.0236	S_8,1	1.1131	S_10,8	– 1.0000
S_2,3	1.2060	S_5,4	1.0844	S_8,2	0.0055	S_10,9	– 1.0000
S_2,4	1,0290	S_6,1	1.0310	S_8,3	1.0133	S_10,10	1.0000
S_2,5	0,5022	S_6,2	0.5000	S_8,4	1.3739	S_10,11	0.0022
S_2,6	1.0148	S_6,3	– 0.1217	S_8,5	– 0.1117	S_10,12	1.0266
S_2,7	1.0266	S_6,4	0.6696	S_8,6	1.2824	S_10,13	1.0266
S_3,1	0.3579	S_6,5	1.0133	S_8,7	0.5000	S_10,14	1.0221
S_3,2	0.6637	S_6,6	1.4425	S_9,1	– 0.0161
S_3,3	1.5600	S_6,7	1.0077	S_9,2	0.8822

In [26], a model for identifying customer preferences based on online reviews is proposed to discover the engineering characteristics that need to be improved. By analyzing the importance and the satisfaction of engineering characteristics, an importance-satisfaction matrix is built, the customer’s preference for engineering characteristics is divided into 4 categories: high-satisfaction and high-importance, low-satisfaction and high-importance, low-satisfaction and low-importance, and high-satisfaction and low-importance. The importance-satisfaction matrix of this study is presented in Fig. 7. As show in Fig. 7, the engineering characteristics that are related to the same functional requirements are generally grouped in the same category, and the importance is sorted according to the frequency in online reviews. The model in this study aims to find the customer requirements that would greatly improve customer satisfaction and promote product updates, even if the frequency is not high. Compared with this study, although this model simultaneously considers the importance and satisfaction of engineering characteristics, there are only two levels of satisfaction, and without the analysis of emotional intensity, and other factors.

Fig. 7

Importance-satisfaction matrix of model in [26].

In summary, Table 5 shows the analysis of the comparison between this article and [7] and [26].

Table 5

Analysis of the comparison between this article and previous studies

	Model of [7]	Model of [26]	Model of this study
Tools of data collection	Train collector	Octopus	Write web crawler with python to crawl online reviews
Collecting content	Review content, user name, number of sub-reviews	Review content	Review content, customer level, number of sub-reviews, number of supporters and review time
Characteristic extraction	Build domain ontology to extract attribute-sentiment words pair	Summarize and classify the keywords	Extract engineering characteristics based on LDA model
Sentiment analysis	4 levels	2 levels	6 levels
Analysis of other factors	number of sub-reviews	no	customer level, number of sub-reviews, number of supporters, review time and new words
Identification of requirements	Apply the mean value of emotional polarity as the gist of dividing requirements hierarchy	Establish an importance-satisfaction matrix by integrating quantitative studies of multiple linear regression and meta-analysis	the KCRs are identified based on the quantitative results and related rules

As shown in Table 5, Compared with the previous model, the model of this study has the following technical and logical advantages:

Firstly, previous studies lack deep mining, as they regard customers as being satisfied with an engineering characteristic when the emotional polarity value is greater than zero; otherwise, customers are regarded as not satisfied. This study offers further analysis. If the product does not have engineering characteristics with emotional polarity values below zero, it can be inferred that the customer was expressing an expectation, and it can be considered as a kind of KCRs and include it among the engineering characteristics for product updates.

In addition, the previous study considered fewer factors that affect customer requirements, which is much more general. The quantification process in this study is more detailed, and more emotional levels, quantifying multiple factors that affect customer requirements, particularly the establishment of a new word lexicon as an essential factor for identification of KCRs.

Finally, though previous studies have identified expected and attractive requirements, or requirements with high-satisfaction and high-importance, the results may be straightforward and not critical KCRs for product updates. This study determines the values of KRI of engineering characteristics by calculating all types of values that reflect the utility of customer requirements and formulating rules for KCRs, which more specifically with product updates.

6 Conclusions and future works

This study takes product updates as a starting point. Based upon relevant online reviews, this article applies text data mining technology to build the identification model, compiles previous research to summarize and develop corresponding rules for KCRs, and aims to identify KCRs. And it is vital to quantify various factors for identifying key requirements. The web crawlers, LDA topic clustering models, and the CA method were introduced to accurately identify KCRs.

Specifically, this article contributes to the literature from the following two aspects.

The text mining and sentiment analysis are integrated to establish the quantification of the customer satisfaction or dissatisfaction utility for engineering characteristics, in which 5 attributes of online reviews that are customer level, number of supports, number of sub-reviews, review time and new word lexicon are simultaneously considered. These structural analysis means can effectively reduce the subjectivity and uncertainty of the traditional customer satisfaction quantification method.

To identify the KCRs, the KRI is defined concurrently considered the SU and DSU which reflects the degree of positive and negative evaluation of customers, respectively. The integration of SU and DSU can overcome the drawback that only takes consideration of the positive preference of customers in the literatures.

The future works for product updates will focus on the evolution of customer requirements and the updating of KCRs considering the periodicity of customer requirements.

References

Kano

, Attractive quality and must-be quality, The Journal of Japanese Society for Quality Control 14(1984), 39–48.

, Li

and Sun

, Hybrid kano-fuzzy-dematel model based risk factor evaluation and ranking of cross-border e-commerce smes with customer requirement, Journal of Intelligent and Fuzzy Systems 37(11) (2019), 1–17.

Dadouchi

and Agard

, Lowering penalties related to stock-outs by shifting demand in product recommendation systems, Decision Support Systems 114(2018), 61–69.

Murray

P.W.

, Agard

and Barajas

M.A.

, Forecast of individual customer’s demand from a large and noisy dataset, Computers & Industrial Engineering 118 (2018), 33–43.

Potra

, Pugna

and Negrea

, Customer perspective of value for innovative products and services, Procedia-Social and Behavioral Sciences 238 (2018), 207–213.

Zhou

, Jiao

R.J.

and Linsey

J.S.

, Latent customer needs elicitation by use case analogical reasoning from sentiment analysis of online product reviews, Journal of Mechanical Design 137(7) (2015), 071401.

H.L.

, Tang

X.B.

and Xie

, Research on User Demand Mining Model Based on Online Comments, Journal of the China Society for Scientific and Technical Information 34 (2015), 1097.

Tung

T.C.

and Chen

H.Y.

, Integrating Conjoint Analysis with TOPSIS Algorithm to the Visual Effect of Icon Design Based on Multiple Users’ Image Perceptions, EURASIA Journal of Mathematics Science & Technology Education 13(3) (2017), 1025–1040.

Yang

, Li

Y.L.

and Chin

K.S.

, An ordinal scale-based GDM approach to prioritize customer requirements in QFD product planning, Journal of Intelligent and Fuzzy Systems 37(3) (2019), 1–19.

10.

Song

and Moon

, Eye-tracking and social behavior preference-based recommendation system, Journal of Supercomputing 75(4) (2019), 1990–2006.

11.

Lee

S.W.

, Jiang

, Kong

H.Y.

and Liu

, A difference of multimedia consumer’s rating and review through sentiment analysis, Multimedia Tools and Applications 1(2020), 1–18.

12.

Mkpojiogu

E.O.C.

and Hashim

N.L.

, Understanding the relationship between Kano model’s customer satisfaction scores and self-stated requirements importance, Springer Plus 5(1) (2016), 1–22.

13.

Song

W.Y.

, Requirement management for product-service systems: Status review and future trends, Computers in Industry 85 (2017), 11–22.

14.

, Chi

and Hao

, Customer demand analysis of the electronic commerce supply chain using Big Data, Annals of Operations Research 268(1-2) (2018), 113–128.

15.

Que

, Nagai

K.Y.

and Minh

N.L.

, Extracting user requirements from online reviews for product design: A supportive framework for designers, Journal of Intelligent & Fuzzy Systems Preprint 37(3) (2019), 7441–7451.

16.

Jiao

and Yang

, A product configuration approach based on online data, Journal of Intelligent Manufacturing 30 (2018). doi: 10.1007/s10845-018-1406-y

17.

Salloum

S.A.

, Al-Emran

, Monem

A.A.

and Shaalan

, Using Text Mining Techniques for Extracting Information from Research Articles, Intelligent Natural Language Processing: Trends and Applications 1 (2018), 373–397.

18.

Sun

W.C.

, Cai

Z.P.

, Li

Y.Y.

, Liu

, Fang

S.Q.

and Wang

G.Y.

, Data processing and text mining technologies on electronic medical records: a review, Journal of Healthcare Engineering 2018 (2018), 1–9.

19.

H.W.

and Moon

, Design of recommendation system for tourist spot using sentiment analysis based on CNN-LSTM, Journal of Ambient Intelligence and Humanized Computing 1 (2019). doi: 10.1007/s12652-019-01521-w

20.

Liu

, Bi

J.W.

and Fan

Z.P.

, Ranking products through online reviews: A method based on sentiment analysis technique and intuitionistic fuzzy set theory, Information Fusion 36 (2017), 149–161.

21.

Y.L.

, Tang

J.F.

, Chin

K.S.

, Jiang

Y.S.

, Han

and Pu

, Estimating the final priority ratings of engineering characteristics in mature-period product improvement by MDBA and AHP, International Journal of Production Economics 131(2) (2011), 575–586.

22.

Blei

D.M.

, Ng

A.Y.

and Jordan

M.I.

, Latent dirichlet allocation, Journal of machine Learning research 3 (2003), 993–1022.

23.

Kim

Y.J.

, Ban

H.J.

, Kim

D.H.

and Kim

H.S.

, Understanding customer experience of airline lounge using text mining of online review, Culinary ence & Hospitality Research 26(2) (2020), 36–44.

24.

Guo

, Barnes

S.J.

and Jia

, Mining meaning from online ratings and reviews: Tourist satisfaction analysis using latent dirichlet allocation, Tourism Management 59 (2017), 467–483.

25.

Jahromi

M.S.

, Bagheri

, Rostami

and Keshavarz

, Feature Extraction in Fractional Fourier Domain for Classification of Passive Sonar Signals, Journal of VLSI Signal Processing Systems for Signal Image and Video Technology 91(5) (2019), 511–520.

26.

, Lan

Q.L.

, Liu

and Yang

, Integrated online consumer preference mining for product improvement with online reviews, Journal of Systems Science & Information 7 (2019), 17–36.

27.

Fernandes

D.S.F.

, Aurélio

D.M.

, Vaccari

S.C.

, Oliveira

D.C.V.

, Fernanda

M.M.

and Oliveira

R.S.

, Latent association rule cluster-based model to extract topics for classification and recommendation applications, Expert Systems with Application 112 (2018), 34–60.

28.

Wang

, Wan

, Cui

, Liu

and Xu

, Personalized recommendation algorithm based on product reviews, Journal of Electronic Commerce in Organizations 16(3) (2018), 22–38.

Identification of key customer requirements based on online reviews

Abstract

Keywords

1 Introduction

2 Literature review

2.1 Customer requirements analysis

2.2 Online reviews and text mining

2.3 Sentiment analysis

3 Methodology

3.1 Framework

3.3 Extraction of product engineering characteristics

3.4.1 Construction of the thesaurus

3.4.2 Sentiment analysis

4.1 Review acquisition and processing

4.3.1 Calculation of the utility value of sentiment Thesaurus construction

4.3.3 Identification of KCRs

Table 3-1 Statistical result of SUj,k j = 0, DSUj,k j < 0 Characteristics DSU j,k j SU j,k j S3,3 – 1.9077 0 S3,4 – 0.9319 0 S6,5 – 1.3647 0 S10,4 – 0.9529 0 S10,8 – 0.9457 0 S10,9 – 0.941 0

References

Table 3-1
Statistical result of SU_{j,k_j} = 0, DSU_{j,k_j} < 0

Characteristics DSU _{j,k_j} SU _{j,k_j}

S_3,3 – 1.9077 0

S_3,4 – 0.9319 0

S_6,5 – 1.3647 0

S_10,4 – 0.9529 0

S_10,8 – 0.9457 0

S_10,9 – 0.941 0