Demand-driven NEV supplier selection: An integrated method based on ontology–QFD–CBR

1 Introduction

The transportation sector accounts for 24% of global CO₂ emissions, while fuel vehicles are an important source of greenhouse gases and pollutants (Wang et al., 2017; Luo et al., 2021). Compared with the fuel vehicle, the new energy vehicle (NEV) has significantly higher fuel efficiency and lower or even zero CO₂ emissions (Teixeira and Sodre, 2018; Liu et al., 2021). Studies have shown that plug-in hybrid electric vehicles and hybrid electric vehicles can reduce CO₂ emissions by about 30%, while in areas with a high proportion of hydro power, pure electric vehicles can reduce CO₂ emissions by 90% (Wang et al., 2022; Yu et al., 2022). NEVs have the advantage of being green, low carbon, energy saving, and convenient (Cano et al., 2018; Xie et al., 2020; Hao et al., 2021). As an effective solution to face the global challenges of environmental pollution and energy shortages caused by traditional fuel vehicles, it has attracted the attention of governments all over the world (He 2016; Bastida-Molina et al., 2021). Governments have proposed various policies and invested a lot of money to support the promotion and application of NEVs (Adnan et al., 2017; Tan and Lin 2019). For example, the Chinese government has attached importance to the development of NEVs since the 1990s and listed NEVs as one of the strategic emerging industries in 2010 (Gong et al., 2013; Li and Jing, 2019; Tan et al., 2021). With the support of various policies of the Chinese government (such as purchase subsidies, special energy conservation, and emission reduction fund), the territory of the NEV industry has continued to expand, and 28 provinces in China have joined the ranks of NEV manufacturing (Gong et al., 2013; Lu et al., 2017; He et al., 2018; Kendall 2018). In 2020, China added more than 68,000 NEV-related enterprises, an increase of 85% compared with 2019. Meanwhile, NEV sales in China have also grown rapidly. In 2014, NEV sales have quadrupled compared with 2013. In 2020, NEV sales were 18.2 times those of 2014, accounting for 41% of the global NEV sales (Hao et al., 2021; Liu et al., 2021). With the rapid growth of NEVs, the competition in the NEV industry is becoming increasingly fierce, and NEV manufacturers are facing more and more challenges. In this case, it is difficult for NEV manufacturers to survive and develop entirely on their own research and development (R&D) investments and innovation capabilities (Huth et al., 2015; Lu et al., 2020). Therefore, the development of cooperation between NEV manufacturers and NEV component suppliers has become an inevitable trend of development (Fan et al., 2020). However, some procurement departments of NEV manufacturers have not fully introduced the supplier competition mechanism, and there are many problems such as an imperfect supplier evaluation system, strong subjectivity, and lack of consideration of user demands. For this, how to select the appropriate supplier from a large number of NEV component suppliers with varying capabilities and levels of the market has become an urgent problem for NEV manufacturers.

As a starting point in the NEV industry supply chain, the selection of an NEV component supplier directly determines the cost and quality of the whole supply chain and even the benefits of the enterprise (Hosseini and Sarder, 2019). Choosing a suitable NEV component supplier can effectively improve the performance, customer satisfaction, and overall benefits of NEV manufacturers; while the improper selection of a component supplier may lead to the loss of time, cost, and market share of NEV manufacturers (Büyüközkan and Çifçi, 2011; Ayağ and Samanlioglu, 2014; Eydi and Fazli, 2019; Hosseini and Sarder, 2019; Wang et al., 2021). For example, in 2016, due to potential safety hazards in Samsung SDI batteries, JAC had to temporarily suspend production of its pure electric vehicle IEV6S, resulting in JAC’s sales of this vehicle being significantly lower than expected. In 2021, due to the risk of fire in the power battery provided by Funeng Technology, BAIC recalled more than 30,000 pure electric vehicles, EX360 and EU400, and suffered huge economic losses. Thus, for NEV manufacturers, it is very important to choose a suitable component supplier from multiple component suppliers, which is the guarantee of their success (Khan et al., 2016; Liu et al., 2019). However, NEV manufacturers are different from traditional automobile manufacturers in terms of technology, sales service, corporate culture, and marketing mode, and their selection requirements and criteria for component suppliers are also different from those of traditional automobile manufacturers. Compared with the selection of a traditional automobile component supplier, NEV manufacturers pay more attention to the component suppliers’ R&D capabilities and technology, environmental protection capabilities, and supply nodes (Liu 2016; Xiao 2020). Meanwhile, for NEV manufacturers, NEV component suppliers are not only simple suppliers but also mutually beneficial partners, sharing information and coordinated development with them. Given this, it is difficult to apply the component supplier selection method of traditional automobile manufacturers to solve the selection problem of NEV manufacturers. Thus, it is necessary to develop a targeted method to select a suitable component supplier for NEV manufacturers.

At present, the studies on the selection of an NEV component supplier are still scarce, but some related research results can be found. These research results can be roughly divided into two aspects: one is the selection or evaluation criteria of suppliers in the automotive industry; the other is the supplier selection method in the automobile industry. Determining the appropriate criteria is an important part of the supplier selection decision. The selection criteria of suppliers in the automotive industry have been widely concerned by scholars. For example, Dargi et al. (2014) developed a framework to support the selection of semi-assembly component suppliers for automobile manufacturers, and they extracted key criteria for evaluating suppliers. Azizi et al. (2015) introduced the most important criteria and sub-criteria for selecting the best supplier in the automotive industry and used Fuzzy technology and TOPSIS to select suitable suppliers. Li et al. (2015) constructed an evaluation index system of component suppliers and proposed a supplier selection model so as to choose the best combination of component suppliers for automobile manufacturers. Lima-junior and Carpinetti (2016) divided the selection criteria of automobile clutch suppliers into priority level, critical level, complementary level, and costly level based on the method of fuzzy quality function deployment and information acquisition difficulty evaluation procedures so as to guide the final selection decision. Manelloa and Calabrese (2019) conducted an ex-post analysis of the main factors influencing the supplier selection process in the automotive industry. They add new empirical evidence by investigating the impact of corporate reputation factors in the supplier selection criteria. For the selection of a headlamp supplier, Jain et al. (2018) determined the selection criteria through literature review and interviews with industry experts and ranked the suppliers using AHP and TOPSIS. Mathiyazhagan et al. (2018) investigated and ranked the environmental criteria for green supplier selection in the Indian automobile industry on the basis of expert judgment and ranked the supplier selection criteria. You et al. (2018) summarized the general evaluation criteria of automobile suppliers through literature screening and found out the categories and key criteria that automobile manufacturers need to consider when selecting component suppliers. Dai and Zhu (2019) established an auto component supplier evaluation system that considers the three dimensions of quality, business, and development and comprehensive capabilities. They used the TOPSIS method to determine the weight of each index and judged the classification of suppliers. Gupta et al. (2019) proposed a framework for the selection of green suppliers in the automotive industry and formulated six environmental standards and three conventional standards. This framework could help decision-makers distinguish the selection criteria of green suppliers. Choosing an appropriate method is very important for the supplier evaluation. Currently, scholars have put forward some relevant methods and a variety of combination methods to evaluate and select suppliers. For example, Park and Lee (2017) proposed a supplier selection method for automotive chassis parts through a hybrid method of DEA and AHP. Dweiri et al. (2016) proposed a supplier ranking model for the automobile industry based on AHP. This method provided managers with insights into various factors that needed to be considered when selecting suppliers. Mou et al. (2018) constructed a sustainable auto-part supplier selection model and proposed a three-stage decision-making method based on a probability distribution-hesitant fuzzy linguistic set and group decision-making theory. Memari et al. (2019) proposed an intuitionistic fuzzy TOPSIS method to select the appropriate automotive catalytic converter supplier. Liu et al. (2019) proposed a fuzzy three-stage integrated MCDM method for sustainable NEV part supplier selection. Zhou et al. (2019) proposed a multi-objective optimization model of auto-part suppliers considering customer complaints, and the multi-attribute utility theory and linear weighting method were used to obtain the optimization results. Wang et al. (2021) proposed a battery supplier selection framework for electric vehicles based on MCDM. In this framework, the MULTIMOORA method is used to sort the alternatives. Wu et al. (2020) proposed a hybrid framework that combines interval type 2 fuzzy sets, k-fuzzy measure, and Choquet integral operator to select the optimal green supplier in the electric vehicle industry. Liu et al. (2021) proposed a hybrid fuzzy symmetric MCDM model, integrating fuzzy linguistic set, the best and worst method, prospect theory, and VIKOR. The model could help NEV manufacturers select innovative power battery suppliers. Ilyas et al. (2021) used the integrated MCDM method based on BWM and FTOPSIS to consider the risks associated with COVID-19 so as to support the selection of suppliers in the automotive industry.

From the abovementioned literature review, we find that the related research provides some methods and ideas that can be used as a reference for this paper, such as the construction of a criteria system for NEV supplier selection. However, relevant methods and technologies in the existing research can only solve part of the problem in the selection of an NEV supplier, and there are still certain limitations, given as follows: 1) most of the existing studies focus on the selection of a traditional automobile component supplier, while the relevant methods and technologies to solve the problem of NEV component supplier selection are still obviously scarce (Galankashi et al., 2016; Hendiani et al., 2020). The method and criteria of component supplier selection for traditional automobile manufacturers are different from those for NEV manufacturers. For example, environmental protection capability and collaborative capability are less considered in the selection process of the traditional automobile supplier. In addition, most of the existing studies only focus on partial problems in the process of supplier selection, rather than systematically solving the problem of supplier selection, and select the best supplier based on the comprehensive performance of suppliers, without considering the indispensable subjective preference of decision-makers. 2) In terms of problem-solving methods, most of the existing studies solve the supplier selection problem based on the MCDM method. However, some common MCDM methods have certain limitations, such as the AHP method is highly subjective, the ranking has irregularities, and the use of additive aggregation may cause information loss; in the TOPSIS method, Euclidean distance is used to measure the distance between two schemes, and the correlation of decision criteria is not considered; some assumptions in DEA may not always be correct in reality; the weight of decision criteria in the VIKOR method is determined by the decision-maker subjectively, and there is no paired comparison between decision criteria and alternatives (Konidari and Mavrakis, 2007; Noori et al., 2021). 3) In terms of the supplier selection decision, most of the existing studies retrieve supplier case information through keyword matching and attribute similarity calculation so as to select component suppliers, which may result in problems such as low retrieval efficiency and low retrieval accuracy due to the lack of consideration of the semantic relations among keywords or attributes (Zhao and Yu 2011). 4) In terms of supplier selection criteria and their weight determination, most of the existing studies use the cited frequency or questionnaires to determine selection criteria, failing to fully consider the logic and semantic relationship of selection criteria (for example, product quality and quality are synonymous), which easily affects the efficiency of the supplier selection decision (Tavana et al., 2021). In addition, the weights of selection criteria in the existing studies are mostly derived from the subjective preferences of decision-makers and fail to fully consider the demands of users or objective data information. 5) In terms of acquiring and processing users’ demands, most of the existing studies collect users’ demands through questionnaires or structured scales. This easily restricts the expression of users’ demands, leading to insufficient consideration of users’ demands. In addition, the existing studies usually only focus on sentiment words of different sentiment polarities expressed in users’ demands or the proportions of different sentiment polarities (Balazs and Velásquez 2016) and fail to fully consider the different strengths of sentiment polarities. It is easy to cause the loss of information and one-sided sentiment consideration. 6) In terms of transforming user demand attributes into expected attributes toward suppliers, most of the existing studies only rely on the domain knowledge and experience of experts (Yang and Chai, 2018; Deng et al., 2021) and fail to integrate objective information, such as academic literature, industry standards, etc., which are highly subjective.

To overcome these limitations, it is necessary to develop a novel method to select a suitable component supplier for NEV manufacturers so as to enrich the relevant theory and method. Specifically, the purpose of this study is to solve the following three questions:

1) How to intelligently construct the ontology of selection criteria of the NEV component supplier so as to support the selection decision of component supplier for NEV manufacturers?

2) How to organize and mine user demand information concerning NEV and incorporate it into the weights of the supplier selection criteria so as to avoid information loss and reduce the interference of subjective factors in the calculation process?

3) How to realize the case information retrieval of component suppliers and improve the retrieval efficiency under the condition of considering the expectations of a decision-maker?

To address the abovementioned questions, we propose an integrated method for demand-driven NEV supplier selection based on ontology–quality function deployment (QFD)–case-based reasoning (CBR). In the method, the literature on the selection of the NEV component supplier is first preprocessed. Second, the domain concepts are extracted using the latent Dirichlet allocation (LDA), and the hierarchical agglomerative clustering (HAC) algorithm and association rules are used to extract the relations between domain concepts so as to construct the domain ontology of selection criteria of the NEV component supplier. Then, demand attributes are extracted from the demand documents provided by users with different demands using LDA, and the initial weights of demand attributes are determined based on the sentiment strength analysis algorithm and information entropy method. On the basis of this, the categories of demand attributes and the final weights of demand attributes are determined based on the Kano model and the Better-Worse coefficient. Furthermore, the expected attributes and their weights of NEV component supplier selection are determined based on QFD, and the semantic similarity of ontology concepts and attribute similarity are calculated based on CBR and ontology theory. Based on the hybrid similarity, the NEV component supplier is selected. Obviously, the proposed method enriches and develops the existing benchmark methods (such as the traditional sentiment analysis method and the traditional hybrid similarity calculation method) and has distinct characteristics and advantages in the construction of selection criteria for the NEV component supplier, the processing of user demand information, and the retrieval of NEV component supplier cases, which can help improve the rationality and efficiency of the selection decision of the NEV component supplier and address other similar issues (e.g., product design scheme optimization and case knowledge retrieval).

The remainder of this paper is organized as follows. Section 2 gives a description of the problem. In Section 3, the architecture of the demand-driven NEV supplier selection system is designed. Section 4 presents a solution framework and a method for demand-driven NEV supplier selection. In Section 5, an empirical study on the selection of a component supplier for the XP NEV manufacturer is given to illustrate the use of the proposed method. Finally, the conclusions of this study and the directions for future research are presented in Section 6.

2 Problem description

To clearly describe the problem of NEV supplier selection, the notations are defined as shown in Table 1.

TABLE 1. Notations frequently used throughout this paper.

The problem concerned in this study is how to construct the domain ontology of the selection criteria for NEV component suppliers, determine $E_b$, $d_a$, ${\omega }_b^{P^{\prime }}$, and ${\omega }_a^E$ and select an appropriate NEV component supplier from $Z_1,Z_1,\dots ,Z_{q_{\mu }}$ according to $D_j$, $P_k$, $z_{\mu a}^P$, and $z_a^o$ so as to support the NEV manufacturer’s procurement decision.

3 Demand-driven NEV supplier selection system

To address the above mentioned problem, the architecture of a demand-driven NEV supplier selection system is designed, as shown in Figure 1. The architecture of the system is divided into four layers: the user layer, the application layer, the business logic layer, and the data access layer. The user layer is the communication link between the user and the system, which is used to connect the main objects served by the system. Different objects can access different functional modules and participate in the requirement survey and the NEV supplier selection and evaluation process. The application layer consists of eight critical functional components: domain concept extraction module, domain ontology management module, user demand acquisition module, QFD module, case retrieval module, system management module, model and information management module, and case management module. This layer provides a series of key procedures that enable the system to support the NEV supplier selection decision. The business logic layer mainly provides the interface for system operation, allowing different users to complete the corresponding operation of the system. This layer uses Protégé to define the domain ontology, extracts the abstract semantic elements of the domain ontology to represent the cases, and implements case retrieval by considering user demands to obtain the ranking results of the suppliers. The data access layer uses MySQL as the main database management system to realize the storage of data by category.

FIGURE 1. Architecture of the demand-driven NEV supplier selection system.

The brief explanation of each functional component is given as follows:

• Domain concept extraction module: the main function of domain concept extraction is to select relevant domain literature from the literature database and extract domain concepts and relationships between domain concepts based on text mining. Since the existing literature on NEV supplier selection and evaluation has the characteristics of large quantity, knowledge dispersion, and multi-source heterogeneity, an effective method to solve the sharing, exchange, and integration of heterogeneous information, that is, ontology, is adopted. According to relevant scientific literature, the selection and evaluation criteria of suppliers are determined based on ontology. Applying knowledge management practices to supply chain management may greatly improve organizational performance.

• Domain ontology management module: the main function of domain ontology management is to model and maintain the domain knowledge about NEV supplier selection criteria so as to make the constructed domain ontology more comprehensive and complete. The quality of the domain ontology affects the determination of the expected attributes and the selection decision of the NEV supplier, so the setup of the domain ontology management module is necessary.

• User demand acquisition module: the main function of user demand acquisition is to collect users’ demands in the form of network documents, extract demand attributes from the demand documents, and use user demands as reference factors for NEV supplier selection. For NEV manufacturers, focusing on user demands can improve their competitive and strategic advantages. Therefore, the acquisition and analysis of user demands are important in supplier selection. In addition, collecting user requirements in the form of network documents can enable users to fully express their demands and make the obtained demand information more comprehensive.

• QFD module: the main function of the QFD module is to investigate and analyze user demands for NEV through a systematic approach and convert them into expected attributes concerning supplier selection. Full consideration of user demands is an effective way to rationally select NEV suppliers. It is helpful for NEV manufacturers to improve product quality and increase customer satisfaction by considering user demands to make an NEV supplier selection decision.

• Case retrieval module: the main function of case retrieval is to search the case base by using the similarity calculation method according to the relevant description of the alternative NEV supplier and then obtain the ranking of the alternative NEV suppliers based on the calculation results. Case retrieval is the most important step in the process of CBR, which can realize the real-time retrieval and related queries of cases in the case base.

• System management module: the main function of system management is user management and system settings. The system divides users into five categories based on permissions. An ordinary user can participate in market research and provide demand documents. An enterprise user can determine the case retrieval conditions according to the terms in the domain ontology and the current case, search the case library, and obtain the ranking results of the NEV supplier. Experts and knowledge engineers can assist in building domain ontology based on the processing results of text information, building case bases according to domain ontology and user demands, and managing and maintaining case bases. The system administrator is responsible for managing user permissions.

• Model and information management module: the information in the databases is the basis for implementing the system functions. The main functions of model and information management are to construct and update various calculation formulas in the model base and update and maintain all types of data in the databases.

• Case management module: the main function of case management is to build the case representation and retrieval mechanism and manage and maintain the case base. Since the case base is the basis for realizing case reasoning, the effectiveness of a supplier selection decision depends on the quality of cases, the representation method of case knowledge, and the retrieval mechanism to a large extent. Therefore, the setting of the module is necessary to ensure the continuous updating of the case base and better serve case retrieval.

4 Method for demand-driven NEV supplier selection

Based on the architecture of demand-driven NEV supplier selection system, we propose a resolution process, as shown in Figure 2. The resolution process is composed of three parts, that is, 1) construction of domain ontology of selection criteria for the NEV component supplier; 2) determination of demand attributes and expected attributes as well as their weights; and 3) selection of the NEV component supplier based on CBR. In the first part, the literature on the selection of the NEV component supplier is preprocessed, and the domain concepts are extracted using LDA. Then, the relations between domain concepts are extracted based on the HAC algorithm and association rules. Furthermore, the domain ontology of the selection criteria for NEV component suppliers can be constructed. In the second part, demand attributes are extracted from the demand documents provided by users with different demands using LDA, and the initial weights of demand attributes are determined based on a sentiment strength analysis algorithm and the information entropy method. Then, the categories of demand attributes and the final weights of demand attributes are determined based on the Kano model and the Better-Worse coefficient. Furthermore, the expected attributes of NEV component supplier selection are determined based on QFD. On the basis of this, the relationship matrix between the demand attributes and the expected attributes is constructed, and the weights of the expected attributes are determined. In the final part, based on CBR and ontology theory, the attribute similarities and hybrid similarities between the alternative cases (i.e., NEV component suppliers) and the target case (i.e., the expectation of decision-makers) are calculated, and the NEV component supplier with the largest hybrid similarity is selected.

FIGURE 2. Resolution process for the problem of demand-driven NEV supplier selection.

According to the resolution process shown in Figure 2, a description of the proposed method for demand-driven NEV supplier selection will be given in this section. The detailed descriptions of each part are, respectively, illustrated in Section 4.1, Section 4.2, and Section 4.3.

4.1 Construction of domain ontology of selection criteria for the NEV component supplier

In this section, the construction process of the domain ontology of selection criteria for the NEV component supplier includes four aspects: 1) representation of case knowledge based on ontology, 2) acquisition and preprocessing of text information, 3) extraction of domain concepts based on LDA, and 4) extraction of relationships between domain concepts.

4.1.1 Representation of case knowledge based on ontology

The ontology representation of case knowledge is the basis for realizing case knowledge retrieval, and it is also the basis for the selection of the NEV component supplier in this paper. In ontology, a framework system is used to describe the relationship between objective concepts, and the semantic association rules between concepts are defined to realize the communication and sharing between knowledge systems (Yoo and No, 2014; Lee et al., 2015). The ontology-based case knowledge representation method is easy to deal with structured knowledge, which can ensure the uniqueness of knowledge understanding. It is suitable for knowledge retrieval with complex semantic relations. The integration of ontology and CBR can improve the efficiency and accuracy of case retrieval (Küçük, 2015).

In this paper, the domain ontology of selection criteria for the NEV component supplier can be represented by the following tuple:

where $C$ is a concept set of cases; $A$ is an attribute set of each concept; $R$ is a relationship set between concepts; and $M$ is a set of instances.

There are three main steps to construct the domain ontology: 1) obtain domain concepts; 2) obtain the relationship between concepts; and 3) construct the ontology of tree structure (Guo and Zhou, 2017). Therefore, it is necessary to acquire domain concepts before constructing a domain ontology.

4.1.2 Acquisition and preprocessing of text information

To ensure the standardization and professionalism of domain concepts, domain concepts are extracted from academic literature in this paper. The academic literature is a type of knowledge resource with high professional value, and it is a relatively standard text form (Liu et al., 2011; Ren, 2012). The domain academic literature contains authoritative data such as domain ontology terms, concepts, and concept relationships. Due to its accessibility and high coverage in the research field, it is possible to construct a relatively complete domain ontology based on academic literature, and the domain ontology constructed is normative and professional (Tang et al., 2020). Therefore, academic literature is used as the source of domain concepts. The detailed description of the acquisition and preprocessing of academic literature is as follows.

First, “NEV,” “selection and evaluation of supplier,” and “automotive industry” are used as keywords, and “2015–2021” was used as search years to search databases, such as China National Knowledge Infrastructure (CNKI), WanFang database, ScienceDirect, and Web of Science. Manually search and download the literature on the selection criteria of the NEV component supplier, and remove duplicate literature in the retrieved literature.

Then, according to the set standards, the source literature of a domain concept is selected from the collected literature, that is, 1) for papers with multiple publications or citations, only one paper is selected; 2) for the review literature, the original literature in the reference list within the retrieval period is taken as the statistical literature; and 3) literature from various core journals or senior authors. For papers that meet the standards, the titles, keywords, abstracts, and text contents are saved, and they are regarded as domain literature. For the convenience of description, let $D=\left\{D_1,D_2,\dots ,D_m\right\}$ be a set of domain literature, where $D_j$ denotes the $j$-th domain literature and $m$ denotes the number of domain literature, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

Furthermore, the domain literature is preprocessed, which includes two steps, that is, 1) word segmentation and part-of-speech (POS) tagging, and 2) text data filtering. The details are given below.

4.1.2.1 Word segmentation and POS tagging

The Chinese Lexical Analysis System (ICTCLAS) and the Natural Language Toolkit (NLTK) are adopted to carry out word segmentation and POS tagging in domain literature $D_j$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$. To improve the accuracy, a secondary tagging method is used for POS tagging, that is, not only the verbs and nouns but also the verbs or adjectives with noun functions and proper nouns are tagged.

4.1.2.2 Text data filtering

To improve the training efficiency of the LDA topic model and the quality of the acquired domain concepts, it is necessary to sequentially perform paragraph filtering, stop word deletion, and POS filtering on domain literature $D_j$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$. The details are described as follows:

1) Paragraph filtering. The text content of each domain literature $D_j$ contains a large amount of text content that has nothing to do with the selection of an NEV component supplier, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$. Therefore, through keyword tagging, the paragraphs irrelevant to the selection criteria of an NEV component supplier are filtered out.

2) Stop word deletion. Stop words are words that appear frequently, but they have no practical meaning and cannot reflect domain-specific knowledge. To improve efficiency, the Chinese stop word list (CSWL) and English stop word list (ESWL) are used to delete the stop words in domain literature $D_j$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$. Based on the results of word segmentation and POS tagging, $D_j$ is compared with the stop words in the CSWL and ESWL. If a word in $D_j$ is the same as the word in CSWL or and ESWL, then it should be deleted; if not, it should be kept, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

After preprocessing the domain literature $D_j$, the preprocessed domain literature set can be obtained, noted as $D^{pro}=\left\{D_1^{\mathrm{p}\mathrm{r}\mathrm{o}},D_2^{\mathrm{p}\mathrm{r}\mathrm{o}},\dots ,D_m^{\mathrm{p}\mathrm{r}\mathrm{o}}\right\}$, where $D_j^{\mathrm{p}\mathrm{r}\mathrm{o}}$ denotes the preprocessed domain literature corresponding to $D_j$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$. Also, the initial concept set can be noted as $W{D}_j=\left\{d_1^j,d_2^j,\dots ,d_{q_j}^j\right\}$, where $d_Q^j$ denotes the $Q$-th initial concept in domain literature $D_j^{\mathrm{p}\mathrm{r}\mathrm{o}}$ and $q_j$ denotes the number of initial concepts in domain literature $D_j^{\mathrm{p}\mathrm{r}\mathrm{o}}$, $Q\in \left\{\mathrm{1,2},\dots ,q_j\right\}$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

4.1.3 Extraction of domain concepts based on LDA

In this section, a brief introduction of the LDA topic model is first given. Then, the process of extracting domain concepts based on LDA is illustrated.

4.1.3.1 LDA topic model

LDA is a three-layer Bayesian probability model that contains a three-layer structure of words, topics, and documents (Blei et al., 2003). It can be used to identify the topic information hidden in large-scale document sets or corpora and to effectively extract the key information from the text (Tirunillai and Tellis, 2014; Guo et al., 2017). The LDA model is a typical directed probability graph model, which is determined by the hyper parameters $\alpha$ and $\beta$, where $\alpha$ denotes the Dirichlet prior parameter of the multinomial distribution of a topic in any document and $\beta$ denotes the Dirichlet prior parameter of the multinomial distribution of the words under this topic. The probability map of the topic extraction model is shown in Figure 3.

FIGURE 3. Probability map of the topic extraction model.

In Figure 3, ${\theta }_d$ denotes the topic probability distribution in the $d$-th document, and ${\phi }_k$ denotes the probability distribution of words under the $k$-th topic. $W_{d,n}$ and $Z_{d,n}$ denote the $n$th word in the $d$-th document and the topic of the $n$th word in the $d$-th document, respectively. $D$ denotes the total number of documents, $K$ denotes the total number of topics, and $N_d$ denotes the number of words in the $d$-th document.

The topic is regarded as the middle layer. Through the probability of a certain topic in the document and the probability of a certain word in this topic, the probability of the word in the document can be obtained (Qian et al., 2016; Bi, Liu, Fan, and Cambria 2019). It can be expressed as a probability formula, that is,

where doc denotes a document, topic denotes a topic, and word denotes a word. $p\left(topic|doc\right)$ and $p\left(word|topic\right)$ can be calculated using $\theta$ and $\phi$, respectively.

4.1.3.2 The process of extracting domain concepts based on LDA

Domain concept extraction is an important part of domain ontology construction. Usually, domain concepts are those terms that are widely distributed and used more frequently in domain-related data and are less used in non-domain-related data (Zheng et al., 2019). Existing studies show that extracting domain concepts from academic literature conforms to the characteristics of LDA training text. Meanwhile, the LDA model has many advantages, such as flexibility and easy expansion, and has good domain portability (Feng and Zhang, 2017). Therefore, the LDA model is used to extract domain concepts in this paper. The details are given as follows.

Step 1. Let $WD{=\{w}_1{,w}_1,\dots ,w_R\}$ denote the set of words concerning the initial concept, where ${\mathrm{w}}_{\mathrm{r}}$ denotes the $\mathrm{r}$-th word in $WD$ and $R$ denotes the number of words, $r\in \left\{\mathrm{1,2},\dots ,R\right\}$. Then, the term frequency of each word in the word set, $t{f}_r^j$, is calculated. The calculation formulas of $WD$ and $t{f}_r^j$ can be represented, respectively, as follows:

where $n_r^j$ denotes the number of occurrences of the word $w_r$ in $W{D}_j$, $r\in \left\{\mathrm{1,2},\dots ,R\right\}$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

Step 2. Use Python and the LDA package in Sklearn to train the LDA model. In general, the hyper parameters $\alpha$ and $\beta$ of the LDA model are set to 50/K and 0.01, respectively (Blei et al., 2003). $K$ denotes the number of topics, and it is determined according to perplexity. The model perplexity is an index used to evaluate the language model. The lower the perplexity is, the stronger the predictive ability of the model will be, and the better the LDA model established will be. The calculation formula of perplexity is given by

where $D^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}=\left\{D_1^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}},D_2^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}},\dots ,D_{\mathrm{l}}^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}\right\}$ denotes a test set of domain literature, $W_d$ denotes the word in the domain literature $D_d^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}$, $N_d$ denotes the number of words in the domain literature $D_d^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}$, and $p\left(W_d\right)$ denotes the probability of word $W_d$ appearing in $D_d^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}$ and $D^{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}\subseteq D^{\mathrm{p}\mathrm{r}\mathrm{o}}$, $d\in \left\{\mathrm{1,2},\dots ,l\right\}$.

Step 3. Solve the trained LDA model to obtain the domain topic distribution. The candidate concept set is composed of all domain topics, noted as $C^{\mathrm{a}\mathrm{l}\mathrm{l}}=\left\{C_1^{\mathrm{a}\mathrm{l}\mathrm{l}},C_2^{\mathrm{a}\mathrm{l}\mathrm{l}},\dots ,C_g^{\mathrm{a}\mathrm{l}\mathrm{l}}\right\}$, where $C_e^{\mathrm{a}\mathrm{l}\mathrm{l}}$ denotes the $\mathrm{e}$-th candidate concept and $g$ denotes the number of candidate concepts, $e\in \left\{\mathrm{1,2},\dots ,g\right\}$.

Step 4. Calculate the term frequency–inverse document frequency (TF-IDF) of the word $w_r$ (Chen et al., 2008; Ray and Chandra, 2012). According to the results of TF-IDF, sort the set of candidate concepts obtained in step 3, screen high-frequency domain topics, and obtain the final domain concept set $C=\left\{C_1,C_2,\dots ,C_n\right\}$, where $C_i$ denotes the $i$ th domain concept and $n$ denotes the number of domain concepts, $i\in \left\{\mathrm{1,2},\dots ,n\right\}$. The calculation formula of TF-IDF of $w_r$ is given by

where $\left|\left\{j:w_r\in W{D}_j\right\}\right|$ denotes the number of initial concept sets containing the word $w_r$ (i.e., the number of sets with $n_r^j\ne 0$).

4.1.4 Extraction of relationships between domain concepts

The extraction of relationships between domain concepts is the most important step in domain ontology construction, which determines the final domain ontology structure (Zheng et al., 2019). There are two kinds of relations between domain concepts, namely, taxonomic relation and non-taxonomic relation. The taxonomic relation refers to the typical classification structure between domain concepts, such as “kind of.” The non-taxonomic relation refers to domain concepts that do not have a typical classification structure between concepts but have certain connections between concepts, such as “synonymous of.” In this paper, the HAC method and association analysis method are used to obtain the relationships between domain concepts (Han et al., 2019). The details are described as follows.

4.1.4.1 Mining taxonomic relation between domain concepts

Due to the variety of words, it is difficult to extract the relationship between many domain concepts based on linguistics, while clustering method can make up for this deficiency (Li et al., 2018). Therefore, this study uses the HAC method to mine the taxonomic relation between domain concepts.

First, the vector space model is used to construct the concept–document matrix (Castells et al., 2007). Since TF-IDF can measure the importance of a word to a document set, the vector space model of a domain concept can be constructed based on TF-IDF so as to describe the vector of a domain concept more comprehensively. The vector of domain concept, ${\overrightarrow{\mathrm{C}}}_{\mathrm{i}}$, is represented as follows:

where $TF-ID{F}_{ij}$ is obtained through Eq. 6, $i\in \left\{\mathrm{1,2},\dots ,n\right\}$ and $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

Then, the cosine distance method is used to calculate the similarity $\mathrm{s}\mathrm{i}\mathrm{m}\left({\overrightarrow{C}}_i,{\overrightarrow{C}}_h\right)$ between ${\overrightarrow{C}}_i$ and ${\overrightarrow{C}}_h$ (Liao and Xu, 2015), where the cosine value of two concept words is larger, the angle between the two vectors will be smaller, and the similarity between the two domain concepts will be greater, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$. The calculation formula of $\mathrm{s}\mathrm{i}\mathrm{m}\left({\overrightarrow{C}}_i,{\overrightarrow{\mathrm{C}}}_h\right)$ is given by

Furthermore, the semantic similarity matrix ${\overrightarrow{S}}_{ih}={\left[sim\left({\overrightarrow{C}}_{\mathrm{i}},{\overrightarrow{\mathrm{C}}}_{\mathrm{h}}\right)\right]}_{\mathrm{n}\times \mathrm{n}}$ between the domain concepts can be obtained, that is,

The specific clustering process is given as follows.

Step 1Construct the vector ${\overrightarrow{C}}_i$ base on the TF-IDF and use each vector as a cluster to form a cluster set $\overline{A}=\left\{A_1,A_2,\dots ,A_n\right\}$, $i\in \left\{\mathrm{1,2},\dots ,n\right\}$.

Step 2Using average-linkage (AL) hierarchical clustering as the distance between $A_i$ and $A_h$ (Xu et al., 2021), and the calculation formula is given by

where ${\overrightarrow{C}}_a$ and ${\overrightarrow{C}}_b$ denote the vectors in cluster $A_i$ and cluster $A_h$, respectively; $\left|A_i\right|$ and $\left|A_h\right|$ denote the number of vectors in the clusters $A_i$ and $A_h$, respectively.

Step 3Select the two clusters $A_i$ and $A_h$ with the highest average similarity. If $Dis\left(A_i,A_h\right)$ is greater than the preset threshold $\delta$ (usually, $\delta =0.5$), then merge the two clusters into a new cluster $A_{ih}$, add $A_{ih}$ to the cluster set $\overline{A}$, and remove $A_i$ and $A_h$ from cluster set $\overline{A}$, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$. Otherwise, go to step 5.

Step 4If the number of clusters in cluster set $\overline{A}$ is greater than 1, then go to step 2. Otherwise, go to step 5.

Step 5Clustering is completed, output cluster set $\overline{A}$.After clustering, the domain concepts can be divided into multiple clusters, but the division of parent concepts and sub-concepts within a cluster needs to be further defined. In this study, the average similarity within the cluster is used to extract the parent concept in the cluster. The calculation formula of average similarity within the cluster of domain concepts ${\mathrm{C}}_{\mathrm{i}}$ can be given by

where $n^{\prime }$ denotes the number of domain concepts within the cluster $A_i$. If the average similarity within the cluster of $C_i$ is greater, it means that $C_i$ is most closely related to other domain concepts in the cluster and has a broader semantics, so it can be used as the parent concept of other concepts in the cluster, $i\in \left\{\mathrm{1,2},\dots ,n\right\}$.

4.1.4.2 Mining non-taxonomic relation between domain concepts

The traditional association rules can only obtain the non-taxonomic relation between domain concepts but cannot obtain the specific relationship name (Agrawal et al., 1996). Usually, the non-taxonomic relation between domain concepts is composed of a verb and a pair of domain concepts, so this structure is used as the main extraction object, and verbs are used as the non-taxonomic relation between domain concepts to improve the traditional association rules in this study (Dong et al., 2013). The related domain concept pair set can be represented as $\mathrm{R}=\left({\mathrm{C}}_{\mathrm{i}},{\mathrm{C}}_{\mathrm{h}},{\mathrm{V}}_{\mathrm{t}}\right)$, where ${\mathrm{V}}_{\mathrm{t}}$ denotes the non-taxonomic relation between domain concepts ${\mathrm{C}}_{\mathrm{i}}$ and ${\mathrm{C}}_{\mathrm{h}}$, $\mathrm{i},\mathrm{h}\in \left\{\mathrm{1,2},\dots ,\mathrm{n}\right\}$, $\mathrm{i}\ne \mathrm{h}$, $\mathrm{t}\in \left\{\mathrm{1,2},\dots ,\mathrm{T}\right\}$. The analysis of the association rules process is described as follows.

Step 1. Randomly select two concepts ${\mathrm{C}}_{\mathrm{i}}$ and ${\mathrm{C}}_{\mathrm{h}}$ without association rule analysis from the domain concept set $\mathrm{C}=\left\{{\mathrm{C}}_1,{\mathrm{C}}_2,\dots ,{\mathrm{C}}_{\mathrm{n}}\right\}$, $\mathrm{i},\mathrm{h}\in \left\{\mathrm{1,2},\dots ,\mathrm{n}\right\}$, $\mathrm{i}\ne \mathrm{h}$. If there is no such domain concept pair, go to step 6.

Step 2. Calculate the support $\mathrm{Sup}$ and the confidence $\mathrm{C}\mathrm{o}\mathrm{n}$ (Le and Lo 2015). $\mathrm{Sup}$ and $\mathrm{C}\mathrm{o}\mathrm{n}$ can be, respectively, calculated by

where $f\left(C_{i,j},C_{h,j}\right)$ denotes the frequency of co-occurrence of concepts $C_i$ and $C_h$ in the domain literature $D_j^{pro}$. $f\left(C_{i,j}\right)$ denotes the frequency of concepts $C_i$ in the domain literature $D_j^{pro}$, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

Step 3. Set the threshold of support and confidence, that is, $Su{p}_{\min }$ and $Co{n}_{\min }$. If $\mathit{Sup}\left(C_i\to C_h\right)\ge Su{p}_{\min }$ and $Con\left(C_i\to C_h\right)\ge Co{n}_{\min }$, then $C_i$ and $C_h$ have a non-taxonomic relation, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$, and go to step 4. Otherwise, go to step 1.

Step 4. Count all the verbs $V_1,V_2,\dots ,V_T$ that appear between $C_i$ and $C_h$ in the domain literature $D_j^{pro}$ and their frequency $F{V}_1,F{V}_2,\dots ,F{V}_{\mathrm{T}}$, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$.

Step 5. If the frequency of $V_t$ is the largest and greater than the threshold $F_{\mathit{min}}$, then the verb $V_t$ is defined as the non-taxonomic relation between $C_i$ and $C_h$ and saved in the domain concept pair set $R$; go to step 1, $i,h\in \left\{\mathrm{1,2},\dots ,n\right\}$, $i\ne h$, $t\in \left\{\mathrm{1,2},\dots ,T\right\}$.

Step 6. After manually checking all the verbs, output the domain concept pair set $R$.According to the obtained domain concepts and the relationship between domain concepts, a domain ontology of tree structure for the selection criteria of NEV component suppliers can be constructed.

4.2 Determination of demand attributes and expected attributes as well as their weights

Based on the constructed domain ontology, demand attributes and expected attributes as well as their weights are determined, including the following three aspects: 1) determination of demand attributes and their weights, 2) determination of demand attribute categories and expected attributes, and 3) construction of a relation matrix and determination of expected attribute weights.

4.2.1 Determination of demand attributes and their weights

In this section, the process of extracting demand attributes from demand documents based on LDA is first illustrated. Then, the process of determining demand attribute weights is given. The detailed descriptions are given as follows.

4.2.1.1 Extracting demand attributes from demand documents based on LDA

The textual contents in demand documents for NEV component suppliers provided by users not only contain the words concerning demand attributes but also contain a lot of noisy and irrelevant words. To improve the effectiveness and efficiency of extracting demand attributes, demand documents are first preprocessed. Let $P=\left\{P_1,P_2,\dots ,P_v\right\}$ denote the set of demand documents, where $P_k$ denotes the $k$-th demand document and $v$ denotes the number of demand documents, $k\in \left\{\mathrm{1,2},\dots ,v\right\}$. Similar to the preprocessing of domain literature $D_j$, ICTCLAS is first used to process word segmentation and POS tagging on the demand document $P_k$, and then the word set concerning $P_k$ can be obtained by removing the stop words, noted as $W{D}_k^P=\left\{W{D}_{k1}^P,W{D}_{k2}^P,\dots ,W{D}_{kq}^P\right\}$, where $W{D}_{kc}^P$ denotes the $\mathrm{c}$th word in $W{D}_k^P$ and $q$ denotes the number of words in $W{D}_k^P$, $j\in \left\{\mathrm{1,2},\dots ,m\right\}$, $k\in \left\{\mathrm{1,2},\dots ,v\right\}$, and $c\in \left\{\mathrm{1,2},\dots ,q\right\}$.

Similar to domain concept extraction, topics can be extracted from preprocessed demand documents based on steps 2–4 in the process of extracting domain concepts. Since there may be some noisy words and topics with similar meanings in the extracted topics, in order to get more reasonable results, the noise words can be filtered manually, the topics with similar meanings can be merged, the important topics can be selected, and each topic can be labeled. Following Poria et al. (2016) and Bi et al. (2019), each extracted topic can be regarded as a demand attribute. Then, the set of labeled topics (i.e., demand attributes) and the set of words concerning each topic can be determined. The determined demand attribute $E_b$ can be denoted as $E_b=\left\{\mathrm{w}\mathrm{o}\mathrm{r}{\mathrm{d}}_{\mathrm{b}1},\mathrm{w}\mathrm{o}\mathrm{r}{\mathrm{d}}_{\mathrm{b}2},\dots ,\mathrm{w}\mathrm{o}\mathrm{r}{\mathrm{d}}_{\mathrm{b}{\mathrm{J}}_{\mathrm{b}}}\right\}$, where $\mathrm{w}\mathrm{o}\mathrm{r}{\mathrm{d}}_{\mathrm{b}\mathrm{J}}$ denotes the $\mathrm{J}$-th frequent word in the $\mathrm{b}$-th demand attribute and ${\mathrm{J}}_{\mathrm{b}}$ denotes the number of frequent words in ${\mathrm{E}}_{\mathrm{b}}$, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$ and $\mathrm{J}\in \left\{\mathrm{1,2},\dots ,{\mathrm{J}}_{\mathrm{b}}\right\}$.

4.2.1.2 Determining the weights of demand attributes

Usually, the demand information reflected in the demand document corresponds to multiple demand attributes. Hence, it is necessary to identify the demand information with respect to each demand attribute so as to determine the weight of each demand attribute. In accordance with the obtained word set $W{D}_k^P$ and the extracted demand attribute ${\mathrm{E}}_{\mathrm{b}}$, demand information concerning ${\mathrm{E}}_{\mathrm{b}}$ can be extracted, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Let $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ be the demand information concerning ${\mathrm{E}}_{\mathrm{b}}$ in $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$, it can be identified by comparing the words in $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ with the demand attribute word ${\mathrm{E}}_{\mathrm{b}}$, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Specifically, if ${\mathrm{E}}_{\mathrm{b}}\in \mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$, then the verbs, adverbs, and adjectives between two adjacent punctuations in the demand information, including demand attribute word ${\mathrm{E}}_{\mathrm{b}}$, are extracted from $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ (Huang and Cheng, 2015); if ${\mathrm{E}}_{\mathrm{b}}\notin \mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$, then $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}{=}^{″}{-}^{″}$, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Thus, $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ can be obtained, that is, $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left\{\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}1},\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}2},\dots ,\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}{\mathrm{u}}_{\mathrm{b}}}\right\}$, where $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{u}}$ and ${\mathrm{u}}_{\mathrm{b}}$ denote the $\mathrm{u}$-th word and the number of words in $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$, respectively, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$, and $\mathrm{u}\in \left\{\mathrm{1,2},\dots ,{\mathrm{u}}_{\mathrm{b}}\right\}$.

Based on the demand information concerning demand attributes, the multi-granularity sentiments with respect to demand attributes reflected by users in their demand documents can be identified. Generally, the word sets of sentiment strength with respect to different demand attributes may be different. Therefore, in order to improve the accuracy of sentiment analysis, a sentiment dictionary with respect to demand attributes needs to be established. Let $\mathrm{W}\mathrm{R}=\left\{\mathrm{W}{\mathrm{R}}_1,\mathrm{W}{\mathrm{R}}_2,\dots ,\mathrm{W}{\mathrm{R}}_{{\mathrm{h}}_{\mathrm{v}\mathrm{i}\mathrm{e}\mathrm{w}}}\right\}$ be the demand preference word set of users with respect to demand attributes, where $\mathrm{W}{\mathrm{R}}_{\mathrm{f}}$ and ${\mathrm{h}}_{\mathrm{v}\mathrm{i}\mathrm{e}\mathrm{w}}$ denote the $\mathrm{f}$-th word and the number of words in $\mathrm{W}\mathrm{R}$, respectively, $\mathrm{f}\in \left\{\mathrm{1,2},\dots ,{\mathrm{h}}_{\mathrm{v}\mathrm{i}\mathrm{e}\mathrm{w}}\right\}$. Also, the word set $\mathrm{W}\mathrm{R}$ can be represented by

To identify the sentiment polarity of users concerning demand attributes, HowNet (http://www.keenage.com/) was introduced to establish positive and negative sentiment dictionaries with respect to demand attributes. Let $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{+}$ and $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{-}$ be the positive and negative evaluation word sets that are commonly used in HowNet, respectively. Then, in accordance with $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{+}$, $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{-}$, and $\mathrm{W}\mathrm{R}$, the positive and negative sentiment dictionaries with respect to demand attributes, $\mathrm{W}{\mathrm{R}}_{\mathrm{H}}^{+}$ and $\mathrm{W}{\mathrm{R}}_{\mathrm{H}}^{-}$, can be constructed, that is,

For some words in $\mathrm{W}\mathrm{R}$ that may belong to neither $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{+}$ nor $\mathrm{W}{\mathrm{R}}_{\mathrm{H}\mathrm{o}\mathrm{w}\mathrm{N}\mathrm{e}\mathrm{t}}^{-}$, the manual identification method is adopted to determine the sentiment dictionary to which the corresponding words belong (Liu et al., 2017). Thus, the positive sentiment dictionary and negative sentiment dictionary with respect to demand attributes, $\mathrm{W}{\mathrm{R}}^{+}$ and $\mathrm{W}{\mathrm{R}}^{-}$, are finally formed.

Let ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ be the sentiment strength of $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ concerning ${\mathrm{E}}_{\mathrm{b}}$, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. In this study, the five-granularity sentiment strengths of users are considered based on actual demands. In particular, if the set $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ does not contain demand information concerning ${\mathrm{E}}_{\mathrm{b}}$, that is, $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}{=}^{″}{-}^{″}$, then ${\mathrm{s}}_{\mathrm{i}\mathrm{k}}^{\mathrm{P}}{=}^{″}{-}^{″}$, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Thus, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ has six possible values, that is, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\in \left\{-2,-\mathrm{1,0,1,2}{,}^{″}{-}^{″}\right\}$, where -2 (2) and -1 (1), respectively, represent the “very” and “slightly” degree of negative (positive) sentiment strengths and 0 represents the neutral sentiment strength, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Let $\mathrm{W}{\mathrm{R}}_{\mathrm{n}\mathrm{e}\mathrm{g}}$ be the set of negative words, $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}$ be the degree word set in HowNet, $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}^{+}$ be the set of enhanced sentiment words in $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}$, and $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}^{-}$ be the set of weakened sentiment words in $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}$. Therefore, the value of ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ can be obtained by comparing the words in set $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ with the words in sets $\mathrm{W}{\mathrm{R}}^{+}$, $\mathrm{W}{\mathrm{R}}^{-}$, $\mathrm{W}{\mathrm{R}}_{\mathrm{n}\mathrm{e}\mathrm{g}}$, and $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}$ (Li, Zheng, Yue, et al., 2021), $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. Let ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}$, and ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}$ be the indicator variables of intersections of $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ and $\mathrm{W}{\mathrm{R}}^{+}$, $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ and $\mathrm{W}{\mathrm{R}}^{-}$, $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ and $\mathrm{W}{\mathrm{R}}_{\mathrm{n}\mathrm{e}\mathrm{g}}$, and $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ and $\mathrm{W}{\mathrm{R}}_{\mathrm{d}}$, respectively. Then, the specific process of the multi-granularity sentiment analysis algorithm can be described as follows:

Step 1. If $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}{=}^{″}{-}^{″}$ or $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}{\leftarrow }^{″}{-}^{″}$, else go to step 2.

Step 2. If $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\cap \mathrm{W}{\mathrm{R}}^{+}\ne \varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}\leftarrow 1$, else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}\leftarrow 0$.

Step 3. If $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\cap \mathrm{W}{\mathrm{R}}^{-}\ne \varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}\leftarrow 1$, else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}\leftarrow 0$.

Step 4. If $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\cap \mathrm{W}{\mathrm{R}}_{\mathrm{n}\mathrm{e}\mathrm{g}}\ne \varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 1$, else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 0$.

Step 5. If $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\cap \mathrm{W}{\mathrm{R}}_{\mathrm{d}}=\varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}\leftarrow 0$; else if $\mathrm{W}{\mathrm{D}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\cap \mathrm{W}{\mathrm{R}}_{\mathrm{d}}^{+}\ne \varnothing$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}\leftarrow 1$; else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}\leftarrow -1$.

Step 6. If ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}={\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=0$ or ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}={\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=1$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\leftarrow 0$; when ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}=1$, if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}=1$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=0$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 1$ or ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}=0$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=1$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 0$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\leftarrow -1-{\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}$, else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\leftarrow 1+{\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}$; when ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{d}}=0$ or $-1$, if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}=1$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=0$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 1$ or ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}+}=0$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}-}=1$, ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}\mathrm{n}\mathrm{e}\mathrm{g}}\leftarrow 0$, then ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\leftarrow -1$, else ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}\leftarrow 1$.Based on the determined results of ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$, the weights of demand attributes can be obtained by further integrating the users’ demand preferences. In order to avoid information loss, this study uses an indicator vector to express sentiment strength information. Let ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left({\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{V}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{V}}\right)$ be the preference indicator vector of sentiment strength ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$, where ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}\mathrm{I}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{V}},{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{V}}=0$ or $1$ and ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}}+{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}}+{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{I}\mathrm{I}}+{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{I}\mathrm{V}}+{\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{V}}\le 1$, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. In accordance with the value of ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$, ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}$ can be determined. Specifically, if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=-2$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{1,0,0,0,0}\right)$; if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=-1$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{0,1,0,0,0}\right)$; if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=0$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{0,0,1,0,0}\right)$; if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=1$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{0,0,0,1,0}\right)$; if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=2$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{0,0,0,0,1}\right)$; especially, if ${\mathrm{s}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}{=}^{″}{-}^{″}$, then ${\mathrm{T}}_{\mathrm{k}\mathrm{b}}^{\mathrm{P}}=\left(\mathrm{0,0,1,0,0}\right)$, this is because most users will not express his/her demand preferences for a demand attribute when he/she holds a neutral sentiment toward the demand attribute in reality (Li, Zheng, Fan, et al., 2021), $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$.Next, the attribute value of $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ concerning demand attribute ${\mathrm{E}}_{\mathrm{b}}$ in the form of sentiment strength distribution, ${\mathrm{P}}_{\mathrm{k}\mathrm{b}}\left(\mathrm{x}\right)$, can be determined, $\mathrm{k}\in \left\{\mathrm{1,2},\dots ,\mathrm{v}\right\}$ and $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$, $\mathrm{x}\in \mathrm{\Omega }$, $\mathrm{\Omega }\in \left\{-2,-\mathrm{1,0,1,2}\right\}$, that is,

Furthermore, the expected values of $\mathrm{W}{\mathrm{D}}_{\mathrm{k}}^{\mathrm{P}}$ concerning demand attribute ${\mathrm{E}}_{\mathrm{b}}$ can be calculated by

On the basis of this, the information entropy method is adopted to calculate demand attribute weights (Chen 2020). Generally, the smaller the information entropy of ${\mathrm{E}}_{\mathrm{b}}$ is, the more information it will provide, that is, the greater the role it plays in the selection of NEV component suppliers, the greater its weight ${\omega }_{\mathrm{b}}^{\mathrm{P}}$ will be, and vice versa (Du and Gao, 2020; Li, Zheng, Yue, et al., 2021). The weight of the demand attribute ${\mathrm{E}}_{\mathrm{b}}$, ${\omega }_{\mathrm{b}}^{\mathrm{P}}$, can be calculated, that is,

where $\mathrm{I}\mathrm{E}\left({\overline{\mathrm{e}}}_{\mathrm{b}}\right)$ represents the information entropy concerning ${\mathrm{E}}_{\mathrm{b}}$, and it can be calculated by

In Eq. 20, ${\widehat{\mathrm{e}}}_{\mathrm{k}\mathrm{b}}$ can be calculated by Eq. 21. In particular, $0\cdot \ln \left(0\right)=0$.

4.2.2 Determination of demand attribute categories and expected attributes

The analysis of demand attributes can promote the determination of expected attributes. In this section, the Kano model is used to further analyze the demand attributes, which helps to determine the categories and priority of demand attributes so as to determine the expected attributes of NEV component suppliers efficiently (Kano et al., 1984).

The Kano category of each demand attribute can be determined by designing a two-factor questionnaire and counting the results of the questionnaire (Matzler and Hinterhuber, 1998). In the designed questionnaire, there are five response options for each demand attribute, and the questions are in the form of forward and reverse, as shown in Table 2, that is, the satisfaction of users when NEV component suppliers have the demand attribute ${\mathrm{E}}_{\mathrm{b}}$ and the dissatisfaction of users when NEV component suppliers do not have the demand attribute ${\mathrm{E}}_{\mathrm{b}}$, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. According to the results of the questionnaire, the attribute category with the largest number corresponding to each demand attribute is regarded as the final category of the demand attribute. Usually, there are five attribute categories (Xu et al., 2009), namely, must-be demand attributes ($\mathrm{M}$), one-dimensional (or performance) demand attributes ($\mathrm{O}$), attractive demand attributes ($\mathrm{A}$), indifferent demand attributes ($\mathrm{I}$), and reverse demand attributes ($\mathrm{R}$).

TABLE 2. Kano categories concerning demand attributes.

However, the traditional Kano model takes “the choice of the majority of users” as the final category of each demand attribute, which fails to consider the influence of the distribution of other attribute categories. In addition, there are some unavoidable problems in the process of data statistics. For example, there are many demand attributes in the same Kano category, and the priority cannot be determined. To solve these problems, the Better-Worse coefficient is calculated based on the traditional Kano model (Berger et al., 1993), and various categories of demand attributes are comprehensively considered. Specifically, according to the questionnaire survey results, the percentages of demand attribute categories can be counted (${\mathrm{A}}_{\mathrm{b}},{\mathrm{O}}_{\mathrm{b}},{\mathrm{M}}_{\mathrm{b}},{\mathrm{I}}_{\mathrm{b}},\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{R}}_{\mathrm{b}}$), and then the satisfaction coefficient and the dissatisfaction coefficient of users toward demand attribute ${\mathrm{E}}_{\mathrm{b}}$, $\mathrm{B}\mathrm{e}\mathrm{t}\mathrm{t}\mathrm{e}{\mathrm{r}}_{\mathrm{b}}$ and $\mathrm{W}\mathrm{o}\mathrm{r}\mathrm{s}{\mathrm{e}}_{\mathrm{b}}$, can be calculated, that is,

where $\mathrm{B}\mathrm{e}\mathrm{t}\mathrm{t}\mathrm{e}{\mathrm{r}}_{\mathrm{b}}\in \left(\mathrm{0,1}\right)$, and the greater the $\mathrm{B}\mathrm{e}\mathrm{t}\mathrm{t}\mathrm{e}{\mathrm{r}}_{\mathrm{b}}$ is, the greater the impact of demand attribute ${\mathrm{E}}_{\mathrm{b}}$ on users’ satisfaction will be; $\mathrm{W}\mathrm{o}\mathrm{r}\mathrm{s}{\mathrm{e}}_{\mathrm{b}}\in \left(-\mathrm{1,0}\right)$, and the smaller the $\mathrm{W}\mathrm{o}\mathrm{r}\mathrm{s}{\mathrm{e}}_{\mathrm{b}}$ is, the greater the impact of demand attribute ${\mathrm{E}}_{\mathrm{b}}$ on users’ dissatisfaction will be. Therefore, according to the calculation results of $\mathrm{B}\mathrm{e}\mathrm{t}\mathrm{t}\mathrm{e}{\mathrm{r}}_{\mathrm{b}}$ and $\mathrm{W}\mathrm{o}\mathrm{r}\mathrm{s}{\mathrm{e}}_{\mathrm{b}}$, the final categories of demand attributes can be determined, as shown in Figure 4, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$. The method of combining the Kano model with the Better-Worse coefficient is more convenient to mine the potential demands of users than the traditional Kano model and helps to further determine the expected attributes.

FIGURE 4. Better-worse coefficient quadrant chart.

Furthermore, the final weight of the demand attribute, ${\omega }_{\mathrm{b}}^{{\mathrm{P}}^{\prime }}$, can be obtained by adjusting the weight of ${\mathrm{E}}_{\mathrm{b}}$ (Högström et al., 2010), that is,

where ${\mathrm{k}}_{\mathrm{b}}$ denotes the adjustment coefficient; if ${\mathrm{E}}_{\mathrm{b}}$ is divided into $\mathrm{M}$, $\mathrm{O}$, $\mathrm{A}$, or $\mathrm{I}$, then ${\mathrm{k}}_{\mathrm{b}}$ is 2, 1.5, 1, or 0.5, respectively.

Based on the domain ontology constructed in Section 4.1 and the categories of demand attributes determined earlier, the demand attribute ${\mathrm{E}}_{\mathrm{b}}$ can be transformed into an expected attribute ${\mathrm{d}}_{\mathrm{a}}$ by an expert group, and each expected attribute can meet certain users’ demands, which makes the selection of expected attributes targeted, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$ and $\mathrm{a}\in \left\{\mathrm{1,2},\dots ,\delta \right\}$.

4.2.3 Construction of the relationship matrix and determination of expected attribute weights

Based on the determined demand attributes and expected attributes, the degrees of correlation between demand attributes and expected attributes can be evaluated by the expert group. Let ${\mathrm{R}}_{\mathrm{b}\mathrm{a}}$ denote the degree of correlation between the demand attribute ${\mathrm{E}}_{\mathrm{b}}$ and expected attribute ${\mathrm{d}}_{\mathrm{a}}$, where ${\mathrm{R}}_{\mathrm{b}\mathrm{a}}\in \left\{\mathrm{1,3,5,7,9}\right\}$, and 1, 3, 5, 7, and 9, respectively, denote the “very weak,” “weak,” “general,” “strong,” and “very strong” degrees of correlation, $\mathrm{b}\in \left\{\mathrm{1,2},\dots ,\mathrm{I}\right\}$ and $\mathrm{a}\in \left\{\mathrm{1,2},\dots ,\delta \right\}$.

Then, the importance weight and the weight of the expected attributes ${\mathrm{d}}_{\mathrm{a}}$, $\mathrm{I}{\omega }_{\mathrm{a}}^{\mathrm{E}}$, and ${\omega }_{\mathrm{a}}^{\mathrm{E}}$, can be obtained. $\mathrm{I}{\omega }_{\mathrm{a}}^{\mathrm{E}}$ and ${\omega }_{\mathrm{a}}^{\mathrm{E}}$ are, respectively, expressed as follows:

Furthermore, the relationship matrix $\stackrel{\sim }{\mathrm{R}}={\left[{\mathrm{R}}_{\mathrm{b}\mathrm{a}}\right]}_{\mathrm{I}\times \delta }$ between ${\mathrm{E}}_{\mathrm{b}}$ and ${\mathrm{d}}_{\mathrm{a}}$ can be constructed as shown in Table 3.

TABLE 3. Relationship matrix between ${\mathrm{E}}_{\mathrm{b}}$ and ${\mathrm{d}}_{\mathrm{a}}$.

4.3 Selection of the NEV component supplier based on CBR

Case retrieval is usually regarded as the most important step in CBR (Tadrat et al., 2012). The target of this process is to retrieve those cases in the case base that are closest to the new problem. To achieve this aim effectively, the similarities between the target case and the stored cases are measured (Homem et al., 2020). In this section, the expectation given by the decision-maker is taken as the target case, and the alternative NEV component suppliers are taken as the historical cases. By calculating the attribute similarity and hybrid similarity between the target case and the historical cases, the NEV component supplier that meets the expectations of the decision-maker is selected.

4.3.1 Attribute similarity measure for each format of attribute value

In case retrieval, there are usually multiple attributes involved, which may have different attribute value formats. In this study, four common attribute value formats are considered in NEV component supplier selection: ontology concepts, crisp numbers, interval numbers, and fuzzy linguistic variables. The attribute values of historical cases and target cases of the same expected attribute are expressed in the same format. Let ${\mathrm{Z}}_{\mu }$ be the $\mu$th historical case, $\mu \in \mathrm{U}$, $\mathrm{U}=\left\{\mathrm{1,2},\dots ,{\mathrm{q}}_{\mu }\right\}$. ${\mathrm{Z}}^{\mathrm{o}}$ denotes a target case, which is also the expectation of the decision-maker concerning NEV component suppliers. Let ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ be the attribute value of the historical case ${\mathrm{Z}}_{\mu }$ concerning the expected attribute ${\mathrm{d}}_{\mathrm{a}}$, $\mu \in \mathrm{U}$ and $\mathrm{a}\in \left\{\mathrm{1,2},\dots ,\delta \right\}$. Let ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$ be the attribute value of the target case ${\mathrm{Z}}^{\mathrm{o}}$ concerning the expected attribute ${\mathrm{d}}_{\mathrm{a}}$, $\mathrm{a}\in \left\{\mathrm{1,2},\dots ,\delta \right\}$. For the four attribute value forms considered, let ${\mathrm{D}}^{\mathrm{I}}=\left\{{\mathrm{d}}_1,{\mathrm{d}}_2,\dots ,{\mathrm{d}}_{{\mathrm{L}}_1}\right\}$, ${\mathrm{D}}^{\mathrm{I}\mathrm{I}}=\left\{{\mathrm{d}}_{{\mathrm{L}}_1+1},{\mathrm{d}}_{{\mathrm{L}}_1+2},\dots ,{\mathrm{d}}_{{\mathrm{L}}_2}\right\}$, ${\mathrm{D}}^{\mathrm{I}\mathrm{I}\mathrm{I}}=\left\{{\mathrm{d}}_{{\mathrm{L}}_2+1},{\mathrm{d}}_{{\mathrm{L}}_2+2},\dots ,{\mathrm{d}}_{{\mathrm{L}}_3}\right\}$, and ${\mathrm{D}}^{\mathrm{I}\mathrm{V}}=\left\{{\mathrm{d}}_{{\mathrm{L}}_3+1},{\mathrm{d}}_{{\mathrm{L}}_3+2},\dots ,{\mathrm{d}}_{\delta }\right\}$ denote the subsets of attributes concerning the formats of ontology concepts, crisp numbers, interval numbers, and fuzzy linguistic variables, respectively, where ${\mathrm{D}}^{\mathrm{I}}\cup {\mathrm{D}}^{\mathrm{I}\mathrm{I}}\cup {\mathrm{D}}^{\mathrm{I}\mathrm{I}\mathrm{I}}\cup {\mathrm{D}}^{\mathrm{I}\mathrm{V}}={\mathrm{D}}^{\mathrm{A}}$, ${\mathrm{D}}^{\mathrm{I}}\cap {\mathrm{D}}^{\mathrm{I}\mathrm{I}}\cap {\mathrm{D}}^{\mathrm{I}\mathrm{I}\mathrm{I}}\cap {\mathrm{D}}^{\mathrm{I}\mathrm{V}}=\varnothing$, and ${\mathrm{D}}^{\mathrm{A}}=\left\{{\mathrm{d}}_1,{\mathrm{d}}_2,\dots ,{\mathrm{d}}_{\delta }\right\}$. Accordingly, the subscripts of these four subsets can be represented as ${\mathrm{\Omega }}^{\mathrm{I}}=\left\{\mathrm{1,2},\dots ,{\mathrm{L}}_1\right\}$, ${\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}}=\left\{{\mathrm{L}}_1+1,{\mathrm{L}}_1+2,\dots ,{\mathrm{L}}_2\right\}$, ${\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}\mathrm{I}}=\left\{{\mathrm{L}}_2+1,{\mathrm{L}}_2+2,\dots ,{\mathrm{L}}_3\right\},$ and ${\mathrm{\Omega }}^{\mathrm{I}\mathrm{V}}=\left\{{\mathrm{L}}_3+1,{\mathrm{L}}_3+2,\dots ,\delta \right\}$, respectively, where ${\mathrm{\Omega }}^{\mathrm{I}}\cup {\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}}\cup {\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}\mathrm{I}}\cup {\mathrm{\Omega }}^{\mathrm{I}\mathrm{V}}={\mathrm{\Omega }}^{\mathrm{A}}$, ${\mathrm{\Omega }}^{\mathrm{I}}\cap {\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}}\cap {\mathrm{\Omega }}^{\mathrm{I}\mathrm{I}\mathrm{I}}\cap {\mathrm{\Omega }}^{\mathrm{I}\mathrm{V}}=\varnothing$, and ${\mathrm{\Omega }}^{\mathrm{A}}=\left\{\mathrm{1,2},\dots ,\delta \right\}$.

Considering the four attribute value formats, the method of measuring expected attribute similarity for each attribute value format is given as follows.

4.3.1.1 Attribute similarity measurement for ontology concept format attribute values

For ${\mathrm{d}}_{\mathrm{a}}\in {\mathrm{D}}^{\mathrm{I}}$, attribute values ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ and ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$ are the ontology concepts. In the ontology of tree structure, the closer the distance between two nodes is, the greater the similarity between the concepts corresponding to the two nodes will be. Therefore, the semantic similarity can be measured according to the path distance between two concept nodes (Huang and Zhou 2007). The semantic distance similarity between ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ and ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$ can be represented as $\mathrm{S}\mathrm{i}{\mathrm{m}}_{\mathrm{c}}\left({\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}},{\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}\right)$, then the calculation formula of $\mathrm{S}\mathrm{i}{\mathrm{m}}_{\mathrm{c}}\left({\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}},{\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}\right)$ is given by

where

In Eq. 28, $\mathrm{D}\mathrm{i}{\mathrm{s}}_{\mathrm{c}}\left({\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}},{\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}\right)$ denotes the semantic distance between ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ and ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$, $\mathrm{L}\mathrm{e}\mathrm{n}\left({\mathrm{L}}_{\mathrm{s}}\right)$ denotes the length of the $\mathrm{s}$-th side in the shortest path between ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ and ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$, and ${\xi }_{\mathrm{a}}$ denotes the total number of sides included in the shortest path between ${\mathrm{z}}_{\mu \mathrm{a}}^{\mathrm{P}}$ and ${\mathrm{z}}_{\mathrm{a}}^{\mathrm{o}}$.

However, the similarity algorithm based on semantic distance does not fully consider the factors that affect the similarity between ontology concepts, which may affect the accuracy of the case retrieval results (Wen et al., 2017). To improve the accuracy of the case retrieval results, the two factors that affect the similarity are considered, that is, the node depth and node density. The concept node depth is the number of sides included in the shortest path between the concept node and the root node. The semantic similarity of ontology concepts can be calculated using the node depth (Zhang et al., 2013), that is,