Fine-Grained Attribute Analysis of User Satisfaction for NEVs: An Interpretable Importance–Performance Analysis Based on Online Reviews

Abstract

The rapid growth of the new energy vehicle (NEV) industry has intensified competition among manufacturers, making user satisfaction a strategic priority. Importance–performance analysis (IPA) provides a viable framework for user satisfaction analysis by examining attribute importance and performance from user reviews. However, existing attribute importance assessment methods lack transparency, and IPA outputs remain coarse-grained. This study aims to develop an improved IPA framework that equips manufacturers with a data-driven approach to product improvement, derived from analysing user satisfaction through consumer reviews. To achieve this, a constrained priority linear programming model is proposed to enhance interpretability in importance measures, and the traditional IPA framework is extended by further diagnosing underperforming attributes through the integration of negative review analysis and the chi-square test, thereby deriving attribute-specific improvement directions. Empirical analysis of 8424 reviews from the Autohome platform demonstrates the effectiveness of the proposed method, and a cross-platform analysis of 4000 reviews from PCauto further confirms its robustness. Results identify that endurance and handling are priority attributes requiring resource allocation, followed by intelligence, interior, and comfort as attributes that need to maintain their advantages.

1. Introduction

Driven by the global energy transition and carbon emission reduction policies, the new energy vehicle (NEV) industry is undergoing unprecedented growth. Governments are accelerating NEV adoption through subsidies, tax incentives, and infrastructure development. According to EVTank, global NEV sales reached 23.542 million units in 2025, representing a year-on-year increase of 29.1% (http://www.evtank.cn/DownloadDetail.aspx?ID=666 (accessed on 20 March 2026)). The Chinese market has been particularly prominent. Data from the China Association of Automobile Manufacturers (CAAM) show that China’s NEV sales reached 16.49 million units in 2025, accounting for 70% of global sales (http://www.caam.org.cn/chn/4/cate_29/con_5237008.html (accessed on 20 March 2026)). In contrast, the European market recorded 3.77 million sales, reflecting a year-on-year increase of 30.5%, whereas the United States market saw modest growth, with sales reaching 1.6 million units, an increase of just 1.72% (http://www.evtank.cn/DownloadDetail.aspx?ID=666 (accessed on 20 March 2026)). Beyond these major markets, other countries are also actively exploring development pathways for the NEV industry. Through technological innovation, policy incentives, and market cultivation, they are continuously improving NEV competitiveness. In the passenger vehicle segment, data from Global NEVS (https://www.globalnevs.com (accessed on 20 March 2026)) indicate that although China continues to dominate the global electric passenger vehicle market, other countries are also experiencing strong sales growth (Figure 1). This suggests that the adoption of electric passenger vehicles is expanding from core markets to a broader global landscape, marking a new phase of diversified growth in the NEV industry.

With rapid technological advances and cost advantages in NEVs, both domestic and international automobile manufacturers have intensified market competition through increased investment. User satisfaction has become central to competitive strategy. To capture market share, manufacturers have shifted from technology-driven strategies to demand-oriented approaches. The concept of “User-ology” proposed by NIO Automobile (a leading Chinese NEV manufacturer) exemplifies this transformation, placing user service at the core of corporate strategy and using user satisfaction metrics to evaluate employee performance. This approach has enabled NIO to stand out in the NEV market. According to J.D. Power’s 2025 China New Energy Vehicle Customer Service Index (NEV-CSI) (https://china.jdpower.com/press-releases/china-new-energy-vehicle-customer-service-index-nev-csi-study (accessed on 20 March 2026)), NIO ranked first among NEV brands with a score of 801 points out of 1000, reflecting strong customer satisfaction. Thus, accurately identifying consumer needs and improving satisfaction are critical for automobile manufacturers.

Existing NEV satisfaction research mainly focuses on factors influencing overall satisfaction [1,2] and the underlying mechanisms [3,4], with specific attribute satisfaction often overlooked. Unlike overall assessments, fine-grained analyses are essential for precisely identifying consumer needs and generating actionable guidance for targeted improvements. Such analyses clarify how specific features affect user experience, enable firms to identify strengths and weaknesses more precisely, and allocate resources strategically, thereby enhancing competitiveness and user loyalty.

The importance–performance analysis (IPA) model diagnoses product or service attributes by simultaneously assessing their importance and performance [5]. Expectancy Confirmation Theory posits that satisfaction arises from the discrepancy between perceived performance and prior expectations [6]; performance that exceeds expectations generates positive affect, whereas performance that falls short yields negative affect. Consequently, performance scores capture the degree to which an attribute fulfils both design targets and consumer expectations. Consumer Choice Theory further indicates that individuals evaluate products or services by weighing their preferences and perceived values across attributes [7]; consequently, importance weights capture an attribute’s relative contribution to the consumer’s utility function and exert a decisive influence on satisfaction judgments. By jointly evaluating performance and importance, the IPA model is particularly suitable for fine-grained assessments of user satisfaction.

However, existing IPA studies often face limitations related to transparency and interpretability when assessing attribute importance. Traditional frequency-based methods only capture how often attributes are mentioned and do not explain their relationship with overall satisfaction [8,9]. Although machine learning models can process large datasets, their results are often difficult to interpret, which limits practical application [10,11,12]. Furthermore, extant research remains coarse-grained, identifying improvement areas without providing concrete, data-driven solutions—thereby leaving a gap between diagnostic output and actionable guidance. Consequently, despite the maturity of the IPA method, its application in the rapidly evolving NEV sector remains limited. Therefore, developing an IPA approach based on online reviews that yields specific improvement strategies for the NEV industry represents an urgent research need.

Based on online review data in the NEV domain, this study proposes an IPA-based approach to prioritise product attribute improvements. The model supports efficient allocation of limited resources to maximise consumer satisfaction. The study addresses four research questions: (1) How can the performance of each product attribute be accurately evaluated from online reviews? (2) How can the importance of consumer perceptions be effectively assessed while improving interpretability? (3) How can specific improvement suggestions for underperforming attributes be extracted? (4) How can performance and importance be jointly evaluated to determine improvement priorities?

To address these questions, the study adopts the following methods. (1) The pre-trained multilingual model tabularisai/multilingual-sentiment-analysis (https://huggingface.co/tabularisai/multilingual-sentiment-analysis (accessed on 9 June 2026)), a fine-tuned version of DistilBERT-base-multilingual-cased, is applied to evaluate attribute performance. (2) A priority-weighted linear programming model is developed to align attribute sentiment scores with overall ratings, where constraint priorities derived from rating discrepancies yield interpretable importance weights for product attributes. (3) For underperforming attributes, word frequency analysis of negative reviews captures consumer feedback and generates specific improvement suggestions, with the chi-square test and odds ratio (OR) verifying statistical significance. (4) Attribute performance and importance are integrated into the IPA framework to determine improvement priorities, enabling targeted resource allocation and enhanced consumer satisfaction.

The main contributions of this study are threefold. (1) It proposes an interpretable attribute importance assessment method based on constrained priority linear programming, addressing the transparency limitations of frequency-based and machine learning methods. (2) It extends the traditional IPA framework by developing a systematic diagnostic procedure for underperforming attributes, integrating negative review frequency analysis and the chi-square test to bridge the gap between traditional IPA outputs and actionable guidance. (3) It extends the improved IPA framework to the NEV industry, providing a fine-grained, data-driven analytical tool that yields specific, implementable improvement strategies for manufacturers.

The remainder of this paper is organised as follows. Section 2 reviews relevant literature. Section 3 presents the proposed IPA methodology based on online reviews. Section 4 applies this model to the NEV industry and presents enterprise-level recommendations. Section 5 conducts a validity assessment. Section 6 provides a discussion, and Section 7 summarises the entire paper.

2. Literature Review

2.1. Consumer Satisfaction Research in the NEV Industry

Existing studies on NEVs mainly focus on technological innovation [13], industrial development [14], marketisation strategies such as purchase intentions [15,16], and pricing strategies [17]. However, consumer satisfaction remains underexplored.

Consumer satisfaction refers to the degree to which consumers’ expectations are met through product or service consumption [18]. NEV satisfaction research has primarily focused on its influencing factors. Su et al. [1] used questionnaire data to analyse the effects of product performance dimensions on consumer satisfaction. However, this approach entails long data collection cycles and high costs. Online reviews have now become a timely and information-rich data source for market research. On this basis, Llopis-Albert et al. [19] analysed the effect of digital transformation on NEV consumer satisfaction using fuzzy set qualitative comparative analysis (fsQCA). Bai et al. [20] examined online NEV reviews from an emotional perspective and identified the mediating roles of perceived efficiency in intelligent experiences and perceived identity in environmental protection experiences in shaping consumer satisfaction. Yang et al. [2] further proposed a dynamic approach that incorporates time series analysis of online reviews to identify key factors across different price ranges and purchase purposes.

Several studies have explored the underlying mechanisms of satisfaction to explain complex market behaviours. Hasan [3] combined questionnaire data with structural equation modelling and found that satisfaction has a significant positive effect on repurchase intention. Ding et al. [21] proposed a large-scale group decision-making method based on sentiment analysis and quantum interference-based prospect theory to rank NEVs and support consumer decision-making. In addition, Bian et al. [4] developed an improved Bass model to examine the dynamic effect of consumer perception on NEV demand, highlighting the central role of satisfaction in purchase decisions.

Prior literature has demonstrated the importance of consumer satisfaction for firm performance. However, fine-grained attribute satisfaction—though proven valuable in other domains [22,23,24]—remains in its nascent stage within the NEV context. Yang et al. [25] proposed a demand analysis method based on BERT-TCBAD-Kano, providing decision-making support for user-centred product development. Zhang et al. [26] developed the FSP-Kano model to dynamically track changes in user requirements across product iterations and to guide attribute upgrade strategies. While the Kano-based studies focus on identifying user demand, this study shifts the analytical lens toward actionable satisfaction improvement within the IPA framework.

2.2. Importance-Performance Analysis Model

IPA is a well-established framework for systematically evaluating multiple product or service attributes in terms of both performance and importance. This approach enables organisations to identify the attributes consumers value most and assess current performance accordingly. IPA has been widely applied in various fields, including electronic products [23,27], tourism [28,29], and public transport [24,30]. The framework generally consists of three core steps, which are described below.

The first step concerns determining attribute performance, which has been measured through various methods in existing IPA studies. Quan and Han [31] employed Likert-scale questionnaires to collect consumer ratings, while Chen et al. [32] applied fuzzy logic to reduce subjectivity in questionnaire data. Sentiment analysis has also been widely adopted. Early approaches employed sentiment dictionaries [33] and rule-based tools such as VADER [27]. In recent years, machine learning methods have gained prominence, including Stanford CoreNLP [11], artificial neural networks [10], and transformer-based models such as BERT [12] and RoBERTa [34]. More recently, large language models, such as GPT, have been employed [35].

The second step is to determine attribute importance. Traditional IPA studies often rely on questionnaires to measure attribute importance, with respondents rating attributes using Likert scales [8,9]. More recently, online reviews have been increasingly used as data sources. Ma et al. [36] treated the frequency of attribute mentions in reviews as a proxy for importance. By contrast, Hauff et al. [37] combined the sufficiency logic of partial least squares structural equation modelling (PLS-SEM) with the necessity logic of necessary condition analysis (NCA) to derive more robust importance assessments. Joung and Kim [38] used deep neural networks to estimate attribute importance based on their effects on overall ratings; however, the black-box nature of such models limits interpretability. Wu et al. [28] proposed a linear programming model for attribute importance assessment that incorporates review credibility, while Wu and Liao [11] further developed a preference learning approach, offering promising alternatives to enhance interpretability.

The final step is to construct the IPA diagram, which visually represents attribute importance and performance on a two-dimensional plane. The traditional IPA matrix provides an intuitive representation that helps organisations identify priorities for product optimisation, service improvement, and strategic planning [39]. Several studies have extended the basic IPA framework. Kim et al. [40] integrated IPA with the Kano model to construct a four-quadrant matrix incorporating Kano-based weights. Ban et al. [41] proposed a 9Qc-IPA method that expands the traditional four-quadrant matrix to nine quadrants by introducing distance criteria and confidence measures. Extending this approach, Băban and Băban [42] developed a 9Qcf-IPA-BPN framework. Other studies have expanded the IPA framework from two to three dimensions to enhance analytical depth. Wang et al. [43] proposed the TIPCA model by introducing the gap between target product performance and importance as a third dimension. Chou [44] added a dissatisfaction dimension and used three-dimensional distances to prioritise attribute improvements, with greater distances from the origin indicating higher priority. Similarly, Qin et al. [45] introduced a competitiveness dimension to construct an IPCA model and support more effective product improvement strategies.

To clearly illustrate the methodological trends across the three stages of IPA research, their distinctions, and how this study diverges from existing studies,

Table 1. Comparison of several state-of-the-art studies.

Methods	Performance Evaluation	Importance Assessment	Improvement of the Classical IPA	Advantages	Disadvantages
Chen et al. [32]	Fuzzy questionnaire	Fuzzy questionnaire, Google search frequency	/	Combining Google search frequency to generate assessment results that better reflect social dynamics.	The accuracy of Google search data depends on keyword selection.
Lai and Wong [29]	Questionnaire, Interview	Questionnaire, Interview	/	Employing the correlation between attribute usage and perceived performance as an indirect performance measure to improve data reliability and validity.	Questionnaire samples may be subject to self-selection bias.
Ranjbari et al. [39]	Questionnaire, Likert scale	Questionnaire, Linear regression	/	Extracting and clustering attributes from online reviews to design questionnaires.	Questionnaire surveys are inherently subjective.
Sun et al. [8]	Questionnaire, Likert scale	Questionnaire, Linear regression	/	Calculating attribute importance by incorporating three-factor theory.	Questionnaire surveys are inherently subjective.
Joung and Kim [38]	IBM Watson	SHAP	/	Conducting IPA for different product segments.	Importance evaluation using deep neural networks and genetic algorithms requires substantial computational resources and time.
Chen et al. [33]	Sentiment dictionary	Association rule mining	/	Utilising association rule mining to assess importance values across multiple destinations.	Data collection relies on geographic coordinates and keywords, which may be one-sided.
Lu et al. [30]	TextBlob	Frequency of feature words	/	Aggregating results monthly to reflect trends in service quality changes.	Word frequency-based importance lacks scientific rigour.
Brochado et al. [10]	Star ratings	ANN	/	Applying a multidimensional segmentation approach.	ANN-based results exhibit limited interpretability.
Joung and Kim [27]	VADER	SHAP	/	Effectively explaining non-linear relationships.	Reliance on machine learning interpretability may still result in limited transparency.
Qin et al. [45]	Stanford CoreNLP	SHAP	Added the third dimension	Integrating multiple data sources and analytical techniques.	Potential computational efficiency issues with complex data.
Tanwar and Agarwal [24]	Questionnaire, Likert scale	Questionnaire, Likert scale	/	Simple calculation with low resource requirements.	Data collection requires substantial human intervention and is highly subjective.
Wang et al. [23]	SnowNLP	Frequency of text	/	Combining satisfaction and importance with time windows to construct a dynamic IPA matrix.	Word frequency-based importance lacks scientific rigour.
Wu et al. [28]	Stanford CoreNLP	Linear programming model	/	Introducing review credibility to improve importance and performance evaluation accuracy.	The definition of review credibility requires contextual adjustment.
Wang et al. [43]	GRU-CAP	Frequency of feature words	Added the third dimension	Introducing a third dimension improves result accuracy.	Word frequency-based importance lacks scientific rigour.
Wu and Liao [11]	Stanford CoreNLP	Preference Learning Model	Added asymmetric IPA	Extending traditional IPA to asymmetric IPA.	Reducing attribute importance to binary reward or penalty weights fails to capture continuous consumer perception.
our paper	DistilBERT-based multilingual sentiment model	Linear programming model with constraint priorities	Improved performance dimension	Proposing a more interpretable attribute importance calculation method and optimising the performance dimension.	Review reliability is not considered because the platform pre-screens reviews; no additional qualitative dimension is added to the traditional IPA framework.

presents a comparative analysis across key methodological dimensions.

For performance evaluation, existing studies mainly rely on questionnaires and Likert scales, sentiment dictionaries, or machine learning, with BERT- and GPT-based approaches gaining increasing popularity. Therefore, this study employs the tabularisai/multilingual-sentiment-analysis model, a fine-tuned version of the DistilBERT-base-multilingual-cased architecture. As a distilled and lighter-weight variant of BERT, DistilBERT’s multilingual cased version supports a wide range of languages, including Chinese.

For importance assessment, prior research has employed various methods. This study constructs an improved constrained priority linear programming model with clear advantages. Compared with questionnaires and Likert scales, combinations of questionnaires with Google search frequency, and word frequency methods, it substantially reduces human intervention and subjective bias, while improving interpretability and scientific validity. Relative to SHAP-based approaches, it requires fewer computational resources, reducing implementation costs, particularly for small and medium-sized enterprises. Unlike association rule mining, which captures only local relationships, the proposed linear programming model yields a global optimal solution. Compared with ANN-based methods, it avoids black-box issues and offers greater interpretability. In contrast to questionnaire-based regression analyses, online review data provide more timely and comprehensive coverage with reduced subjectivity. Finally, compared with existing linear programming approaches, the proposed model emphasises consistency between review sentiment values and user ratings. By introducing constraint priorities, it ensures that sentiment values associated with higher ratings substantially exceed those associated with lower ratings, thereby improving model construction and result interpretation.

Regarding IPA improvement, unlike studies that extend the traditional framework by adding a third dimension, this study enhances the performance dimension by conducting a further analysis of underperforming attributes. This makes the conclusions more actionable. In addition, this study applies the IPA model to the NEV industry at a fine-grained attribute level, providing strong support for targeted product improvement and consumer satisfaction enhancement.

3. Materials and Methods

3.1. Outline

The overall framework is illustrated in Figure 2 and consists of five steps. (1) Data preprocessing: Collected reviews are deduplicated and filtered to ensure data quality. (2) Performance evaluation: A fine-tuned DistilBERT-based multilingual sentiment model is applied to analyze review sentiment, and performance values for each attribute are derived from the sentiment analysis results. (3) Importance assessment: A linear programming model is constructed to estimate attribute importance, providing interpretable results that address limitations of many machine learning algorithms. (4) Low-performance attribute analysis: Word frequency analysis is conducted on reviews for underperforming attributes to highlight key issues and improvement directions. (5) IPA model construction: An IPA matrix is constructed to inform product improvement strategies. Section 3.2, Section 3.3, Section 3.4 and Section 3.5 describe these methods in detail.

3.2. Performance Evaluation

Many e-commerce platforms, such as Amazon, Ctrip, and Autohome, encourage consumers to provide both star ratings and textual reviews. Star ratings directly represent consumers’ overall satisfaction, whereas textual reviews contain richer content that reflects detailed evaluations of specific attributes. In natural language processing, sentiment analysis serves as a core tool for analysing consumer feedback [46]. From this perspective, online reviews can be regarded as reflections of product performance in real-world use. Accordingly, this study determines attribute performance by applying the tabularisai/multilingual-sentiment-analysis model to online reviews.

The tabularisai/multilingual-sentiment-analysis model, a fine-tuned version of the distilled BERT variant DistilBERT-base-multilingual-cased, has been proven to possess good cross-linguistic generalizability. The sentiment analysis procedure consists of two main steps. (1) The text is segmented and meaningless words, such as stop words, are removed to improve efficiency and accuracy. (2) The pre-trained model assigns sentiment scores to the processed text. The sentiment value of each attribute is calculated by analysing all sentences related to that attribute within a given review.

Specifically, the sentiment analysis model is first applied to obtain the sentiment value $S_{lj}$ of attribute $c_j$ in the $l\text{-}th$ review. Note that $S_{lj}$ is assigned a null value if the $l\text{-}th$ review does not contain any textual content related to attribute $c_j$. Then, an information fusion approach is employed to derive the overall performance value $Pe{r}_j$ of attribute $c_j$. Let R denote the set of all reviews, and $R_j$ denote the subset mentioning attribute $c_j$, where $R_j\subseteq R$. The performance $Pe{r}_j$ is computed as follows:

$$Pe{r}_j={\displaystyle \frac{{\displaystyle {\sum }_1^{\left|R_j\right|}{S}_{lj}}}{\left|R_j\right|}}$$

(1)

The resulting score $Pe{r}_j$ is a floating-point number between 0 and 1, with higher values indicating greater overall consumer satisfaction with attribute $c_j$. By comparing performance values across attributes, the strengths and weaknesses of the product can be identified. The median performance value across all attributes, denoted as $\overline{Per}$, is used as a benchmark. If $Pe{r}_j\ge \overline{Per}$, the performance of product attribute $c_j$ is considered positive; otherwise, it is deemed negative.

3.3. Importance Assessment

Drawing on preference learning research [11,28,47], it is recognized that increasing the importance of satisfaction-related attributes tends to enhance overall satisfaction, whereas greater importance assigned to dissatisfaction-related attributes intensifies overall dissatisfaction. Based on this rationale, this study proposes an improved constrained priority linear programming model. The model is built upon two key assumptions.

Assumption 1 (Additivity).

The overall satisfaction of a consumer with a product is an additive aggregation of his/her evaluations of individual attributes.

Thus, for the $l\text{-}th$ online review, the sentiment value $S_l$ is expressed as follows:

$$S_l={\displaystyle {\sum }_{j=1}^n{\mathrm{w}}_j{S}_{lj}}$$

(2)

where $n$ represents the number of product attributes, and the weight $w_j$ satisfies $w_j\in (0,1)$ and ${\displaystyle {\sum }_{j=1}^n{w}_j}=1$.

Assumption 2 (Weak Ordinality).

Although higher star ratings are generally associated with higher overall satisfaction at the population level, the relationship at the individual review level may be non-monotonic.

Assumption 2 distinguishes our model from previous models that impose strict monotonicity. By allowing for possible non-monotonicity at the individual review level, our model better reflects real-world consumer behaviour, where a user giving a rating of t may be more critical than a user giving a rating of t + 1, without necessarily implying lower satisfaction.

Accordingly, the model maps attribute performance to corresponding overall star ratings and introduces priority constraints based on rating differences. By minimizing deviation variables to estimate attribute weights, the approach quantifies each attribute’s contribution to the overall rating and provides a transparent and interpretable framework for importance assessment.

The parameter $w_j$ is obtained by solving a linear programming model that optimises the consistency between review sentiment values and corresponding overall star ratings. Online reviews typically use a five-point star rating system. Let the set of rating levels be {1, …, t, …, 5}, where $t$ denotes the $t\text{-}th$ rating level. Let $r_t\in R$ denote an arbitrary review with $t\text{-}th$ rating level. In this system, for any two adjacent rating levels $t$ and $t+1$, the sentiment value of $r_t$ is assumed to be lower than that of $r_{t+1}$. That is, $S_{r_t}<S_{r_{t+1}}$, where $t\in \{1,2,3,4\}$. In practice, not all reviews strictly satisfy this condition. To address this inconsistency, a non-negative deviation variable $d^{+}$ is introduced to allow flexibility while preserving a positive association between star ratings and sentiment values. The constraint is formulated as:

$$S_{r_t}-d^{+}\le S_{r_{t+1}}$$

(3)

where $d^{+}\ge 0$, $t\in \left\{1,2,3,4\right\}$.

In addition, rating discrepancies are explicitly incorporated into constraint prioritisation. Larger differences in star ratings should correspond to greater separation in sentiment values to ensure more accurate estimation of attribute importance. Accordingly, different priority levels are assigned to constraints based on rating differences. This design allows the model to better capture sentiment variation across rating levels and yields attribute weights that more closely reflect observed data. The resulting linear programming model is defined as:

$$\min z=M_1{\displaystyle \sum _{r_t,r_{t+4}\in R}{d}_{r_t,r_{t+4}}^{1+}}+M_2{\displaystyle \sum _{r_t,r_{t+3}\in R}{d}_{r_t,r_{t+3}}^{2+}}+M_3{\displaystyle \sum _{r_t,r_{t+2}\in R}{d}_{r_t,r_{t+2}}^{3+}}+M_4{\displaystyle \sum _{r_t,r_{t+1}\in R}{d}_{r_t,r_{t+1}}^{4+}}$$

(4)

$$s.t.\left\{\begin{array}{l}{S}_{r_t}-d_{r_t,r_{t+4}}^{1+}\le S_{r_{t+4}}, t=1, r_t,r_{t+4}\in R\\ S_{r_t}-d_{r_t,r_{t+3}}^{2+}\le S_{r_{t+3}}, t=1,2, r_t,r_{t+3}\in R\\ S_{r_t}-d_{r_t,r_{t+2}}^{3+}\le S_{r_{t+2}}, t=1,2,3 r_t,r_{t+2}\in R\\ S_{r_t}-d_{r_t,r_{t+1}}^{4+}\le S_{r_{t+1}}, t=1,2,3,4 r_t,r_{t+1}\in R\\ {\displaystyle {\sum }_{j=1}^n{w}_j=1, w_j\in (0,1), j=1,2,\dots ,n}\\ d_{r_t,r_{t+4}}^{1+},d_{r_t,r_{t+3}}^{2+},d_{r_t,r_{t+2}}^{3+},d_{r_t,r_{t+1}}^{4+}\ge 0\end{array}\right.$$

(5)

The objective function minimizes the weighted sum of positive deviation variables associated with different rating gaps. To prioritise constraints with larger rating differences, the coefficients are set such that $M_1\gg M_2\gg M_3\gg M_4$. Among them, $M_1$ represents the weight coefficient of the positive deviation variable $d_{r_t,r_{t+4}}^{1+}$, and the same logic applies to the other coefficients. The first constraint enforces sentiment separation for reviews differing by four stars, whereas the second to fourth constraint apply analogous conditions for smaller rating gaps. If $S_{r_t}\le S_{r_{t+4}}$, then $d_{r_t,r_{t+4}}^{1+}=0$; otherwise $d_{r_t,r_{t+4}}^{1+}>0$. The same principle applies to the remaining deviation variables. Solving this linear programming model yields the optimal importance weight $w_j$ for each attribute. The median of all attribute weights serves as the threshold, differentiating high-importance attributes from low-importance ones.

3.4. Low-Performance Attribute Analysis

Using the sentiment analysis method described in Section 3.2, sentiment values were obtained for each attribute in each review. By examining the distribution of these values, negative reviews associated with underperforming attributes were systematically identified, allowing specific sources of consumer dissatisfaction to be revealed.

After extracting the negative review dataset, additional preprocessing was conducted. First, the Jieba Chinese word segmentation tool was applied to the review texts to extract semantic information accurately. Next, a customised stop-word list was developed by combining general and domain-specific terms to remove irrelevant terms and improve data quality. After data cleaning, the remaining valid words were statistically analysed to calculate word frequencies.

To present the results intuitively, word cloud visualisation was used, where word size is positively associated with frequency of occurrence. This highlights the most frequently mentioned negative issues and potential associations among them. For example, frequent co-occurrence of a specific attribute with certain negative terms may indicate general quality problems or design deficiencies related to that attribute.

We further subjected the link between negative vocabulary and negative sentiment to a chi-square test to ascertain its statistical significance. A contingency table was constructed (

Table 2. Contingency table of negative vocabulary and review sentiment.

	Negative Reviews	Positive Reviews	Total
Negative Vocabulary Present	$n_1$	$n_2$	$n_1+n_2$
Negative Vocabulary Absent	$n_3$	$n_4$	$n_3+n_4$
Total	$n_1+n_3$	$n_2+n_4$	$n_1+n_2+n_3+n_4$

), in which the presence or absence of negative vocabulary was treated as one categorical variable and review sentiment (positive or negative) as the other. Here, $n_1$ represents the number of negative comments containing a specific negative vocabulary, while $n_2$ denotes those in positive comments. Conversely, $n_3$ represents the number of negative comments without that vocabulary, and $n_4$ represents the number of positive comments without it.

For the chi-square test, we propose the following hypotheses.

Null Hypothesis:

H0. 

There is no significant association between negative vocabulary and negative reviews.

Alternative Hypothesis:

H1. 

There is a significant association between negative vocabulary and negative reviews.

According to the chi-square test model, the chi-square statistic formula for negative vocabulary and comment sentiment is:

$${\chi }^2={\displaystyle \sum _{k=1}^4\frac{{(n_k-E_k)}^2}{E_k}}$$

(6)

where $E_k$ is the expected frequency of the cell containing $n_k$. Referring to the chi-squared distribution table, when ${\chi }^2>{\chi }_{\alpha }^2[(2-1)(2-1)]={\chi }_{\alpha }^2(1)$, the null hypothesis H0 is rejected and the alternative hypothesis H1 is accepted, indicating a significant association between negative vocabulary and negative reviews. A larger chi-square value indicates greater statistical significance.

The association strength between negative words and negative emotion could be quantified by the odds ratio (OR), which is defined as the ratio of the odds of the negative emotion occurring in the presence of negative words to the odds of it occurring in the absence of negative words:

$$OR={\displaystyle \frac{n_1{n}_4}{n_2{n}_3}}$$

(7)

OR > 1 means that negative words are positively correlated with negative emotion.

3.5. IPA Model Construction

IPA visualises the relationship between attribute performance and importance using a quadrant diagram, providing managers with an intuitive basis for decision-making. In this study, the methods described in Section 3.2 and Section 3.3 are used to quantify attribute performance and importance, which are then mapped into the four quadrants using their respective mean values as axes baselines, each associated with a distinct strategic implication.

Priority improvement areas: Attributes in the fourth quadrant are highly important to consumers but exhibit low performance. These attributes should be prioritised for improvement to enhance consumer satisfaction.

Resource optimisation suggestions: Attributes in the second quadrant demonstrate high performance but relatively low importance. Resources allocated to these attributes can be reduced or reallocated to improve overall efficiency.

Advantage maintenance strategy: Attributes in the first quadrant represent core strengths that are highly important and perform well. These attributes should be maintained at a high level and further refined to sustain competitive advantage.

Low-priority features: Attributes in the third quadrant have low importance and low performance. Decisions regarding the retention or improvement of these attributes can be made based on specific strategic considerations.

4. Results

4.1. Data Collection and Preprocessing

The data were collected from Autohome (https://www.autohome.com.cn, Beijing, China), one of China’s leading automotive internet platforms. The platform hosts a large volume of user reviews covering a wide range of vehicle models. BYD, a major participant in China’s NEV market with strong domestic and international performance, has attracted extensive user feedback across several models. This provides rich and representative data for empirical analysis. Therefore, this study selected the BYD Han, a major BYD model with the highest number of user reviews on the Autohome platform, to examine consumer evaluations and identify user satisfaction characteristics.

To collect the relevant data efficiently, a web crawler was developed using Python 3.10. The crawler simulated user browsing behaviour to sequentially access the BYD Han review pages on the Autohome platform and extract review content. Owing to platform constraints, a maximum of 200 pages could be accessed, with 10 reviews displayed per page. We tracked data from January 2024 to February 2026, collecting a total of 9165 comments. After removing duplicates and invalid data, 8424 valid reviews remained. Each review contains key fields such as reviewer ID, review title, model information, review date, review content, and star rating.

The platform’s evaluation system categorises reviews into 11 dimensions. Excluding “most satisfied” and “least satisfied”, the remaining nine attributes were selected for analysis: space ($c_1$), driving experience ($c_2$), endurance ($c_3$), appearance ($c_4$), interior ($c_5$), quality–price ratio ($c_6$), intelligence ($c_7$), handling ($c_8$), and comfort ($c_9$). These attributes closely correspond to commonly used NEV quality dimensions: $c_1$, $c_3$, and $c_8$ relate to vehicle performance and functionality; $c_2$ and $c_9$ capture user experience; $c_4$ and $c_5$ reflect aesthetic design; $c_6$ and $c_7$ represent functional value and technological level. They have also been widely identified as key factors influencing consumer satisfaction in NEV-related studies [2,21,25,26]. The frequency distribution of attribute evaluations is reported in

Table 3. Frequency of evaluations for each attribute.

Attribute	${\mathit{c}}_1$	${\mathit{c}}_2$	${\mathit{c}}_3$	${\mathit{c}}_4$	${\mathit{c}}_5$	${\mathit{c}}_6$	${\mathit{c}}_7$	${\mathit{c}}_8$	${\mathit{c}}_9$
Frequency	1	0.980	0.982	0.976	0.984	0.970	0.677	0.303	0.301

. The results indicate that attributes $c_1$, $c_2$, $c_3$, $c_4$, $c_5$, and $c_6$ are mentioned most frequently in user reviews.

Each review was segmented into attribute-specific passages based on the nine defined themes. After segmentation, Jieba was used for word tokenisation, and a stop-word list was applied to remove irrelevant terms. Sentiment analysis was then conducted separately for each attribute-specific dataset to support performance and importance evaluation.

4.2. Estimation of Attribute Performance

The tabularisai/multilingual-sentiment-analysis model was applied to the preprocessed review texts to quantify sentiment scores for clauses related to different attributes. For each review, the sentiment scores of all clauses corresponding to a given attribute were aggregated by calculating their mean, yielding an attribute-level sentiment value. Partial results are reported in

Table 4. Sentiment analysis results for each review.

Index	${\mathit{c}}_1$	${\mathit{c}}_2$	${\mathit{c}}_3$	${\mathit{c}}_4$	${\mathit{c}}_5$	${\mathit{c}}_6$	${\mathit{c}}_7$	${\mathit{c}}_8$	${\mathit{c}}_9$
R1	0.871	0.852	0.836	0.448	0.742	0.225	0.811	/	/
R2	0.922	0.618	0.501	0.866	0.587	0.551	/	/	/
…	…	…	…	…	…	…	…	…	…
R8424	0.733	0.461	0.498	0.400	/	0.274	0.513	/	/

, where “/” indicate that the corresponding review does not mention that attribute.

To evaluate the accuracy of sentiment scoring, all reviews from the “most satisfied” and “most dissatisfied” categories were used. As predefined platform categories, these groups inherently carry positive and negative sentiment labels, respectively, enabling validation without manual annotation and reducing subjective bias. Sentiment scores were classified as negative (0) or positive (1) using a threshold of 0.5. The resulting performance metrics were Recall = 0.885, Precision = 0.918, and F1-score = 0.901. These results indicate that the sentiment analysis model can effectively capture sentiment tendencies on this dataset.

Using Equation (1), sentiment analysis results from all reviews were aggregated to evaluate the performance of the BYD Han across different attributes (

Table 5. Composite attribute performance based on online reviews.

Attribute	${\mathit{c}}_1$	${\mathit{c}}_2$	${\mathit{c}}_3$	${\mathit{c}}_4$	${\mathit{c}}_5$	${\mathit{c}}_6$	${\mathit{c}}_7$	${\mathit{c}}_8$	${\mathit{c}}_9$
Performance	0.609	0.651	0.527	0.674	0.645	0.586	0.654	0.627	0.632

). The median performance score $\overline{Per}$ was 0.632. Based on this benchmark, the BYD Han exhibits positive performance in driving experience ($c_2$), appearance ($c_4$), interior ($c_5$), and intelligence ($c_7$), with comparatively weaker performance in the remaining attributes. Appearance ($c_4$) shows the highest performance, whereas endurance ($c_3$) shows the lowest.

To further assess the robustness of these findings, the distribution of sentiment scores was analysed. Heatmaps illustrating the distribution for each attribute are presented in Figure 3, where colour intensity represents review frequency. For appearance ($c_4$), scores are highly concentrated in the upper range, which is consistent with its highest performance ranking and indicates broadly positive consumer evaluations. By contrast, scores for endurance ($c_3$) are more widely dispersed with fewer high values, supporting the conclusion that endurance exhibits the weakest performance.

4.3. Determination of Attribute Importance

In the dataset, total review scores are decimal values ranging from 0 to 5. To facilitate analysis, these scores were categorised into five star-rating levels based on their distribution (

Table 6. Correspondence between total review scores and star rating levels.

Rating (x)	4.5 < x ≤ 5	3.8 < x ≤ 4.5	2.8 < x ≤ 3.8	1.5 < x ≤ 2.8	0 < x ≤ 1.5
Categorisation	5	4	3	2	1

). Attribute importance was then estimated by solving the linear programming model defined in Equations (4) and (5). Following common practices for setting priority coefficients, $M_1, M_2, M_3$ and $M_4$ were set to 1000, 100, 10, and 1, respectively. This configuration ensures that deviations associated with larger differences in star ratings receive higher priority during optimisation. The resulting attribute importance values are reported in

Table 7. Importance of each attribute based on online reviews.

Attribute	${\mathit{c}}_1$	${\mathit{c}}_2$	${\mathit{c}}_3$	${\mathit{c}}_4$	${\mathit{c}}_5$	${\mathit{c}}_6$	${\mathit{c}}_7$	${\mathit{c}}_8$	${\mathit{c}}_9$
Importance	0.085	0.058	0.136	0.025	0.129	0.059	0.135	0.200	0.174

.

The results indicate clear differences in the importance of product attributes. Handling ($c_8$) is identified as the most important attribute, whereas appearance ($c_4$) is identified as the least important. A comparison between Table 3 and Table 7 shows that frequently mentioned attributes (e.g., $c_1$ and $c_4$) do not exert a strong influence on consumers’ overall evaluations. Conversely, attributes that are less frequently mentioned (e.g., $c_8$ and $c_9$) may have a greater impact on overall satisfaction.

To examine model robustness, 80% of reviews were randomly sampled to recalculate attribute importance. The specific results are shown in Appendix A. In addition, a sensitivity analysis was conducted on the priority coefficients $M_1$, $M_2$, $M_3$ and $M_4$. The specific results are shown in Appendix B. The analysis shows that although absolute weight values vary slightly and the order of some intermediately ranked attributes shows minor variations in a few scenarios, the highest- and lowest-importance attributes remain unchanged across all scenarios.

4.4. Further Analysis of Underperforming Attributes

To further explore the causes of underperformance, an in-depth analysis was conducted for attributes exhibiting low performance. First, all negative reviews associated with each underperforming attribute were extracted. After data preprocessing, word frequency analysis was applied to the top 200 most frequent words in these reviews. The resulting word clouds are shown in Figure 4.

For the attribute space ($c_1$), consumers mainly express dissatisfaction with internal space, particularly regarding wheelbase, vehicle length and head clearance. Vehicle class may be another concern, with the sedan being the primary class causing user dissatisfaction.

Dissatisfaction with endurance ($c_3$) mainly relates to fuel consumption, electricity use, and electricity cost, which reflect user concerns about mileage. Additional factors such as air-conditioning use and winter conditions further reduce perceived endurance.

For the quality–price ratio ($c_6$), terms such as price, maintenance, and ownership cost indicate that users are highly concerned about purchase and maintenance costs. Additionally, references to sales advisor, discount, purchase tax and subsidy suggest that promotional activities and insufficient purchase subsidies may also be focal points for users.

Negative feedback on handling ($c_8$) primarily concerns U-turn and turning radius. Additional complaints relate to the vehicle’s body size affecting manoeuvrability. These issues suggest that consumers find the vehicle less agile in specific scenarios, such as cornering.

To investigate whether negative vocabulary is significantly associated with negative reviews, chi-square tests were conducted for each underperforming attribute. Results in

Table 8. Chi-square values for negative vocabulary and review sentiment.

Attribute	Negative Vocabulary	df	Chi-Square Value	OR
Space (c1)	Wheelbase	1	4.818	1.202
	Head Clearance	1	12.328	1.446
	Sedan	1	5.705	1.306
	Vehicle Length	1	16.719	1.699
	Vehicle Class	1	13.199	1.671
Endurance (c3)	Fuel Consumption	1	6.981	1.151
	Kilometres	1	23.276	1.253
	Air Conditioning	1	23.484	1.386
	Fuel	1	21.716	1.438
	Winter	1	71.725	2.228
Quality-Price Ratio (c6)	Price	1	4.687	1.112
	Discount	1	273.061	2.564
	Maintenance	1	20.703	1.387
	Subsidy	1	43.595	1.922
	Purchase Tax	1	30.365	1.690
Handling (c8)	Body Size	1	14.683	1.627
	U-Turn	1	50.194	4.187
	Electric Vehicle	1	15.922	2.353
	Turning Radius	1	12.035	2.086
	Fuel	1	11.328	2.415

${\chi }_{0.05}^2(1)=3.841$.

show that negative vocabulary related to each attribute is significantly correlated with negative reviews. For example, for space ($c_1$), the chi-square value for “head clearance” is 12.328, which substantially exceeds the critical value of 3.841. Furthermore, Table 8 demonstrates that OR values are all greater than 1. This indicates a strong positive correlation with negative reviews and suggests that improvements in these aspects should be prioritised.

Overall, by combining high-frequency negative word analysis with significance testing, this study identifies the main sources of consumer dissatisfaction for each underperforming attribute.

4.5. Construction of the IPA Model

Figure 5 presents the IPA diagram constructed using the computed performance and importance, which illustrates the improvement strategies for the BYD Han. To examine the stability of attribute positions in the IPA model, additional diagrams were reconstructed based on importance results from Appendix A and Appendix B. The results in Appendix C show that attribute positions remain largely stable under repeated random sampling and weight coefficient variations.

Based on the IPA results, the nine attributes are classified into four quadrants.

(1) Quadrant IV: Concentrate here

Endurance ($c_3$) and Handling ($c_8$) are located in the priority-improvement quadrant. These attributes require focused resource allocation and immediate improvement actions. Word frequency analysis of negative reviews provides specific directions for improvement.

(2) Quadrant I: Keep up the good work

Interior ($c_5$), intelligence ($c_7$) and comfort ($c_9$) are located in the advantage maintenance quadrant. Enterprises should sustain their competitive strengths through continuous investment and monitoring. Notably, it is worth noting that interior ($c_5$) lies close to the importance threshold, suggesting that enterprises should invest prudently in maintaining its competitive advantage. Comfort ($c_9$) lies near the performance boundary, suggesting that minor shifts in user perception could elevate its priority.

(3) Quadrant III: Low priority

Space ($c_1$) and quality-price ratio ($c_6$) are positioned in the low-priority quadrant. These attributes should be maintained at a basic acceptable level, with resources allocated only after higher-priority attributes have been addressed.

(4) Quadrant II: Possible overkill

Driving experience ($c_2$) and appearance ($c_4$) fall within the resource optimisation quadrant. This suggests that firms should allocate resources rationally to maintain acceptable performance levels without excessive investment, thereby maximising overall consumer satisfaction while controlling optimisation costs.

5. Cross-Platform Verification

To examine the generalisability of the proposed methodology across different data sources, this study applied the same analytical framework to 4000 user reviews of the BYD Han collected from the PCauto platform (Guangzhou, China). PCauto classifies user evaluations into ten dimensions: advantages, disadvantages, appearance, interior, space, configuration, dynamics, handling, fuel consumption/endurance, and comfort. Handling-related reviews were completely absent, and fuel consumption/endurance reviews accounted for less than 19% of the sample. Therefore, the analysis focused on six core attributes: appearance, interior, space, configuration, dynamics, and comfort. Using the methodology described in Section 3, an IPA model was constructed, as shown in Figure 6. A comparison of the common attributes across the two datasets, as illustrated in Figure 5 and Figure 6, yields several consistent findings.

First, interior and comfort fall into the Quadrant I (high importance, high performance) in both figures, whereas appearance remains in the Quadrant II (low importance, high performance). Meanwhile, configuration in Figure 6 is conceptually similar to quality–price ratio in Figure 5, and both attributes are located in the Quadrant III (low importance, low performance).

Second, although space shows low performance in both figures, its quadrant position differs: it appears in the Quadrant III in Figure 5 but shifts to Quadrant IV in Figure 6. This likely reflects variation in consumer preferences across platforms and market segments. The result suggests that space still has room for improvement and requires targeted optimisation tailored to different consumer groups.

Third, dynamics in Figure 6 can be regarded as conceptually related to driving experience in Figure 5. This interpretation is supported by the word frequency analysis on the Autohome platform, where comments related to driving experience frequently mention dynamics performance. Although driving experience in Figure 5 falls into the Quadrant II (relatively high performance), while dynamics in Figure 6 lies in the Quadrant III (low performance), both findings collectively indicate that users from different platforms do not regard this attribute as a key factor influencing their satisfaction.

Overall, the cross-platform comparison yields largely consistent conclusions for the same vehicle model. The discrepancies observed in the analysis of space and the driving experience/dynamics attributes may reflect variation in consumer preferences across platforms and market segments.

6. Discussion

Based on the above findings, this study proposes a three-tiered resource allocation strategy.

(1)

Priority improvement suggestions

Endurance is a key determinant of user satisfaction. Current dissatisfaction with fuel consumption and winter conditions indicates that immediate improvement actions are required from enterprises. First, substantial R&D investment should focus on advancing core battery and energy management technologies. Second, enterprises should offer practical ranges across different price segments to better align product offerings with consumer expectations. Third, managing consumer trust is equally critical. This can be achieved through transparent disclosure of battery degradation and proactive management of range expectations.

Handling is positioned on the boundary between priority improvement and advantage maintenance, suggesting that enterprises should also strengthen resource allocation for this attribute. To address specific consumer complaints—such as U-turn and turning radius—firms should invest in next-generation handling technologies. Furthermore, the advanced technologies could be marketed as core safety features that enhance vehicle stability and driver control, rather than merely as performance enhancements.

Although space is identified as an attribute with relatively low impact on user satisfaction in the dataset from Autohome, the results from PCauto show slightly different patterns. This discrepancy suggests that enterprises should still pay attention to user complaints regarding internal space, and accordingly optimise space design.

(2)

Advantage maintenance suggestions

Comfort lies near the performance boundary, suggesting that minor shifts in user perception could elevate its priority. Therefore, after addressing the immediate concerns of endurance and handling, enterprises should focus on user needs related to comfort. Continuous collection of user feedback and iterative optimisation may help enhance its performance.

Intelligence is highly rated by users, positioning it as a key competitive advantage for the BYD Han model. Firms should maintain their competitive strengths through continuous investment and monitoring. Marketing efforts should highlight the practicality and advancement of intelligent functions via test-drive events and omnichannel campaigns.

Interior directly affects daily user experience. Offering diversified interior styles can meet personalised needs and broaden the target market. It is worth noting that interior lies close to the importance threshold, suggesting that enterprises should invest prudently in maintaining its competitive advantage.

(3)

Resource reduction suggestions

Although BYD Han has gained user recognition in appearance and driving experience, enterprises should reduce investment in these attributes under resource constraints and instead focus on improving overall performance. Regarding the quality–price ratio, user perception is moderate and its impact on satisfaction is limited, so resource allocation need only be kept at a reasonable level.

7. Conclusions

This study proposes an IPA-based framework for analysing consumer satisfaction in the NEV industry. The approach evaluates consumer satisfaction at a fine-grained attribute level by jointly assessing attribute performance and importance, establishing an optimisation hierarchy for targeted improvement strategies. Leveraging online reviews, a DistilBERT-based multilingual sentiment model is employed for sentiment analysis to quantify attribute performance. Attribute importance is estimated using a linear programming model linking attribute-level sentiment values with overall review ratings. Differential weights for positive deviation and prioritised constraints for larger rating gaps improve estimation accuracy and yield interpretable importance measures. For underperforming attributes, word cloud visualisation and chi-square test examine consumer feedback and derive specific improvement suggestions. Cross-platform validation confirms the method’s effectiveness. Overall, this study addresses the lack of fine-grained attribute satisfaction analysis and provides actionable implications for NEV enterprises. The framework is inherently vehicle- and platform-agnostic, and can be applied to online reviews of other vehicle models or platforms.

This study has several limitations. First, product attribute themes are derived from predefined review dimensions on the selected platform. For platforms lacking such structures, attribute extraction would be required beforehand. Future research could explore more generalisable attribute extraction techniques to extend applicability. Second, the dataset was collected from online reviews without explicit screening for authenticity. Fabricated or low-quality reviews may affect result accuracy. Incorporating review validity assessment into the analytical process would improve reliability. Third, user demographics, vehicle categories, regional characteristics and temporal factors may influence satisfaction perceptions. In future research, analysing heterogeneity across different groups may help identify which attributes matter most to each group. Finally, this study applies the traditional two-dimensional IPA model. Although effective, this framework has inherent limitations. Future studies may extend it by introducing additional dimensions and variables to enhance explanatory power and generalisability, thereby providing deeper insights for NEV research and practice.