Introduction

Climate change has become a globally recognized focal point, drawing extensive attention from governments and the international community. In 2020, China introduced the “dual-carbon” goals, committing to peak carbon emissions by 2030 and achieve carbon neutrality by 2060 (Liu et al. 2022). This vision presents significant opportunities and challenges for China’s energy structure transformation. The transportation sector is one of the fastest-growing fields of carbon emissions in China, contributing ~10% to the national total annually (M. Liu et al. 2021). Electric vehicles (EVs), powered by electricity or other clean energy sources and utilizing battery, fuel cell, or other renewable energy technologies, hold immense potential to replace traditional fuel-powered vehicles (Zhang et al. 2020). Consequently, promoting EVs has become a pivotal measure in China’s road transportation sector to enhance energy efficiency, reduce emissions, and accelerate industrial transformation and upgrading (Zeng et al. 2023; Zhang and Qin, 2018). In support of this initiative, multiple government departments, including the Ministry of Finance and the Ministry of Science and Technology, have successively released a series of policies. These policies aim to comprehensively promote the development of EVs by lowering purchase costs, improving charging convenience, fostering technological innovation, and enhancing market access regulation. The implementation of these policies has yielded significant results, leading to a rapid increase in EV sales in China.

In recent years, with the continuous expansion of the EV market and the maturation of technology, the development of EVs has gradually shifted from policy-driven to market-driven. The noticeable deceleration in the growth of EV sales has emerged alongside the reduction in policy subsidies (Wang et al. 2019). To address this challenge, the Chinese EV industry needs to adopt more market-oriented approaches to enhance public awareness and acceptance of EVs, attracting a larger pool of potential users, and facilitating the sustained expansion of the EV market.

Analyzing post-purchase satisfaction among consumers is a crucial aspect of marketing, as it reflects the gap between consumer demands and expectations (Zhao et al. 2023), providing a vital channel for understanding user attitudes. A profound understanding of the factors influencing consumer satisfaction can assist businesses and policymakers in determining directions for improving EVs. This, in turn, allows the formulation of more effective strategies to further enhance consumer satisfaction. With the development of social media, potential EV consumers often rely on online evaluations to gain insights into product attributes and functionalities (Park et al. 2007; Zeng et al. 2023). Consumer satisfaction comments shared through online information platforms can influence the perceptions and expectations of potential consumers regarding specific EV products, thereby impacting their willingness and decision-making processes when considering the purchase of such vehicles (Okada et al., 2019). The improvement of post-purchase satisfaction not only enhances consumer loyalty and repeat purchases (Hasan, 2021; Zeithaml and Bitner, 2000) but also generates positive word-of-mouth, facilitating the acquisition of new users (v. Wangenheim and Bayón, 2007). Accordingly, this paper examines post-purchase satisfaction among EV consumers and the factors influencing it. In the context of subsidy withdrawal, addressing these issues is essential for accurately identifying market penetration pathways for EV development, further enhancing consumer satisfaction, and ultimately driving broader EV adoption.

To accurately unveil consumer post-purchase satisfaction and its driving factors for EVs in China, this study collected Chinese consumer online review data and objective product attributes of vehicles from China’s largest automotive website, Autohome, spanning approximately eight years. Utilizing data mining and sentiment analysis (SA) techniques, the study extracted overall post-purchase satisfaction from a vast number of user comments and identified the subjective product attributes that consumers prioritize. Diverging from most literature that confines its analysis to a few dimensions of vehicle product attributes, this paper employed ERNIE 3.0 (Sun et al. 2021), a natural language processing (NLP) model developed by Baidu—the world’s largest Chinese search engine platform—to study and unearth 28 dimensions of subjective attributes, constructing the most extensive and granular Chinese EV user text review structured database to date. With this database, the study could capture consumers’ more nuanced emotional attitudes and perceptions, enabling a more precise identification of market factors affecting the development of EVs.

Finally, using a machine learning (ML) explainability framework, the study delved into the objective and subjective automotive product attributes that influence consumer satisfaction with EVs. The study ranked the impact strength of automotive attributes both statically and dynamically, aiming to comprehensively understand changes in consumer preferences and demands for various automotive product attributes across different market environments. Unlike previous research, which primarily focused on the fitting accuracy and predictive precision of ML models, this research placed a particular emphasis on model interpretability. In addition to ensuring robust model performance, advanced interpretability techniques—particularly SHapley Additive exPlanations (SHAP)—were integrated to quantitatively assess the marginal contribution of each attribute in predicting consumer satisfaction. This approach not only enhances the interpretability of ML models but also offers prioritized policy recommendations for EV manufacturers, retailers, and government policymakers. In summary, the findings contribute to a deeper understanding of consumer preferences for EV attributes. They serve as empirical evidence to guide government policymakers in adjusting promotion policies for EVs and assist automotive manufacturers in producing EVs that better align with consumer preferences.

The remainder of the paper is structured as follows. Section “Literature review” briefly discusses the motivation and review of prior literature. Section “Methods” presents the methodology and data, while the results and discussion are presented in the section “Results and discussion”. Finally, the section “Conclusion” concludes.

Literature review

Factors influencing the development of EVs

Fostering the growth of EVs is a pivotal measure in mitigating carbon emissions. Against the backdrop of gradually diminishing subsidies, maintaining steady growth of EVs remains crucial. Consequently, scholars have shifted their focus towards a market-oriented perspective to study the factors influencing the development of EVs. Predominantly, existing literature has focused on factors impacting EV sales, categorizing them into three dimensions: environmental factors, consumer-specific factors, and the product attributes of automobiles. Literature considering external factors suggests that crucial elements affecting consumer purchasing decisions include automobile technology, energy prices, charging infrastructure, charging costs, and government incentives (Gnann et al., 2018; Ma, Xu et al., 2019; Shu et al., 2022). On the other hand, research from the perspective of consumer-specific factors identifies personality traits (Irfan and Ahmad, 2021), household income, the number of cars owned by the household (Jensen and Mabit, 2017), and environmental awareness (W. Liu et al., 2021) as significant factors influencing consumers’ decisions to purchase EVs. Studies investigating the inherent product attributes of automobiles have found that features such as price, size, battery specifications, charging capabilities, powertrain type, classification, fuel economy, and brand play pivotal roles in influencing consumers’ purchase decisions (Jena, 2020; Ma, Xu et al., 2019; Wang et al., 2020).

Limitations of sales data and the importance of consumer satisfaction

Previous studies have predominantly focused on factors influencing the sales of EVs to identify market elements critical to their development. However, the implementation of stringent car purchase restrictions across most first-tier cities in China (W. Li et al., 2019; Liu et al., 2020; S.-C. Ma et al., 2017) suggests that EV sales data may not accurately capture consumers’ true needs. In cities such as Beijing, Shanghai, and Shenzhen, prospective EV buyers encounter significant barriers, including license plate quotas and requirements for social insurance or personal income tax payments (He et al., 2018; J. Li et al., 2019; Stephens et al., 2018). These restrictions limit the effective supply of EVs, thereby distorting the actual demand. As a result, conclusions derived from sales data regarding market influence may be biased (Ma, Fan et al., 2019). In contrast, consumer comments, which are typically less affected by market policies, provide a more direct reflection of consumer opinions. With the rapid growth of social media, an increasing number of consumers are freely sharing their views online, making these comments a valuable source of qualitative insights into consumer preferences and pain points (Chen et al., 2004; Cui et al., 2012). Therefore, this study leverages consumer comments as a more reliable source for uncovering the true demands and preferences of EV consumers. User satisfaction refers to a function of perceived quality and pre-purchase expectations (Anderson and Sullivan, 1993). User satisfaction with EVs reveals a gap between consumers’ perceptions and expectations (Zhao et al. 2023), highlighting unmet needs among EV consumers. Examining post-purchase satisfaction and the factors influencing it allows for a deeper understanding of EV consumer preferences and demands. This insight helps the industry clarify the direction for product improvements, enhance product competitiveness, and drive sustained growth in the EV market (Hasan, 2021).

Consumer satisfaction research and methodological advances

Currently, research on consumer satisfaction with EVs can be broadly categorized into two types. One focuses on examining consumers’ satisfaction levels with specific features, such as EV-related policies (Li et al. 2016), charging station location selection (Liu et al. 2018), distribution of charging facilities (Lin and Yang, 2024), charging efficiency (Hajforoosh et al. 2016), and user-perceived mileage (Franke et al. 2017). The other type of research centers on overall consumer satisfaction with EVs. Wang and Liu (2020) collected online comments from different brands of EVs on the Autohome website to assess overall consumer satisfaction. They found that 32% of the sampled consumers were not very satisfied with EVs. Meanwhile, Zhao et al. (2023) conducted a customer satisfaction survey in the EV market and discovered that overall satisfaction was slightly higher among mini-EV users compared to full-size EV users. However, both user groups demonstrated positive overall satisfaction, indicating a favorable attitude among EV users.

Furthermore, scholars have delved into factors influencing consumer satisfaction with EVs. Firstly, consumers’ individual characteristics, such as the number of vehicles in a household (Hamed et al. 2023), consumers’ environmental consciousness (Chu et al. 2019), and the future expectation of EVs (Kwon et al., 2020) impact their satisfaction with EVs. As products serve as the primary means of communication and interaction between consumers and businesses, they play a crucial role in consumer satisfaction. Given that EVs represent an innovative product challenging traditional fuel vehicles, Chu et al. (2019) consider their innovativeness as a significant determinant affecting consumer satisfaction. Additionally, the design, safety performance (J. Ma et al. 2017), price, space, energy consumption (Wang and Liu, 2020), and noise insulation (Zhao et al. 2023) can also impact consumer satisfaction. Furthermore, complementary facilities related to EVs can influence consumer satisfaction. Examples include after-sales service (Sun et al. 2017), charging time (Jedin and Li, 2023), the number of charging stations and charging costs (Kwon et al. 2020).

However, most of the aforementioned studies on consumer satisfaction primarily employ survey questionnaires or simulated experiments to investigate consumer satisfaction. These methods, relying on fixed questionnaire options or specific real-life scenarios, to some extent, may limit consumers’ self-expression, with the possibility of omitted options that consumers may not have considered (Malhotra, 2015). In contrast, comments voluntarily generated by users on social platforms often offer a more genuine reflection of consumer sentiments. Big data has transcended the temporal and spatial constraints inherent in traditional declarative preference research (Büschken and Allenby, 2016), thereby presenting significant opportunities for mining genuine consumer preferences and demands from such commentary. Additionally, it furnishes valuable recommendations for product enhancements (Alsaeedi and Khan, 2019). For instance, Xu et al. (2023) analyzed user-generated content on the Autohome website regarding BYD’s EVs, suggesting specific improvements in the interior, energy consumption, and space for BYD’s EVs. Zhang et al. (2021), by collecting data from various automotive websites and combining it with survey questionnaires, formed a multi-source heterogeneous dataset. Based on this dataset, they conducted an analysis of consumer satisfaction with EVs, revealing significant impacts on satisfaction from product performance and consumer attributes. However, most of these studies have primarily focused on researching a limited number of dimensions of automotive attributes that have already been categorized on websites. To overcome this limitation, this study employs cutting-edge Chinese NLP models to explore more detailed and diverse dimensions of vehicle attributes, allowing for a more accurate revelation of consumers’ latent preferences and needs.

Lastly, current research on consumer satisfaction with EVs has identified important factors influencing satisfaction. However, it falls short of quantifying the specific impact differences of various product attributes on consumer satisfaction. Consequently, providing priority recommendations for the further improvement of EVs is challenging. While some scholars have started exploring this area, such as Chen et al. (2021), who employed large-scale group decision-making methods to determine user demand rankings based on identified user needs, and Xu et al. (2023), who used the AATAE_BiLSTM model to calculate consumer satisfaction and attention scores, categorizing the results into four quadrants to provide prioritized improvement suggestions for BYD’s EVs. However, even in these cases, they did not precisely quantify the marginal impact of each automotive product attribute on consumer satisfaction. This study addresses this gap by utilizing ML models in conjunction with the SHAP interpretable analysis method.

Methods

Data acquisition

A browser-based driver is used in this research, employing the Selenium package, which is programmed in Python (version 3.8.3 used in this study; Van Rossum and Drake, 1995), to simulate user browsing behavior on web pages for the acquisition of webpage source code. Various techniques, including ID, Xpath, and CSS Selector, were employed for data localization, enabling the extraction of objective product attributes and consumer subjective online reviews pertaining to EVs from the Autohome website.

Based on the aforementioned approach, this study gathered a comprehensive dataset from the Autohome website, covering nearly 8 years, from March 2014 to February 2022. The dataset comprises 38,422 instances of subjective online reviews, which include both satisfaction scores and Chinese review texts provided by consumers. These reviews correspond to a total of 1321 distinct EV models. Additionally, 14 objective product attributes for these 1321 models were collected, as they reflect fundamental technical characteristics of the vehicles and are widely recognized as factors most likely to influence consumer perception (Jena, 2020; Ma, Xu et al. 2019; Wang et al. 2020). These objective attributes include powertrain type, multipurpose, luxury brand, vehicle size, battery type, SUV, price, range, maximum speed, total motor power, battery capacity, warranty period, warranty mileage, and launch date. Detailed variable descriptions and descriptive statistics for the objective product attributes are provided in Tables S1 and S2, respectively, of the Data section in the attached Supplementary Information.

SA and structured processing of review texts

SA is an NLP technology used to analyze opinions, attitudes, and emotions toward specific entities. In the context of consumer product reviews, SA helps identify key product attributes that consumers focus on and assess their sentiments (such as positive, negative, or neutral) toward these attributes.

In this study, we employed the ERNIE NLP model to extract subjective product attributes and their corresponding sentiment polarity from consumer online review texts. The ERNIE model was selected for its exceptional ability to process Chinese text, leveraging extensive pre-training on diverse Chinese language data. Moreover, its “comment opinion extraction” feature, trained on review texts from 12 industries, including the automotive sector, ensures precise and relevant analysis of Chinese automobile consumer reviews. Specifically, we input a total of 38,422 online review texts into the ERNIE model. The model then returns the results, which are in GBK-encoded JSON format, comprising four elements: attribute word (noun), descriptive word (adjective), sentiment polarity (0 for negative, 1 for neutral, 2 for positive), and short phrases. An example output from the model might be: {“items”: [{“prop”: “appearance”, “adj”: “stylish”, “sentiment”: 0, “abstract”: “appearance is stylish”}]}. In this context, “prop” refers to the product attribute that consumers focus on, while “sentiment” indicates the corresponding sentiment polarity. The specific steps for implementing the ERNIE model with a local Python script are provided in Supplementary Information, Section S2.1.

Given the variability in language use across consumer reviews, consumers often describe the same attribute using different terms. To ensure consistency and clarity, we grouped the subjective product attributes extracted from the ERNIE model into 28 categories based on shared meanings and attribute descriptions. This consolidation of varied expressions into coherent categories allowed us to provide a comprehensive yet simplified representation of the attributes discussed by consumers. These categories include car paint, texture, handling features, handling agility, handling stability, steering control, braking performance, driving comfort, air conditioning performance, interior smell, interior color, legroom, exterior design, overall appearance, power, handling, handling precision, touch, shock absorption effect, noise reduction capability, interior materials, seating space, storage capacity, interior craftsmanship, interior design, passenger comfort, overall comfort, and overall space.

To better capture the intensity of consumer sentiment and amplify these sentiments across the dataset, we weighted the subjective product attributes based on the frequency of sentiment polarity occurrences within each review. Specifically, if a sentiment polarity associated with a subjective product attribute is positive, the attribute is assigned a score of 2 if it appears once, 4 if it appears twice, and 6 if it appears three or more times. Similarly, negative sentiment is scored inversely, while neutral or unmentioned attributes are scored 0. After structured processing, the attributes are categorized into seven sentiment polarities, ranging from very satisfied to very dissatisfied, with corresponding scores of 6, 4, 2, 0, –2, –4, and –6, respectively. The assignment method is shown in Table 1. This method allows for a nuanced understanding of consumer sentiment while preserving the original data distribution. The process maintains the relative order of sentiments by applying proportional expansion locally within each review, ensuring that no new relationships are introduced and the original distribution characteristics are preserved. In a large dataset like ours, with 38,422 reviews, local frequency variations tend to balance out, with weighted results regressing toward the mean, minimizing global distribution distortion. The variable descriptions and descriptive statistics for the 28 categories of subjective product attributes are detailed in Tables S1 and S2 in the Data section of the Supplementary Information.

Table 1 Structured processing method for sentiment polarity of subjective evaluations in EVs.
Full size table

ML explainability framework

ML, as a core technique in the field of artificial intelligence and data science, is highly regarded for its ability to efficiently model complex relationships between multivariate input features and output targets. However, a common issue with ML algorithms is their lack of interpretability and “black-box” nature, which severely limits their widespread application in real-world tasks. To address this limitation, many explanation methods have been proposed. SHAP, an additive feature attribution method, measures the impact of each feature on the predictions of the output variable by calculating the marginal contributions (SHAP values) of each feature to each possible feature subset.

Compared to traditional ML explainability methods, SHAP draws on the equity theory in cooperative games, allowing it to calculate the marginal contributions of features while fully considering the interactions among variables. This approach benefits from a more comprehensive theoretical foundation and more rigorous mathematical formulas. Such capabilities enhance our understanding of ML models’ decision-making processes and facilitate the analysis of their results. Consequently, this article introduces the ML-SHAP method to analyze the marginal impact of various EV attributes on consumer satisfaction, revealing consumers’ most genuine demands. The core computational principles of the SHAP method are provided in Supplementary Information, Section S2.2.

Although ML algorithms significantly outperform traditional statistical models, previous studies have argued that ensemble models perform better than individual approaches (Du et al. 2020). Ensemble models improve the overall prediction performance and generalization ability of ML systems by combining the predictions of multiple base models. Hence, certain representative ensemble methods, including eXtreme Gradient Boosting (XGBoost), light gradient boosting machine (LightGBM), and random forest (RF), have been widely applied in various prediction scenarios. In this study, we also apply these ensemble models to build a predictive model for consumer satisfaction.

XGBoost (version 1.5.2 used in this study; Chen and Guestrin, 2016) is an optimized gradient boosting algorithm that iteratively adds new models to correct the residuals of the previous model, progressively improving overall prediction accuracy. The process involves initializing model predictions, calculating residuals, training new decision trees to fit these residuals, and updating predictions with weighted learning rates until the desired accuracy is achieved. Similarly, LightGBM (version 3.1.1 used in this study; Ke et al. 2017) improves computational efficiency by converting continuous features into discrete histograms and using a “leaf-wise” growth strategy for tree construction. It further improves performance by selecting high-gradient samples through Gradient-based One-Side Sampling (GOSS) and incrementally adding new trees. On the other hand, RF (implemented with sklearn version 1.0.2 in this study; Breiman, 2001) takes a different approach by generating multiple samples using bootstrap aggregation, building numerous decision trees, and randomly selecting features at each split. The final prediction is then made by majority voting in classification tasks, ensuring robustness and accuracy. For detailed computational principles and key formulas of each algorithm, refer to Supplementary Information Sections S2.3, S2.4, and S2.5.

Specifically, this study used 14 objective product attributes and 28 subjective product attributes of EVs as input features and the satisfaction score as output to build a predictive model for consumer satisfaction. However, in the empirical analysis of real-world data, observations are often affected by stochastic error or noise. When ML algorithms endeavor to discern the relationships among variables, an excessive focus on fitting the data may inadvertently lead to the learning of idiosyncratic, non-generalizable patterns, manifesting as overfitting. To mitigate this issue, we transformed the continuous variable of consumer satisfaction scores into two discrete categories: a value of 0 indicates a low satisfaction level, while a value of 1 signifies a high satisfaction level. By grouping these values into predefined intervals, the impact of noise and outliers is substantially reduced. As demonstrated by Hu et al. (2009) and Liu et al. (2002), discretization significantly improves both the training speed and accuracy of ML methods. Decision tree-based algorithms (such as XGBoost, LightGBM, and RF utilized in this paper), when applied to discrete values, tend to produce more compact and accurate models. Additionally, discretization simplifies the model structure, enabling it to better capture essential trends and detect significant patterns more effectively, without being influenced by outliers or noise. This improvement better aligns with the goals of exploring factors influencing consumer satisfaction in this study. The discretization process will be explained in the section “Analysis of the distribution of sentiment polarity and consumer satisfaction”, where the discretization outcomes based on the actual distribution of consumer satisfaction scores are reported.

Secondly, to further mitigate overfitting and verify the generalization capabilities of each model, we partitioned the dataset into 80% for training and 20% for testing. This approach enabled systematic training and evaluation of the models. During the training stage, we optimized model performance through a comprehensive grid search combined with 10-fold cross-validation. This process systematically explored various hyperparameter combinations to determine the optimal settings. Specifically, we first defined the possible value ranges for each hyperparameter and then validated them through 10-fold cross-validation. In this process, the original training set was further divided into 10 equal parts; each time, 9 parts were used for training, and the remaining 1 part was used for validation. This process was repeated 10 times, rotating the validation set each time to ensure that model performance did not rely on any particular data subset.

The tuned hyperparameters and their corresponding grid search ranges for the ML models are detailed in Section S2.6 of the Supplementary Information. For each hyperparameter combination, we trained the model on the training folds and evaluated its performance on the validation fold, then averaged the performance across all folds to obtain a reliable estimate for that combination. We ultimately selected the best-performing hyperparameter combination from cross-validation and trained the final model on the entire training set.

After optimization, we conducted a comprehensive performance evaluation of the final selected model on the test set, with evaluation metrics, including accuracy (ACC), area under the curve (AUC), precision, recall, and F1-score. The specific calculation principles and formulas for these evaluation metrics are detailed in Section S2.7 of the Supplementary Information. We then applied the SHAP method to interpret each decision made by the model. By calculating SHAP values, we quantified the contribution of each feature to the model’s output and ranked the input features in both static and dynamic terms based on their SHAP values, providing an in-depth analysis of how various features impact EV consumer satisfaction. Figure 1 presents the methodological framework of this study.

Fig. 1: Overview of the methodological framework.
figure 1

In step (1), this study uses Autohome, an online automotive information-sharing platform, as the data source. In step (2), two types of data are gathered: subjective online reviews (comprising review text and satisfaction score) and objective product attributes. In step (3), the ERNIE model is used to extract the subjective product attributes that consumers focus on and their corresponding sentiment polarity. In step (4), statistical analysis of the structured subjective data is conducted, and machine learning with SHAP is applied to identify the most impactful subjective and objective product attributes on consumer satisfaction. In step (5), recommendations are proposed to advance the EV industry and guide the formulation of government incentive policies.

Full size image

Results and discussion

Analysis of the distribution of sentiment polarity and consumer satisfaction

Aiming to gain insights into consumers’ emotional attitudes, we conducted a statistical analysis on the frequency of sentiment polarities associated with the subjective product attributes of concern to EV consumers. As illustrated in Fig. 2a, the frequency of positive sentiment polarities (2, 4, 6) far exceeds that of negative sentiment polarities (–2, –4, –6). Furthermore, the majority of consumer evaluations fall into the categories of ‘satisfied’ and ‘quite satisfied’, while occurrences of ‘very unsatisfied’ and ‘neutral’ are notably rare. This suggests that the subjective product attributes of EVs generally meet consumers’ basic expectations and needs. Overall, consumers demonstrate a positive and favorable emotional inclination.

Fig. 2: Distribution of consumers’ subjective reviews.
figure 2

a Sentiment polarity distribution of subjective product attributes. b Distribution of EV consumer satisfaction scores.

Full size image

In terms of overall satisfaction among EV consumers (as shown in Fig. 2b), scores of 33, 34, and the perfect score of 35 overwhelmingly dominate the satisfaction ratings. Together, these three scores constitute ~80% of the total frequency, while scores below 32 are notably infrequent. This indicates that the majority of consumers provide high overall satisfaction ratings for EVs, and these satisfaction scores are highly concentrated, reflecting a consensus among most individuals regarding their views on EVs.

However, it is noteworthy that during the rating process, consumers tend to employ a deduction approach. In other words, if EVs fail to meet certain expectations or exhibit shortcomings in certain aspects, consumers apply a deduction system, subtracting points from the maximum score of 35. As a result, consumer ratings typically cluster in the range of moderate to high scores. Therefore, although the majority of evaluations fall within the high range (33–35 points), it does not necessarily indicate that existing EVs entirely meet consumer expectations. On the contrary, it implies that certain attributes still fail to fulfill consumer needs, leaving room for improvement. Thus, the overall satisfaction scores align with the distribution of sentiment polarities in this study. The majority of consumers maintain a satisfied or quite satisfied stance towards EVs.

As mentioned in the section “ML explainability framework”, the discretization of consumer satisfaction scores is based on their actual distribution. Observing the distribution in Fig. 2b, 34 points serve as a “turning point” in the data, acting as a natural demarcation. Around this threshold, the distribution characteristics change significantly, which may indicate a noticeable shift in consumer satisfaction. Therefore, in this study, scores of 34 and above are categorized as the high-score region for consumer satisfaction, assigned a value of 1, while scores below 34 are classified as the low-score region, assigned a value of 0.

Subjective product attributes of interest to consumers

Through a statistical analysis of the frequency of 28 subjective product attributes within consumer review texts, we generated a keyword cloud depicting key attributes (Fig. 3). This visualization reveals the pivotal factors that capture consumers’ attention in their experience with EVs. Notably, the most frequently mentioned terms in consumer evaluations are overall space, appearance, handling, driving comfort, and power. These subjective product attributes emerge as the primary focal points of consumer attention and perception in their use of EVs.

Fig. 3: Keyword cloud of subjective product attributes.
figure 3

a Chinese corpus. b English corpus. In each figure, word size is proportionate to frequency, highlighting the key subjective product attributes that capture consumer attention in their EV experiences.

Full size image

Firstly, overall space is the primary concern for consumers when using EVs. The spatial design of EVs directly impacts the overall driving experience, making it the most noticeable subjective product attribute for consumers. However, overall appearance, as the second most significant attribute of interest, has been largely overlooked in previous literature. We believe that an eye-catching exterior provides consumers with a better esthetic experience and can boost driver confidence, thus enhancing the overall driving experience. Consumer evaluations of overall appearance convey attitudes toward vehicle design, esthetics, and brand image. Handling, the third major focus, reflects consumers’ emphasis on the driving experience, while driving comfort is the fourth key attribute that concerns consumers. As cars have become indispensable in daily life, driving comfort is increasingly valued. Lastly, power reflects consumer expectations for efficient and stable power output, underscoring the desire for performance and reliability in EVs.

Marginal impact analysis of automotive product attributes on consumer satisfaction

Table 2 presents the classification performance of three ML ensemble models. Detailed receiver operating characteristic (ROC) curves and confusion matrices are provided in the section S3 of the Supplementary Information. These models demonstrate comparable predictive capabilities and solid classification performance across the board. However, the XGBoost model stands out in terms of overall performance, achieving an AUC of 84%, with an accuracy and precision of 77%, a recall of 84%, and an F1 score of 80%. This indicates that XGBoost effectively captures the nonlinear relationships between variables, exhibiting not only strong predictive performance but also robust generalization abilities. Therefore, this model can be combined with the SHAP method to provide valuable interpretability insights.

Table 2 Classification performance comparison across ML models.
Full size table

Based on the selected optimal predictive model, XGBoost, the SHAP method was incorporated to analyze the factors influencing consumer satisfaction with EVs. Figure 4a illustrates the feature importance ranking obtained through the XGBoost-SHAP explainability method. The horizontal axis represents SHAP values, with positive values indicating a positive impact and negative values indicating a negative impact on the output variable. The vertical axis displays features ranked by importance, as determined by their cumulative SHAP values. Notably, features with the same value may exhibit different SHAP values across individuals, with some being positive and others negative. Examining the relationship between feature values and SHAP values provides a deeper understanding of the non-linear dynamics between each feature and consumer satisfaction.

Fig. 4: Feature importance results of different methods.
figure 4

a Top 20 important features based on SHAP values. b Top 20 important features based on the XGBoost total gain method. In panel (a), purple and green indicate higher and lower feature values, respectively. The horizontal axis represents the SHAP values, where positive values indicate a positive impact and negative values indicate a negative impact on the output variable. The vertical axis displays features ranked by their cumulative SHAP values. In panel (b), features are arranged by importance according to the XGBoost total gain method.

Full size image

It is essential to highlight that XGBoost inherently offers a method for assessing feature contributions. To assess the robustness of the SHAP ranking results with respect to different feature importance methods, we also present the outcomes of XGBoost based on the total gain feature importance calculation method, shown in Fig. 4b. Upon scrutiny, we observe that the feature importance outcomes of XGBoost align with the conclusions drawn from SHAP. Despite the substantial differences in their calculation principles, these two methodologies yield remarkably consistent feature importance rankings. This implies that the SHAP ranking results exhibit relative insensitivity to diverse feature importance calculation approaches, showcasing high stability and reliability.

Figure 4a illustrates the top 20 product attributes most correlated with consumer satisfaction. With the exception of price, all other attributes show a positive correlation with consumer satisfaction. From the attribute rankings, it is clear that the primary factor influencing consumer satisfaction with EVs is total motor power, with range and warranty mileage also appearing in the top five. Additionally, the research identified launch date and price as significant factors, ranking among the top three in terms of impact on consumer satisfaction. The subjective product attributes that most influence consumer satisfaction include overall space, exterior design, handling, overall comfort, and interior design, which can be grouped into the categories of space, design, handling, and comfort. This aligns with the findings from the word cloud analysis of consumer reviews, highlighting a consistent focus on these key subjective attributes.

Total motor power, a distinctive product attribute of EVs, is the most critical factor influencing consumer satisfaction. The electric motor, one of the three core components of EVs, directly reflects the vehicle’s power performance, constituting a crucial technical parameter. EVs with higher motor power offer faster acceleration, providing passengers with more robust and stable power support, thereby enhancing the driving experience. Consequently, consumers of EVs with higher power are more likely to attain greater satisfaction. Range, as the second most influential automotive performance parameter affecting consumer satisfaction, signifies the prevalent range anxiety among EV consumers. Range directly impacts both the convenience of vehicle use and the frequency of charging. Vehicles with longer ranges offer greater travel distances and require fewer charging cycles, reducing the time and opportunity costs associated with waiting for charging and enhancing overall convenience. Similarly, warranty mileage is a crucial product attribute influencing consumer satisfaction. A longer warranty mileage implies that, during the warranty period, vehicle owners can enjoy free maintenance services without additional costs. This effectively alleviates the maintenance burden on owners, reducing the overall cost of vehicle use. For EVs, longer range and warranty mileage contribute to a heightened sense of driving security and convenience, thus delivering an improved user experience.

The launch date also emerges as a crucial factor influencing consumer satisfaction. It reflects broader market conditions that influence various aspects of vehicles, such as performance, technological advancements, and market maturity. Vehicles introduced at a later stage often incorporate more advanced technologies, benefiting from the opportunity to learn from the failures of earlier models. This allows for meticulous optimization and adjustments that better align with consumers’ driving preferences. Additionally, over time, consumer acceptance of EVs has increased, meaning that newer EV models are more likely to receive higher satisfaction ratings from consumers. Furthermore, price is another essential factor. Price is directly related to consumers’ budget constraints. According to the theory of diminishing marginal utility and consumer surplus, as prices increase, the satisfaction derived from the same level of consumption tends to decrease. In the EV market, consumers evaluate the price they are willing to pay for a specific level of performance, comfort, and technology. When the price is lower than a consumer’s willingness to pay, greater post-purchase satisfaction is typically achieved.

Moreover, this study has uncovered intriguing and significant findings. Contrary to the intuition that alleviating range anxiety requires larger battery capacities, the results reveal that consumers have a limited perception of battery capacity. This is evident in the driving satisfaction predictions, where battery capacity does not exhibit a high marginal contribution, and larger capacities do not necessarily lead to greater consumer satisfaction. The study suggests that range, rather than technical parameters like battery capacity, serves as a more intuitively perceived indicator for consumers. Consumers tend to prioritize performance aspects they can directly relate to, such as range, over technical specifications. Furthermore, excessively large battery capacities may contribute to higher vehicle costs, increased risks of battery safety incidents, and potentially more severe environmental impacts during the battery recycling process (Dik et al. 2022). Thus, larger battery capacities may conflict with factors that consumers prioritize, such as price—ranked third in terms of marginal impact on consumer satisfaction—and concerns about vehicle safety. Additionally, this finding aligns with the environmental consciousness observed among Chinese consumers (Chu et al. 2019). Therefore, marketing EVs by emphasizing range over battery capacity may better align with consumer interests.

Furthermore, EV manufacturers should enhance consumer satisfaction with the vehicle’s overall appearance through improved exterior design. As revealed in the analysis in the section “Subjective product attributes of interest to consumers”, the overall appearance of EVs is the second most significant factor that consumers pay attention to and perceive during the vehicle usage process. However, by combining these findings with the results in Fig. 4a, the research can conduct a more detailed analysis. The study observes that the impact of overall appearance on consumer satisfaction is asymmetrical. In other words, an exceptionally attractive appearance is not essential for consumers—a visually appealing appearance does not necessarily result in a substantial increase in overall satisfaction. However, the necessity lies in ensuring that the appearance is “not ugly”—an unattractive appearance can lead to a sharp decline in overall consumer satisfaction. Therefore, the design of EVs does not need to be overly outstanding, but it should at least ensure that the overall appearance meets an average standard. Moreover, the study finds in Fig. 4a that exterior design contributes much more to consumer satisfaction than overall appearance. This suggests that while consumers care about appearance, exterior design is the key factor influencing satisfaction. Therefore, if EV enterprises aim to enhance overall appearance to improve consumer satisfaction, the key lies in improving exterior design.

An interesting insight emerges regarding the overall appearance of vehicles. By comparing the word cloud in the section “Subjective product attributes of interest to consumers” (Fig. 3) with Fig. 4a, it can be observed that while consumers frequently discuss overall appearance, its marginal impact on consumer satisfaction is lower than that of objective product attributes, such as total motor power, price, and range. This may suggest that the current offerings in the Chinese EV market already meet consumers’ expectations for appearance. To further enhance consumer satisfaction and boost market competitiveness, EV companies should focus on achieving technological breakthroughs in other product attributes.

Finally, consumers seem to have a lukewarm attitude toward the multipurpose attributes of EVs. Many automotive companies emphasize multi-scenario applications of EV products, such as serving as a power source during camping, featuring a meditation mode, or offering a one-button bed transformation, to highlight the product’s intelligence. However, as shown in Fig. 4a, this particular attribute does not significantly enhance consumer satisfaction. This could be due to the fact that these features primarily serve as “smart” marketing elements, while their actual functionality often falls short of meeting the expectations and genuine needs of various consumer groups, which may lead to a lower frequency of use. Consequently, this attribute has a limited impact on the actual user experience, as reflected in its minimal influence on consumer satisfaction and the inconsistent attitudes consumers display toward it. To address this, EV manufacturers should refine their marketing strategies for multi-scenario applications, aligning them more closely with consumers’ practical needs to genuinely enhance satisfaction. Alternatively, given limited resources, EV companies should prioritize improving other vehicle attributes that have a higher marginal impact on consumer satisfaction to achieve maximum and more immediate improvements in consumer satisfaction.

Temporal trends in the importance of key product attributes

Based on the analysis in the previous section, the launch date has been identified as a crucial factor influencing consumer satisfaction with EVs. The development of China’s EV industry over the past decade has been primarily driven by government policies and subsidies. Different launch dates not only reflect different levels of maturity in vehicle performance but also represent distinct macro-market environments. Analyzing the impact of various vehicle attributes on consumer satisfaction according to the launch date is essential for understanding the dynamic changes in consumer preferences and demands across different market conditions. This approach allows for an examination of how consumer preferences and needs for EV attributes evolve in response to the changing market landscape.

Excluding the launch date, the top five objective product attributes influencing consumer satisfaction with EVs are total motor power, price, range, warranty mileage, and whether the vehicle is an SUV. Figure 5a illustrates the temporal trends of these five major objective attributes. Overall, with the exception of total motor power and the SUV classification, the impact of other objective attributes on consumer satisfaction remains relatively stable. Nevertheless, the graph reveals some interesting insights.

Fig. 5: Temporal trends of the key product attributes.
figure 5

a Temporal trends of the five core objective product attributes. b Temporal trends of the five core subjective product attributes. In both panels, “2015 and before” on the horizontal axis represents the period from March 2014 to December 2015, and “2021 and beyond” represents the period from January 2021 to February 2022. Each time point on the horizontal axis denotes the average SHAP value over the respective time range.

Full size image

With the widespread implementation of subsidy withdrawal policies for Chinese EVs in the second half of 2019, the marginal impact of EV prices on consumer satisfaction surpassed that of total motor power, rising to the top position as the most sensitive factor for consumers. This finding underscores the critical role that government subsidy policies have played in promoting the EV market. The implementation of these subsidies initially reduced the economic barriers for consumers to experience EVs, thereby facilitating the growth of the Chinese EV market. However, as the subsidies were phased out, price once again emerged as the primary concern for consumers.

In addition, Fig. 5a reveals consumers’ continuous focus on range. Over the past 8 years, the increasing emphasis on range highlights a persistent concern with addressing range anxiety. The results indicate that, despite the rapid growth of the EV market, concerns about range remain one of the primary obstacles to the broader development and adoption of EVs. Furthermore, it is noteworthy that the SUV attribute was the most influential factor in consumer satisfaction prior to 2016. This trend aligns with the surge in popularity of SUV models among Chinese automotive consumers around that time. However, as the “SUV trend” waned, consumer interest in this attribute noticeably declined.

Similarly, Fig. 5b presents the temporal trends of the key subjective product attributes. Overall, the trends for these five subjective attributes are relatively consistent. However, it is noteworthy that between 2018 and 2019, the impact of these subjective attributes on consumer satisfaction reached a low point. Comparing this with Fig. 5a over the same period reveals that, with the exception of price, all other objective product attributes exhibited an upward trend, even displaying noticeable peaks. This suggests that, during this time, consumers were more focused on the objective attributes of automobiles than on the subjective ones. This shift in emphasis may have been influenced by the tightening of subsidy policies. Since EV subsidy policies are often directly linked to objective product attributes, the increase in subsidy thresholds and the gradual withdrawal of subsidies may have prompted consumers to prioritize these attributes to maximize their policy benefits. Consequently, the period from 2018 to 2019 marked a trough for subjective product attributes, during which consumers placed a greater emphasis on the economic practicality of their vehicles.

Furthermore, by comparing the temporal trends of objective and subjective attributes, the study found that the overall space’s impact on consumer satisfaction follows a trend remarkably similar to the objective attribute SUV. Although the popularity of SUV models declined after 2020, consumer expectations for ample space remain high. This suggests that consumers’ core attraction to SUV models is rooted more in the desire for spaciousness than in the model type itself. Additionally, the study notes that, after 2019, while the marginal contribution of various objective attributes has gradually stabilized, consumers’ interest in subjective attributes has shown an overall upward trend. This implies that, over time, EVs have undergone substantial optimization and improvements in basic performance, sufficiently meeting consumers’ fundamental needs. As the EV market continues to develop, consumers are increasingly seeking additional value in the form of personalization, comfort, and esthetic appeal, reflecting a shift toward more diverse transportation needs. This trend signals the EV market’s evolution toward greater maturity and diversity, presenting manufacturers with opportunities for innovation and differentiation to better meet the varied preferences of different consumer groups.

Conclusion

This study collected objective product attributes and consumer online review data for 1321 EV models on the Autohome website between 2014 and 2022. Using data mining and SA techniques, the study assessed overall consumer satisfaction with EVs, gaining insights into the subjective product attributes that consumers are most concerned about. By employing an advanced ML explainability framework, the study pinpointed the most impactful objective and subjective product attributes on consumer satisfaction and ranked their impact intensity both statically and dynamically.

Compared to prior studies, this research employs an ML explainability framework to quantitatively analyze the marginal effects of various automotive product attributes on consumer satisfaction. It delves into the finer preferences and demands of consumers, offering a more thorough analysis of how these preferences and demands evolve over time. The study reveals that current consumers of EVs in China generally hold either satisfied or very satisfied attitudes toward their EVs. Objective product attributes unique to EVs, such as total motor power and range, exert the most significant influence on consumer satisfaction. Regarding subjective product attributes, spatial considerations garner the most attention and perception from consumers, followed by design, handling, and comfort, all of which significantly impact consumer satisfaction.

Furthermore, this study uncovers several nuanced yet crucial findings: (1) range anxiety persists despite gradually increasing consumer satisfaction as the EV market matures; (2) while battery capacity largely determines range, consumers prioritize range closely associated with their actual usage experiences over battery capacity, with excessive battery capacity even eliciting dissatisfaction; (3) EV exterior design need not be excessively outstanding, but it must at least meet average esthetic standards, as poor EV design strongly evokes consumer dissatisfaction. If EV companies aim to enhance satisfaction through design, they should allocate more resources to improving exterior esthetics; (4) price remains an important influencing factor, especially after the government implemented large-scale subsidy policies, making it the most sensitive factor for EV consumers.

Against the backdrop of the gradual transition from policy-driven to market-driven development in the EV sector, the results of this study hold significant implications for gaining a deeper understanding of consumer preferences regarding EV attributes. This can assist relevant authorities and automotive companies in adjusting and optimizing their strategies for the market promotion of EVs. Specifically, the objective attributes of EVs are crucial, with special emphasis on total motor power and range. Meanwhile, subjective attributes such as vehicle space, design, handling, and comfort should not be overlooked. As objective attributes of automobiles become increasingly homogeneous, automotive manufacturers should pay greater attention to subjective consumer demands, providing more value-added vehicles to enhance the market competitiveness of EVs. At the government level, closely aligning with industry development, encouraging green innovation, and expanding infrastructure, such as charging stations, are key to driving the entire EV industry toward higher quality and greater environmental sustainability.

Due to the delayed popularity of the internet in China, early EV online review data is relatively limited, resulting in an imbalance in sample sizes across different years. Additionally, this study primarily focuses on online reviews and may not fully consider offline consumer evaluations of EVs, potentially introducing sample selection bias. The prevalent issue of the “silent majority”, where the voices of most consumers are not adequately recorded and analyzed, exacerbates this problem, potentially impacting the comprehensiveness of the study’s conclusions. Furthermore, the lack of access to private data, such as consumer gender and age, hinders a thorough examination of potential influences from diverse demographic characteristics on the study results. Future research is expected to address these limitations, offering a more comprehensive and in-depth analysis of consumer satisfaction with EVs.