In this paper, we develop a new multi-attribute decision support framework for electric vehicle purchase decisions by exploiting the rich textual information contained in online consumer reviews. The rapid growth of the electric vehicle market has increased the difficulty of evaluating competing products, so consumers frequently seek electronic word-of-mouth from multiple automotive platforms. However, unstructured online comments are diffuse, heterogeneous, and difficult to compare. In order to address this challenge, we quantify free-form text reviews as probabilistic linguistic term sets, aggregate multi-source website evaluations by a webpage-credibility weighting procedure, derive attribute weights through an improved ITARA method enriched with prospect-theory parameters, and finally extend the MARCOS ranking logic to the probabilistic linguistic environment. A case study based on real electric vehicle reviews gathered from two popular Chinese automotive websites is presented to demonstrate the usefulness of the proposed model. Sensitivity analyses on the risk coefficients and comparisons with classical PL-TOPSIS and PL-TODIM ranking methods are also implemented. The results show that the proposed model is both scientifically valid and practically valuable for supporting consumer-oriented electric vehicle selection.
1. Introduction
Energy consumption around the world has continued to rise, and the transportation sector is one of the largest contributors. This concern has encouraged many countries to promote sustainable mobility. In particular, electric vehicles have become one of the most promising solutions because of their high efficiency and low tailpipe emissions. In recent years, favorable policy environments, rapid technological innovations, and declining battery costs have accelerated the electrification of the automotive industry. Indeed, the market share of electric vehicles in new car sales has increased sharply, and this trend is especially evident in China, where millions of battery-electric vehicles are sold every year.
Although the increasing number of electric vehicle models provides more choices, it also produces a serious information-overload problem. Consumers are not only concerned about price and range, but also about driving control, interior quality, intelligent configuration, spaciousness, comfort, and after-sales experience. Therefore, the electric vehicle purchase process can be characterized as a typical multi-attribute decision-making (MADM) problem. Because most customers lack professional mechanical knowledge, online reviews left by previous owners have become a major source of decision support. A large proportion of prospective electric vehicle buyers consult online forums, review platforms, and social media before making a final purchase decision. Positive comments can significantly increase purchase intention, while negative remarks can dramatically discourage consumers. E-commerce platforms and automotive vertical portals thus serve as important intermediaries between vehicle manufactures and potential customers.
Nevertheless, interpreting online reviews is not straightforward. The number of comments for a single electric vehicle model can be enormous, and their contents are colloquial, multi-dimensional, and often inconsistent. Star ratings only provide aggregate satisfaction scores and cannot accurately represent specific attributes. Textual reviews, by contrast, convey strong semantic and sentiment information. For example, a consumer may write that an electric vehicle has “fast acceleration but limited rear space,” which is impossible to encode in a five-star rating system. Therefore, an automated and structured analysis of review text is needed to enable fair comparisons among electric vehicle candidates.
Previous research has examined product ranking based on online reviews, especially in the hospitality domain. However, the number of studies dedicated to electric vehicle selection is relatively small. In particular, conventional studies often transform online comments into only three coarse sentiment grades, such as positive, neutral, and negative. This coarse granularity loses much valuable decision information. In addition, the operational details related to website trustworthiness, subjective consumer risk attitude, and robust multi-criteria aggregation are often ignored in existing frameworks. To fill these gaps, this paper answers four research questions:
- How can online text comments regarding electric vehicles be transformed into a fine seven-grade probabilistic linguistic representation?
- How can the weights of multiple online information sources be determined by considering both the quantity and quality of reviews?
- How can attribute weights be derived objectively while reflecting consumer loss aversion and risk preference?
- How can the MARCOS method be extended to the probabilistic linguistic environment in order to generate an accurate and stable ranking of electric vehicle alternatives?
In response to these questions, we construct an end-to-end decision model. First, Python-based crawlers are used to capture a large number of user comments for a set of candidate electric vehicles. The textual data are cleaned, segmented, filtered, and aggregated into eight core product attributes through word-frequency analyses. A sentiment-analysis engine computes a Likert score for each attribute-related keyword, and the scores are mapped to a seven-level probabilistic linguistic term set (PLTS). Second, to fuse comments from multiple websites, we calculate a webpage weight based on the number of reviews as well as an evaluation-consistency measure. The intuition is that a website with more comments should be more influential, while a website whose comments display larger disagreement should be considered less reliable. Third, the ITARA (Indifference Threshold-based Attribute Ratio Analysis) method is improved by incorporating prospect-theory value functions when determining the indifference threshold, thus capturing the psychological bias of consumers. Finally, the MARCOS method is extended to the probabilistic linguistic domain so that both the proximity to the ideal solution and the distance from the anti-ideal solution are optimized in a utility-based ranking framework.
The remainder of this paper is organized as follows. Section 2 reviews relevant studies. Section 3 recalls the basic concepts of probabilistic linguistic term sets, ITARA, prospect theory, and MARCOS. Section 4 proposes the complete electric vehicle purchase decision model. Section 5 illustrates the model by a case study using real online reviews. Section 6 reports sensitivity and comparative analyses. Section 7 concludes the paper and discusses limitations.
2. Literature Review
2.1 Product Ranking Based on Online Reviews
A considerable body of literature has used online reviews to support product ranking and recommendations. Current approaches can be generally classified into three categories. The first category concentrates on free-text contents. In these studies, product features are extracted first and then sentiment polarity or numerical sentiments are assigned to each feature. For instance, fuzzy inference, sentiment dictionaries, and hesitant fuzzy sets have been widely employed to convert text into preference values. Many researchers extended traditional MADM methods such as TOPSIS, TODIM, and VIKOR to the online-review context. This category of research demonstrates that textual reviews are capable of revealing multi-faceted product strengths and weaknesses that cannot be reflected by coarse overall ratings.
The second category concentrates on numerical online ratings. Several scholars use the percentage distribution of star scores for each product attribute, then apply stochastic dominance rules or regret theory to rank competing items. Online ratings are easy to obtain and compare, and they provide credible quantitative signals. However, they suffer from insufficient semantic detail. The third category combines textual contents with numerical ratings. For example, graph-based models and sentiment-aware PageRank methods have been proposed to fuse heterogeneous information. Such hybrid approaches usually outperform models that rely on a single data source because the complementary nature of text and ratings reduces bias and improves robustness.
Despite these advances, most product-ranking studies still target hotels, restaurants, electronic devices, or general consumer goods. The application of online-review-based decision support for electric vehicle purchase remains sparse. Given the unique complexity of electric vehicle evaluation, more domain-specific research is necessary.
2.2 Retail-Oriented Electric Vehicle Research
The electric vehicle literature from the retail perspective can be divided into three streams. The first stream explores factors that shape purchase intention. Studies show that perceived environmental benefits, government subsidies, charging infrastructure, social influence, and consumer demographics all play important roles. Some scholars construct integrated models using structural equation models or hybrid choice models, while others apply meta-analysis to summarize prior empirical findings.
The second stream focuses on sales forecasting. Researchers have applied total-cost-of-ownership models, sentiment analysis of social-media content, grey models, principal-component neural networks, and systems-of-systems frameworks to predict electric vehicle sales. Correct sales forecasts are important for supply chain planning and policy making.
The third stream deals with direct product selection. For example, several studies compare electric vehicles using criteria such as acceleration, price, battery capacity, and range. Multi-attribute techniques including entropy, TOPSIS, MARCOS, and other compensatory methods have been tested. More recently, scholars have combined online review sentiment with fuzzy MADM to evaluate battery electric vehicles. However, such studies are still relatively rare and often omit the decision-maker’s psychological risk attitude. We therefore position our work precisely at the intersection of online review mining, MADM, and electric vehicle selection.
2.3 Probabilistic Linguistic Term Set and Decision Methods
Probabilistic linguistic term sets were initially proposed to overcome an important limitation of hesitant fuzzy linguistic term sets, namely the implicit assumption that all linguistic terms are equally important. A probabilistic linguistic term set attaches probability values to linguistic terms, making it possible to express the distribution of opinions in a natural and flexible way. Since its introduction, many operational laws, aggregation operators, distance measures, similarity measures, and outranking methods have been developed. The probabilistic linguistic representation is especially suitable for summarizing large numbers of online comments, because the probability on each linguistic term can be directly interpreted as the proportion of consumers holding a certain opinion.
Several classical MADM methods have been extended to the probabilistic linguistic environment, including TOPSIS, TODIM, VIKOR, PROMETHEE, ELECTRE, and LINMAP. In addition, group consensus frameworks and regret-theory extensions have been proposed. Nonetheless, only a few attempts have integrated the MARCOS method with probabilistic linguistic evaluation information. MARCOS is attractive because it produces rankings that are robust to the alternative-adding phenomenon and can handle trade-offs among many criteria. Therefore, an extension to the probabilistic linguistic context is a potentially valuable contribution.
3. Preliminaries
3.1 Probabilistic Linguistic Term Sets
Let \(S=\{s_l\mid l=-\tau,\ldots,-1,0,1,\ldots,\tau\}\) be a discrete linguistic term set with odd granularity \(2\tau+1\). A probabilistic linguistic term set on \(S\) is defined as
$$
L(p)=\left\{L^{(k)}(p^{(k)})\mid L^{(k)}\in S,\; p^{(k)}\in[0,1],\; k=1,2,\ldots,\#L(p),\;\sum_{k=1}^{\#L(p)}p^{(k)}\le1\right\},
$$
where \(L^{(k)}\) is the \(k\)-th linguistic term, \(p^{(k)}\) is its corresponding probability, and \(\#L(p)\) is the total number of different linguistic terms contained in \(L(p)\). In most review-summarization tasks, the probabilities are obtained from the observed frequency of each opinion grade among all relevant comments.
The score function of a probabilistic linguistic term set is used for comparison and defuzzification. If \(r^{(k)}\) is the subscript of the linguistic term \(L^{(k)}\), then the score is expressed as
$$
E(L(p)) = s_{\alpha}, \qquad \alpha = \frac{\sum_{k=1}^{\#L(p)}r^{(k)}p^{(k)}}{\sum_{k=1}^{\#L(p)}p^{(k)}}.
$$
To describe the degree of disagreement inside a PLTS, the deviation degree is defined by
$$
\sigma(L(p)) = \frac{\sum_{k=1}^{\#L(p)}p^{(k)}\left(r^{(k)}-\alpha\right)^2}{\sum_{k=1}^{\#L(p)}p^{(k)}}.
$$
For two PLTSs with the same number of linguistic terms, the Euclidean distance between them is frequently used:
$$
d\left(L_1(p),L_2(p)\right)=\sqrt{\frac{1}{\#L(p)}\sum_{k=1}^{\#L(p)}\left(r^{(k)}_1p^{(k)}_1-r^{(k)}_2p^{(k)}_2\right)^2}.
$$
When the probabilities in a PLTS do not sum to one, a normalization operation can be applied by computing
$$
\hat{p}^{(k)} = \frac{p^{(k)}}{\sum_{k=1}^{\#L(p)}p^{(k)}}.
$$
To aggregate probabilistic linguistic information, the probabilistic linguistic weighted average (PLWA) operator is defined as
$$
\text{PLWA}_{\omega}\left(L_{1}(p),L_{2}(p),\ldots,L_{n}(p)\right)
=
\bigcup_{\substack{L_1^{(k)}\in L_1(p)\\ L_2^{(k)}\in L_2(p)\\ \vdots\\ L_n^{(k)}\in L_n(p)}}
\left\{
\sum_{i=1}^{n} \omega_i p_i^{(k)} L_i^{(k)}
\right\},
$$
where \(\omega=(\omega_1,\ldots,\omega_n)^\mathrm{T}\) is the weight vector satisfying \(\omega_i\in[0,1]\) and \(\sum_{i=1}^n\omega_i=1\).
3.2 ITARA Weighting Method
ITARA is a recently proposed objective weighting method based on indifference thresholds. The central idea is that if the performances of alternatives under one attribute are very similar, then this attribute has little discriminative power and should receive a lower weight. The method involves several steps: acquire the indifference threshold \(T_j\) for criterion \(C_j\); normalize the decision matrix; order the performance values; compute the ordered distances between adjacent alternatives; determine which ordered distances are significantly larger than the indifference threshold; and finally derive the attribute weight from the aggregated differences. One of the shortcomings of the original ITARA is that the indifference threshold is usually derived only from the positive ideal solution, which assumes that the decision-maker is perfectly rational. In our model, we improve this step by invoking the value function of prospect theory.
3.3 Prospect Theory
Prospect theory was introduced to capture bounded rationality in human judgments. Its value function is concave in the gain domain and convex in the loss domain, with a steeper slope in the loss domain. Mathematically, the value function can be written as
$$
v(x)=
\begin{cases}
x^{\alpha}, & x\ge 0,\\
-\theta (-x)^{\beta}, & x<0,
\end{cases}
$$
where \(\alpha\) and \(\beta\) are the attitude parameters toward gains and losses, respectively, and \(\theta\) is the loss-aversion coefficient. Empirically obtained values such as \(\alpha=\beta=0.88\) and \(\theta=2.25\) are commonly used. Prospect theory has been widely integrated into MADM because different consumers may evaluate the same information differently depending on their reference points.
3.4 MARCOS Method
MARCOS is a multi-criteria decision-making method built around the idea of compromise ranking. It first constructs ideal and anti-ideal solutions, then measures each alternative’s utility with respect to both references, and finally combines these utility values into an overall utility function. MARCOS is known to be robust in the presence of a large decision matrix and less sensitive to rank reversal compared to some other distance-based methods.
4. Electric Vehicle Purchase Decision Model Based on Improved ITARA
The framework proposed in this paper is depicted conceptually in four phases: data preparation, PLTS construction, weighting, and ranking. In the first phase, user reviews about several electric vehicles are collected from multiple websites. In the second phase, text mining and sentiment analysis produce individual PLTS decision matrices. In the third phase, webpage weights and attribute weights are computed. In the fourth phase, the probabilistic linguistic MARCOS method is applied to generate the final rank order.
4.1 Text Crawling and Preprocessing
Online comments for each candidate electric vehicle are collected from two representative automotive review websites using Python web crawlers. The target period is limited to the most recent year so that the information reflects the current product version as much as possible. Duplicate comments, advertisements, and meaningless strings are removed. The remaining texts are then subjected to Chinese word segmentation. Since consumer language is colloquial and noisy, we extend the segmentation dictionary with domain-specific terms. Stop words such as auxiliary particles and frequently occurring adverbs that do not contribute to product evaluation are added to a customized stop-word list. In addition, brand names and model names are excluded so that the remaining vocabulary mainly describes product attributes.
The texts are then processed by the content-mining software to generate word-frequency lists. After descending sorting, high-frequency attribute-related words are manually merged. For example, words such as “acceleration,” “steering,” and “braking” are grouped under the general attribute “handling.” Following this procedure, eight attributes are selected for electric vehicle evaluation: handling, range, appearance, interior, intelligent configuration, space, price, and comfort. The grouping dictionary in the case study is summarized in the following table.
| No. | General attribute | Representative merged terms |
|---|---|---|
| C1 | Handling | power, engine, driving, acceleration, throttle, brake, steering, control |
| C2 | Range | kilometers, battery, mileage, long distance, energy consumption |
| C3 | Appearance | body, shape, tail lamp, paint, door, exterior |
| C4 | Interior | seat, screen, central console, instrument panel, sunroof |
| C5 | Features | function, system, infotainment, mode, air conditioning, performance, tire, voice |
| C6 | Space | rear row, trunk, storage, front row, spacious, compact |
| C7 | Price | cost, fee, price level, value for money, discount, saving |
| C8 | Comfort | wind noise, tire noise, odor, insulation, experience, damping, noise |
4.2 Conversion of Text Reviews into Seven-Grade PLTS Decision Matrices
Sentiment analysis is implemented at the attribute-phrase level. For each attribute-word family, the sentiment-analysis engine returns a Likert score \(\xi_f\) between \(-3\) and \(+3\). We map these scores onto a seven-level linguistic scale \(l_{-3},l_{-2},l_{-1},l_0,l_1,l_2,l_3\). The exact mapping is summarized in Table below.
| Likert score | Semantic meaning | Linguistic term |
|---|---|---|
| -3 | extremely negative | \(l_{-3}\) |
| -2 | moderately negative | \(l_{-2}\) |
| -1 | slightly negative | \(l_{-1}\) |
| 0 | neutral | \(l_0\) |
| 1 | slightly positive | \(l_1\) |
| 2 | moderately positive | \(l_2\) |
| 3 | extremely positive | \(l_3\) |
After the sentiment scores for all relevant attribute expressions are obtained, the relative frequency of each linguistic grade is computed. These relative frequencies are treated as probabilities, which yields a normalized seven-grade PLTS for every combination of candidate vehicle and attribute. Thus, if agent \(e_h\) provides reviews about alternative \(P_i\) with respect to attribute \(C_j\), the aggregated evaluation is denoted by \(g^h_{ij}\), and the associated decision matrix is
$$
D^h = \left(g^h_{ij}\right)_{m\times n}.
$$
This conversion process preserves the distribution of consumer opinions rather than reducing them to an average sentiment score. For example, if 60% of comments are extremely positive and 20% are moderately positive while 20% are negative, this distribution is directly represented in the PLTS.
4.3 Determination of Webpage Weights
Because comments are collected from several websites, the reliability of each webpage should be considered. A simple arithmetic mean of various websites would overweight sparsely reviewed or extremely aggressive reviews. In our model, the webpage weight is composed of two components.
The first component is the quantity share of reviews. Suppose \(N_h\) denotes the number of valid comments collected from webpage \(e_h\), for \(h=1,\ldots,H\). The quantity share is
$$
U_h = \frac{N_h}{\sum_{h=1}^{H}N_h}.
$$
The second component is an evaluation-credibility measure. A webpage whose PLTS elements have large deviation degrees indicates severe disagreement among commenters, so it is considered less reliable. For every element \(g^h_{ij}\), the deviation degree \(\sigma(g^h_{ij})\) can be calculated from the definition in Section 3.1. Then the credibility component is defined by
$$
\widetilde{U}_h = 1 – \frac{\sum_{i=1}^{m}\sum_{j=1}^{n}\sigma\left(g^h_{ij}\right)}
{\sum_{h=1}^{H}\sum_{i=1}^{m}\sum_{j=1}^{n}\sigma\left(g^h_{ij}\right)}.
$$
By combining these two components with a balancing parameter \(\gamma\in[0,1]\), we define the importance of webpage \(e_h\) as
$$
\Theta(e_h) = \gamma U_h + (1-\gamma)\widetilde{U}_h,
$$
and the normalized webpage weight is
$$
w_h = \frac{\Theta(e_h)}{\sum_{h=1}^{H}\Theta(e_h)}.
$$
In the remainder of this paper, \(\gamma=0.5\) is adopted, meaning that review quantity and credibility are considered equally important. The individualized web-based decision matrices can then be aggregated into a group decision matrix by the PLWA operator:
$$
g_{ij} = \operatorname{PLWA}_{w}\left(g^1_{ij},g^2_{ij},\ldots,g^H_{ij}\right).
$$
4.4 Attribute Weights Based on Improved ITARA with Prospect Theory
In traditional ITARA, the indifference threshold is computed mainly according to the positive ideal solution. This is inconsistent with the fact that consumers do not always evaluate information in a neutral way. To introduce psychological realism into electric vehicle selection, we next improve the threshold-construction stage.
For each attribute \(C_j\), the positive ideal solution \(g_j^+\) and negative ideal solution \(g_j^-\) are defined according to the score function of PLTS. Specifically,
$$
g_j^+ =
\begin{cases}
\arg\max_{1\le i\le m} E(g_{ij}), & \text{for benefit criterion},\\
\arg\min_{1\le i\le m} E(g_{ij}), & \text{for cost criterion},
\end{cases}
$$
$$
g_j^- =
\begin{cases}
\arg\min_{1\le i\le m} E(g_{ij}), & \text{for benefit criterion},\\
\arg\max_{1\le i\le m} E(g_{ij}), & \text{for cost criterion}.
\end{cases}
$$
Then the average distances from all alternatives to the positive and negative ideal points are computed. These two distances are interpreted differently in prospect theory. The distance to the positive ideal solution represents a possible loss because an alternative fails to match the best performance, while the distance to the negative ideal solution represents potential gain because the alternative can deviate away from the worst performance. Thus,
$$
\varsigma_j^+ = \frac{1}{m}\sum_{i=1}^{m}d\left(g_{ij},g_j^+\right),
\qquad
\varsigma_j^- = \frac{1}{m}\sum_{i=1}^{m}d\left(g_{ij},g_j^-\right).
$$
Applying the prospect value function yields
$$
v_j^+ = \left(\varsigma_j^-\right)^{\alpha},
\qquad
v_j^- = -\theta\left(\varsigma_j^+\right)^{\beta}.
$$
The final indifference threshold for attribute \(C_j\) is then computed as the average of these two prospect values:
$$
T_j = \frac{1}{2}\left(v_j^+ + v_j^-\right).
$$
When \(\alpha\), \(\beta\), and \(\theta\) change, the threshold \(T_j\) reflects different attitudes toward gains and losses. This is the key difference between the improved ITARA and the standard ITARA method.
After the thresholds are obtained, the alternatives are sorted in ascending order under each attribute based on their PLTS scores. For brevity, let \(\beta_{ij}\) and \(\beta_{i+1,j}\) be two adjacent alternatives in the sorted order for attribute \(C_j\). The ordered distance between them is
$$
\gamma_{ij}=d\left(\beta_{ij},\beta_{i+1,j}\right),
\quad i=1,2,\ldots,m-1.
$$
Only the ordered distances that exceed the indifference threshold contribute to the attribute weight. The excess difference is defined as
$$
\phi_{ij}=
\begin{cases}
\gamma_{ij}-T_j, & \gamma_{ij}>T_j,\\
0, & \gamma_{ij}\le T_j.
\end{cases}
$$
Using the \(l_2\)-metric aggregation, we obtain
$$
\xi_j = \left(\sum_{i=1}^{m-1}\phi_{ij}^2\right)^{1/2},
$$
and the final attribute weights are normalized by
$$
\omega_j = \frac{\xi_j}{\sum_{j=1}^{n}\xi_j}.
$$
Attributes with very small differences among the electric vehicle models receive smaller weights and sometimes even zero weight if they cannot help consumers distinguish among alternatives. In contrast, attributes that generate substantial differences receive large weights.
4.5 Ranking via Probabilistic Linguistic MARCOS
The MARCOS method is extended here to PLTS in order to obtain the final ranking. First, the group PLTS matrix \(G=(g_{ij})_{m\times n}\) is augmented by inserting the positive ideal vector and the negative ideal vector as two extra rows. Using the attribute weights, each row is aggregated into a PLTS via the PLWA operator. Then the score function is used to obtain a crisp value for each candidate electric vehicle and for the ideal and anti-ideal rows.
Let \(S_i\) denote the score of the aggregated evaluation for electric vehicle alternative \(i\). Further let \(S_{\text{AI}}\) and \(S_{\text{AAI}}\) denote the scores of the positive ideal and negative ideal solutions, respectively. The utility degrees are
$$
K_i^+ = \frac{S_i}{S_{\text{AI}}},
\qquad
K_i^- = \frac{S_i}{S_{\text{AAI}}}.
$$
Because \(S_{\text{AAI}}\) is derived from negative ideal ratings, its value is usually the lowest among all alternatives. A high \(K_i^+\) means that the alternative is close to the ideal solution, while a high \(K_i^-\) means that the alternative is far away from the anti-ideal solution. These two utility degrees are then integrated using the utility functions
$$
f\left(K_i^+\right) = \frac{K_i^-}{K_i^+ + K_i^-},
\qquad
f\left(K_i^-\right) = \frac{K_i^+}{K_i^+ + K_i^-}.
$$
Eventually, the overall utility function for each electric vehicle alternative is calculated by
$$
f(K_i)=\frac{K_i^+ + K_i^-}{1+\frac{1-f(K_i^+)}{f(K_i^+)}+\frac{1-f(K_i^-)}{f(K_i^-)}}.
$$
Larger \(f(K_i)\) values imply better electric vehicle alternatives. The complete model therefore combines textual review semantics, multi-source reliability, psychological risk attitudes, and stable utility-based ranking.
5. Case Study
5.1 Data Description
To demonstrate the practical application of the proposed framework, we simulated a consumer who intends to purchase an electric vehicle for daily commuting. After preliminary screening, five battery electric vehicles are selected as candidate alternatives; they are denoted by \(P_1,P_2,P_3,P_4,P_5\) in order to keep the analysis neutral. Online reviews published in the most recent year were collected from two large automotive consumer-review websites. After removing duplicates and invalid entries, a total of \(3{,}658\) valid reviews were retained. The review counts for each website and model are shown in the next table, where Website \(E_1\) and Website \(E_2\) represent the two sources.
| Website | Model \(P_1\) | Model \(P_2\) | Model \(P_3\) | Model \(P_4\) | Model \(P_5\) | Total |
|---|---|---|---|---|---|---|
| \(E_1\) | 65 | 290 | 511 | 1,396 | 839 | 3,101 |
| \(E_2\) | 87 | 105 | 192 | 69 | 104 | 557 |
| Total | 152 | 395 | 703 | 1,465 | 943 | 3,658 |
5.2 Data Preprocessing and PLTS Construction
The raw review text was processed using the workflow described in Section 4.1. Segmentation and word-frequency analysis generated the eight attributes shown earlier. For instance, almost all high-frequency words related to battery, battery life, mileage, and charging efficiency were merged under the attribute “range”; words such as air-conditioning, voice control, assisted driving, and intelligent system were merged under the attribute “features”; and words related to tire noise, seat comfort, and suspension were merged under the attribute “comfort.” After obtaining the multi-dimensional sentiment scores, we constructed the probabilistic linguistic decision matrices. Because the original PLTS arrays are lengthy, the following small excerpt illustrates the evaluation information for models \(P_1\) and \(P_2\) under attribute \(C_1\).
| Attribute | Alternative | PLTS evaluation |
|---|---|---|
| \(C_1\): Handling | \(P_1\) | \(\{l_{-3}(0.118), l_{-2}(0.053), l_{-1}(0.026), l_0(0.010), l_1(0.092), l_2(0.036), l_3(0.665)\}\) |
| \(P_2\) | \(\{l_{-3}(0.016), l_{-2}(0.007), l_{-1}(0.019), l_0(0.024), l_1(0.047), l_2(0.042), l_3(0.845)\}\) |
These elements are produced after aggregating the two website-specific matrices with the webpage weights computed in the next subsection. The full decision matrix has a dimension \(5\times 8\), and every cell has the same PLTS structure.
5.3 Webpage Weights and Aggregated Decision Matrix
Using the procedure in Section 4.3, the quantity shares and credibility shares of the two websites are calculated. The results are shown below.
| Website | Quantity share \(U_h\) | Credibility share \(\widetilde{U}_h\) | Final weight \(w_h\) |
|---|---|---|---|
| \(E_1\) | 0.848 | 0.592 | 0.72 |
| \(E_2\) | 0.152 | 0.408 | 0.28 |
The website with substantially more electric vehicle reviews therefore receives a larger weight. The aggregated group decision matrix is then obtained as a weighted PLTS sum of the two website matrices.
5.4 Attribute Weights
For each attribute, the positive and negative ideal PLTSs were identified based on their score values. The average distances to these ideal points were computed and transformed into indifference thresholds by the prospect-theory function. The final attribute weights computed by the improved ITARA are shown in the table.
| Attribute | \(C_1\) Handling | \(C_2\) Range | \(C_3\) Appearance | \(C_4\) Interior | \(C_5\) Features | \(C_6\) Space | \(C_7\) Price | \(C_8\) Comfort |
|---|---|---|---|---|---|---|---|---|
| Weight \(\omega_j\) | 0.2087 | 0.0666 | 0.0402 | 0.0000 | 0.3561 | 0.0429 | 0.0029 | 0.2826 |
The largest weights are assigned to \(C_5\) (features) and \(C_8\) (comfort), followed by \(C_1\) (handling). This result indicates that consumers in the dataset pay more attention to intelligent configuration, comfort, and driving experience in electric vehicles. The zero weight of attribute \(C_4\) (interior) suggests that the five candidate electric vehicles do not show sufficiently discriminative differences in this dimension across the evaluated online comments.
5.5 Ranking with Probabilistic Linguistic MARCOS
By applying the probabilistic linguistic weighted average operator to the group decision matrix, each alternative was aggregated into a single PLTS. We then calculated the score values of all alternatives together with the ideal and anti-ideal scores. The table below gives the resulting scores and utility values.
| Alternative | Score \(S_i\) | \(K_i^+\) | \(K_i^-\) | \(f(K_i)\) | Rank |
|---|---|---|---|---|---|
| \(P_1\) | 1.483 | 0.612 | 1.005 | 0.723 | 5 |
| \(P_2\) | 2.375 | 0.979 | 1.610 | 1.461 | 1 |
| \(P_3\) | 2.139 | 0.882 | 1.450 | 1.255 | 4 |
| \(P_4\) | 2.325 | 0.959 | 1.576 | 1.416 | 2 |
| \(P_5\) | 2.283 | 0.941 | 1.547 | 1.380 | 3 |
| AI | 2.426 | — | — | — | — |
| AAI | 1.476 | — | — | — | — |
The final ranking is \(P_2 \succ P_4 \succ P_5 \succ P_3 \succ P_1\). According to the online-review-based model, alternative \(P_2\) is the most suitable electric vehicle choice for the simulated consumer. This result is intuitively reasonable because \(P_2\) has the highest proportion of positive comments and a very good balance among the highly weighted attributes.
6. Sensitivity and Comparative Analysis
6.1 Sensitivity of Prospect-Theory Coefficients
Because the improved ITARA method depends on the prospect-theory parameters \(\alpha\), \(\beta\), and \(\theta\), it is important to observe how these parameters influence the final attribute weights and ranks. We performed three separate sensitivity streams.
In the first stream, the loss-aversion coefficient \(\theta\) was changed from 1 to 5 while keeping the other parameters fixed. It was observed that \(C_5\) gained increasing weight when \(\theta\) increased, and eventually dominated all other criteria as \(\theta\) approached 5. By contrast, the weights of \(C_2\), \(C_3\), and \(C_6\) decreased gradually to zero, while \(C_1\) and \(C_8\) first increased and then decreased. This behaviour confirms that a consumer who is extremely sensitive to potential loss will tend to rely heavily on the “features” attribute when choosing an electric vehicle.
In the second stream, the gain-attitude coefficient \(\alpha\) was increased from 0 to 1. In this scenario, \(C_8\) initially dominated at very small \(\alpha\), whereas higher \(\alpha\) increased the weights of \(C_1\) and \(C_5\). Thus, when the consumer pays more attention to gains, both handling and configuration become more influential.
In the third stream, the loss-attitude coefficient \(\beta\) was adjusted from 0.5 to 1. The attribute \(C_5\) dominated when \(\beta\) was small, while \(C_8\) and \(C_1\) increased gradually. Overall, the weight vector changes noticeably with each risk parameter, demonstrating that the proposed attribute-weighting model is sufficiently sensitive to reflect different consumer risk preferences.
6.2 Comparison with Probabilistic Linguistic TOPSIS
To assess the reliability of the extended MARCOS ranking, we compared it with the well-known PL-TOPSIS method using the same group decision matrix and the same attribute weights. In PL-TOPSIS, the distance from each alternative to the positive ideal solution is minimized and the distance to the negative ideal solution is maximized. The resulting ranking derived from the closeness coefficient in PL-TOPSIS is \(P_2 \succ P_4 \succ P_3 \succ P_5 \succ P_1\).
The ranking obtained by PL-MARCOS is \(P_2 \succ P_4 \succ P_5 \succ P_3 \succ P_1\). The first two alternatives and the last alternative of the two rankings are identical, and only the middle positions of \(P_3\) and \(P_5\) are slightly swapped. The close agreement verifies that the probabilistic linguistic MARCOS method is a valid ranking tool. Moreover, MARCOS has several advantages: it does not require an extra normalization step in a simple probabilistic framework, it explicitly builds utility functions for both ideal and anti-ideal directions, and it tends to be more robust against rank-reversal problems when new alternatives are introduced.
6.3 Comparison with Probabilistic Linguistic TODIM
We also compared the proposed model with PL-TODIM, a widely cited method for capturing the loss-avoidance behavior of decision-makers. PL-TODIM computes the dominance degree of each pair of alternatives under each attribute. In the present case, the PL-TODIM ranking is exactly \(P_2 \succ P_4 \succ P_5 \succ P_3 \succ P_1\), which is identical to the ranking obtained by the proposed PL-MARCOS method. This consistency is an encouraging signal that the proposed framework captures the same fundamental preference structure as TODIM but with much lower computational complexity. Because PL-MARCOS does not require pairwise comparisons for every pair of alternatives, the computation burden is substantially reduced in large-scale electric vehicle selection problems.
| Method | Ranking |
|---|---|
| Proposed PL-MARCOS | \(P_2\succ P_4\succ P_5\succ P_3\succ P_1\) |
| PL-TOPSIS | \(P_2\succ P_4\succ P_3\succ P_5\succ P_1\) |
| PL-TODIM | \(P_2\succ P_4\succ P_5\succ P_3\succ P_1\) |
7. Conclusion and Outlook
In this paper, we established an online-review-based multi-attribute decision model to support the purchase decision of electric vehicles. The main contributions are fourfold. First, the proposed method converts textual online comments into a seven-grade probabilistic linguistic representation, thereby preserving the distribution of consumer opinions rather than compressing them into an average sentiment score. This approach captures richer attitudinal differences among consumers. Second, we considered webpage weights by jointly examining the number and the consistency of comments across different online car review platforms. This treatment effectively reduces the bias caused by websites with a small number of comments or high disagreement. Third, the ITARA weighting method is improved by applying the prospect-theory value function to the construction of indifference thresholds. As a result, the attribute weights reflect not only the dispersion of performance values but also the psychological risk attitudes of electric vehicle buyers. Fourth, the MARCOS method was extended to the probabilistic linguistic environment, which provides a robust final ranking that considers both proximity to the positive ideal solution and distance from the negative ideal solution. The application to real review data demonstrates the feasibility of the entire decision pipeline.
Despite its strengths, the proposed model has several limitations that provide directions for future research. First, for newly launched electric vehicle models, the number of online reviews may be too small to yield a reliable PLTS. Cold-start situations will alter the webpage balance and increase uncertainty. Future studies could integrate expert judgments or market prices to mitigate this issue. Second, our model captures loss aversion and risk attitude through prospect theory, but consumer psychology is multi-dimensional. Regret aversion, social conformity, and fairness concerns may also influence the purchase of electric vehicles. Third, we focus on textual comments; however, numerical star ratings often accompany text. A richer decision model that simultaneously uses both structured ratings and unstructured text can further improve the robustness of the electric vehicle ranking. In addition, future applications may employ dynamic sentiment analysis to track opinion shifts over time and help consumers observe how a vehicle model evolves through product update cycles.
In conclusion, the growing complexity of the electric vehicle market demands systematic decision support. The probabilistic linguistic model proposed here is a reliable and practical step toward helping consumers mine large-scale online comments, understand product differences, and ultimately select an electric vehicle that best fits their preferences.