1. Introduction
In recent years, the automotive industry has witnessed a remarkable transition towards electrification. Governments around the world have introduced supportive policies, while battery technology, electric drive systems, and smart cockpit technologies have continuously advanced. As a result, the number of electric cars available to consumers has grown rapidly, and competition among automotive manufacturers has become increasingly intense. Although electric cars offer notable benefits in energy saving and emissions reduction, consumers are often confronted with considerable difficulty when selecting a suitable electric car. The purchasing process involves many technical specifications and subjective attributes, such as range, driving comfort, appearance, interior quality, intelligent configuration, space, price, and handling performance. A consumer must process a large amount of information before making a confident choice.
Online reviews have become one of the most important information sources in this decision process. Potential electric car buyers frequently consult user-generated reviews on automotive websites before visiting dealerships or making a purchase. Unlike simple star ratings, text reviews provide detailed descriptions of real driving experience. For example, a review may describe how the car accelerates on highways, how the suspension absorbs rough roads, how quiet the cabin is at high speed, or how spacious the rear seats are. These qualitative clues cannot be fully captured by a single rating score. However, online reviews also present a challenge: the amount of text can be enormous, and individual comments may contain ambiguous or even contradictory opinions. Therefore, modeling consumer preferences from online reviews and transforming them into a reliable ranking of electric car alternatives is both practically relevant and theoretically important.
This study considers the purchase decision of an electric car as a multi-attribute decision-making problem. I propose a framework that integrates text mining, emotion analysis, probabilistic linguistic modeling, improved objective weighting, and multi-criteria ranking. My work starts with collecting electric car reviews from multiple automobile websites. The text is segmented and the key product attributes are extracted by frequency analysis. Sentiment analysis then converts the text comments into seven-grade probabilistic linguistic evaluations. After that, I determine the weight of each website by considering both the number of comments and review consistency. For the attribute weights, I improve the ITARA method with a prospect-theory-based threshold setting. Finally, I extend the MARCOS method to the probabilistic linguistic environment to rank all candidate electric cars.
The main contributions of this paper are as follows:
- I propose a fine-grained conversion method from online review text to seven-grade probabilistic linguistic term sets. The method preserves richer emotional information than coarse three-grade sentiment classification.
- I construct a website-weight calculation model that simultaneously accounts for review quantity and review reliability. The model reduces bias caused by sparse comments on newly launched electric cars.
- I improve the ITARA method by using prospect theory to derive indifference thresholds. The improved method can reflect consumers’ risk attitudes in the weight determination process.
- I extend the MARCOS method to the probabilistic linguistic framework. The resulting method provides a robust ranking of electric car alternatives based on their utility relative to both positive and negative ideal solutions.
2. Preliminaries
This section introduces the basic concepts used throughout the study. The probabilistic linguistic term set is the core tool for representing uncertain evaluation information extracted from online reviews. I also present the score function, deviation degree, and normalized Euclidean distance used in later computations.
2.1 Probabilistic linguistic term sets
The qualitative evaluation of an electric car is usually expressed with linguistic terms. Let \(S=\{s_t \mid t=-3,-2,-1,0,1,2,3\}\) be a linguistic term set with seven grades. A probabilistic linguistic term set is defined as:
$$
L(p)=\left\{s_{(k)}\left(p^{(k)}\right)\; \middle| \; s_{(k)}\in S,\ p^{(k)}\ge 0,\sum_{k=1}^{\#L(p)}p^{(k)}\le 1\right\}.
\tag{1}
$$
Here \(s_{(k)}\) is the \(k\)th linguistic term, \(p^{(k)}\) is its corresponding probability, and \(\#L(p)\) is the total number of linguistic terms that have nonzero probability. In the normalized form, the probabilities are adjusted so that their sum equals one:
$$
\tilde{p}^{(k)}=\frac{p^{(k)}}{\sum_{t=1}^{\#L(p)}p^{(t)}}.
\tag{2}
$$
Let \(r^{(k)}\) be the subscript of \(s_{(k)}\), so \(s_{(k)}=s_{r^{(k)}}\). The score function is:
$$
E\left(L(p)\right)=s_{\bar{\alpha}},\quad
\bar{\alpha}=\frac{\sum_{k=1}^{\#L(p)}r^{(k)}p^{(k)}}{\sum_{k=1}^{\#L(p)}p^{(k)}}.
\tag{3}
$$
A larger score value indicates better overall evaluation. To measure disagreement among the linguistic terms, the deviation degree is used:
$$
\sigma\left(L(p)\right)=\frac{\sum_{k=1}^{\#L(p)}p^{(k)}\left(r^{(k)}-\bar{\alpha}\right)^2}{\sum_{k=1}^{\#L(p)}p^{(k)}}.
\tag{4}
$$
A smaller deviation implies that reviewers have more consistent opinions.
2.2 Distance between probabilistic linguistic term sets
The distance between two probabilistic linguistic term sets is an important tool for measuring differences between product alternatives under a given attribute. In this study, all probabilistic linguistic evaluations are built on the same seven-grade language scale, so the distance formula can be written as:
$$
d\left(L_1(p),L_2(p)\right)=\sqrt{\frac{1}{\#L}\sum_{k=1}^{\#L}\left(r_1^{(k)}p_1^{(k)}-r_2^{(k)}p_2^{(k)}\right)^2},
\tag{5}
$$
where \(r_i^{(k)}\) and \(p_i^{(k)}\) are the subscript and probability of the \(k\)th term in \(L_i(p)\), respectively. This distance function satisfies the basic conditions of nonnegativity, symmetry, and identity of indiscernibles.
3. Proposed Purchase Decision Model
The proposed electric car purchase decision model consists of four stages: online review conversion, website weight calculation, attribute weight determination, and final ranking. I describe these stages in detail below.
3.1 Online review conversion into probabilistic linguistic evaluation
The first step is to collect text reviews of the candidate electric cars from automobile websites. A Python-based crawler is used to retrieve review pages, cleaning the text by removing duplicate records, emoticons, and meaningless characters. Then, text segmentation is performed with the ROST Content Mining software. Domain-specific words are added to a custom dictionary to improve segmentation accuracy, while stop words such as particles and high-frequency adverbs are removed. Standard electric car model names and brand names are also excluded from the vocabulary to avoid unnecessary interference.
Word-frequency statistics are then used to identify the attributes that are most frequently discussed. Related words are clustered into broader attribute categories. In this study, eight attributes are identified: driving control, battery range, exterior appearance, interior decoration, intelligent configuration, cabin space, price, and driving comfort. These attributes form the evaluation criteria set \(C=\{C_1,C_2,\ldots,C_8\}\).
To quantify the sentiment orientation of each review attribute, the Emotion Calculator software calculates a Likert score between \(-3\) and \(3\) for each attribute word in each review. The score is then converted into one of the seven linguistic grades. For example, a score of \(3\) corresponds to \(s_3\), a score of \(0\) corresponds to \(s_0\), and a score of \(-3\) corresponds to \(s_{-3}\). For a given electric car \(P_i\), attribute \(C_j\), and website \(e_h\), the probability of grade \(s_t\) is calculated by counting how many reviews express that grade and dividing the result by the total number of valid reviews in the corresponding evaluation cell. Thus, the website-level evaluation is:
$$
g_{ij}^{(h)}=\left\{s_t\left(p_{ij,t}^{(h)}\right)\mid t=-3,-2,\ldots,3\right\}.
\tag{6}
$$
3.2 Website weights based on review quantity and reliability
Not all websites have equal importance. A website with many detailed comments should generally be given greater weight, because its evaluation has a broader sample base. At the same time, a website whose comments deviate sharply from the average may be less reliable. I combine both aspects into a website-weight formula.
Let \(N_h\) be the number of valid reviews collected from website \(e_h\). The quantity factor is:
$$
u_h=\frac{N_h}{\sum_{t=1}^{H}N_t}.
\tag{7}
$$
The reliability factor is based on the average deviation of all probabilistic linguistic evaluations on that website:
$$
u_h’=1-\frac{\sum_{i=1}^{m}\sum_{j=1}^{n}\sigma\left(g_{ij}^{(h)}\right)}
{\sum_{h=1}^{H}\sum_{i=1}^{m}\sum_{j=1}^{n}\sigma\left(g_{ij}^{(h)}\right)}.
\tag{8}
$$
Here \(\sigma(g_{ij}^{(h)})\) is the deviation degree defined in Equation (4). The overall importance of website \(e_h\) is calculated as:
$$
\Theta(e_h)=\gamma u_h+(1-\gamma)u_h’,
\tag{9}
$$
where \(\gamma\in [0,1]\) balances review quantity and reliability. In this paper I set \(\gamma=0.5\), so both factors are treated as equally important. The normalized website weight is:
$$
w_h=\frac{\Theta(e_h)}{\sum_{t=1}^{H}\Theta(e_t)}.
\tag{10}
$$
After obtaining \(w_h\), the individual matrices are aggregated into a group decision matrix through the probabilistic linguistic weighted-average approach. Since all seven grades are present in the framework, the aggregated probability of grade \(s_t\) for alternative \(i\) and attribute \(j\) is:
$$
p_{ij,t}=\sum_{h=1}^{H}w_h p_{ij,t}^{(h)}.
\tag{11}
$$
3.3 Attribute weights based on an improved ITARA method
The ITARA method is a well-known objective weighting technique in multi-attribute decision-making. It measures how strongly alternatives differ under each attribute. If all alternatives have very similar values on an attribute, that attribute receives a low weight. The key parameter in ITARA is the indifference threshold, which controls whether a small difference between alternatives is meaningful. I improve the conventional threshold-setting procedure from the perspective of prospect theory.
First, the positive ideal evaluation and negative ideal evaluation are defined for each attribute \(C_j\). For a benefit-type attribute such as battery range, the positive ideal solution is the one with the maximum score, while the negative ideal solution has the minimum score. Because all attributes considered in this case are benefit criteria, the definitions are straightforward. Let \(g_j^{+}\) and \(g_j^{-}\) be the positive and negative ideal probabilistic linguistic evaluations for attribute \(C_j\). The average distance to the positive ideal is called the extremely good indifference threshold:
$$
T_j^{+}=\frac{1}{m}\sum_{i=1}^{m}d\left(g_{ij},g_j^{+}\right),
\tag{12}
$$
and the average distance to the negative ideal is called the extremely poor indifference threshold:
$$
T_j^{-}=\frac{1}{m}\sum_{i=1}^{m}d\left(g_{ij},g_j^{-}\right).
\tag{13}
$$
Conventional ITARA methods often derive thresholds only from the positive ideal solution. However, consumers perceive losses and gains differently. Prospect theory states that losses have a larger psychological impact than gains of the same magnitude. Therefore, I treat \(T_j^{+}\) as a loss-like quantity and \(T_j^{-}\) as a gain-like quantity. Using the value function of prospect theory:
$$
v^{+}=\left(T_j^{-}\right)^{\alpha},\quad
v^{-}=-\theta\left(T_j^{+}\right)^{\beta}.
\tag{14}
$$
Based on empirical evidence, I choose \(\alpha=\beta=0.88\) and \(\theta=2.25\). The final indifference threshold is obtained by the additive weighted aggregation:
$$
T_j=\frac{1}{2}\left(v^{+}+v^{-}\right).
\tag{15}
$$
Then, all evaluation values under attribute \(C_j\) are sorted in ascending order. Let \(g_{ij}\) and \(g_{i+1,j}\) be two consecutive evaluations after sorting. Their ordered distance is:
$$
o_{ij}=d\left(g_{ij},g_{i+1,j}\right),\quad i=1,2,\ldots,m-1.
\tag{16}
$$
If this ordered distance exceeds the final indifference threshold, the difference is considered meaningful and contributes to the attribute weight. Otherwise, the difference is excluded. The significant gap is:
$$
\phi_{ij}=\begin{cases}
o_{ij}-T_j, & \text{if } o_{ij}>T_j,\\
0, & \text{if } o_{ij}\le T_j.
\end{cases}
\tag{17}
$$
Finally, the weight of attribute \(C_j\) is obtained by the \(l_2\)-metric aggregation:
$$
\zeta_j=\left(\sum_{i=1}^{m-1}\phi_{ij}^{2}\right)^{1/2},\quad
\omega_j=\frac{\zeta_j}{\sum_{t=1}^{n}\zeta_t}.
\tag{18}
$$
3.4 Probabilistic linguistic MARCOS ranking method
I extend the MARCOS method to rank the alternatives after the attribute weights are obtained. The MARCOS method is built on the idea that the best alternative should have maximum utility relative to the positive ideal solution and minimum utility relative to the negative ideal solution.
Let \(S_i\) be the weighted score value of alternative \(P_i\). The weighted score of the positive ideal solution is denoted by \(S_{AI}\), and the weighted score of the negative ideal solution is denoted by \(S_{AAI}\). These scores are computed by the probabilistic linguistic score function after the weighted aggregation of each alternative under all attributes.
The utility degree relative to the positive ideal solution is:
$$
K_i^{+}=\frac{S_i}{S_{AI}},
\tag{19}
$$
and the utility degree relative to the negative ideal solution is:
$$
K_i^{-}=\frac{S_i}{S_{AAI}}.
\tag{20}
$$
The final utility function is:
$$
f(K_i)=\frac{K_i^{+}+K_i^{-}}{1+\frac{1-f(K_i^{+})}{f(K_i^{+})}+\frac{1-f(K_i^{-})}{f(K_i^{-})}},
\tag{21}
$$
where the auxiliary utility functions are:
$$
f(K_i^{+})=\frac{K_i^{-}}{K_i^{+}+K_i^{-}},\quad
f(K_i^{-})=\frac{K_i^{+}}{K_i^{+}+K_i^{-}}.
\tag{22}
$$
A larger value of \(f(K_i)\) indicates a better electric car alternative. The alternatives are therefore ranked in descending order of \(f(K_i)\).
4. Case Study of Electric Car Selection
To demonstrate the feasibility of the proposed model, I apply it to a real electric car selection problem. The consumer has owned a fuel vehicle and wants to buy an electric car for daily commuting. After consulting friends and carrying out test drives, five candidate electric cars are selected: the AION S, AION Y, Tesla Model Y, BYD Dolphin, and BYD Yuan PLUS. I denote them by \(P_1,\ldots,P_5\). The review information is collected from two Chinese automobile websites: the first website is a large traditional vertical automobile portal, and the second is an emerging automotive information platform. These two information sources are denoted by \(E_1\) and \(E_2\), respectively.
4.1 Data collection and preprocessing
A Python crawler is used to collect online text reviews published from December 2023 to December 2024. A total of 3672 comments are obtained. After deleting duplicate comments and invalid records, 3658 valid comments remain. Table 1 reports the number of comments for each electric car on each website.
| Website | AION S | AION Y | Model Y | BYD Dolphin | Yuan PLUS |
|---|---|---|---|---|---|
| \(E_1\) | 65 | 290 | 511 | 1396 | 839 |
| \(E_2\) | 87 | 105 | 192 | 69 | 104 |
In the preprocessing stage, the review texts are segmented with custom dictionaries. The stop-word list removes pronouns, auxiliary words, modal particles, and pure numbers. After segmentation, word frequencies are calculated. Related words are grouped into eight attribute categories. Table 2 presents the attribute structure and representative words.
| Attribute | Representative words |
|---|---|
| \(C_1\): Driving control | power, acceleration, throttle, brake, steering, handling |
| \(C_2\): Battery range | battery, mileage, energy consumption, long-distance travel |
| \(C_3\): Exterior appearance | body, styling, paint, lights, door lines |
| \(C_4\): Interior decoration | seat, central screen, instrument panel, roof |
| \(C_5\): Intelligent configuration | system, infotainment, voice control, air conditioning, tires |
| \(C_6\): Cabin space | rear seats, front seats, trunk, storage space |
| \(C_7\): Price | cost, cost-effectiveness, discount, car maintenance cost |
| \(C_8\): Driving comfort | wind noise, tire noise, odor, vibration, damping, ride quality |
4.2 Formation of the probabilistic linguistic decision matrix
After the sentiment analysis of each attribute-related word, the Likert scores are converted into the seven-grade probabilistic linguistic evaluations. I count the frequency of each language grade in each evaluation cell and normalize the probabilities. Because two websites are used, two \(5\times 8\) decision matrices are constructed. Each matrix contains probabilistic linguistic term sets. I then aggregate these two matrices into a combined matrix using the website weights described in the next section.
4.3 Calculation of website weights
Using Equations (7)–(10), the website weights are obtained. Table 3 lists the quantity factor, reliability factor, and final website weight.
| Website | Quantity factor \(u_h\) | Reliability factor \(u_h’\) | Final weight \(w_h\) |
|---|---|---|---|
| \(E_1\) | 0.8477 | 0.5922 | 0.72 |
| \(E_2\) | 0.1523 | 0.4078 | 0.28 |
Table 3 shows that website \(E_1\) has a substantially larger number of comments than website \(E_2\). However, \(E_2\) still contributes useful information because its reviews are relatively consistent and provide a complementary perspective. The aggregated decision matrix is summarized by the combined probabilistic linguistic evaluations. Let \(g_{ij}\) denote the final evaluation of alternative \(P_i\) on attribute \(C_j\). Each \(g_{ij}\) is a probabilistic linguistic term set with seven language grades.
4.4 Calculation of attribute weights
Using the combined decision matrix, I compute the positive ideal evaluation and the negative ideal evaluation for every attribute. Equations (12)–(15) are used to obtain the extremely good indifferent threshold \(T_j^{+}\), the extremely poor indifferent threshold \(T_j^{-}\), and the final prospect-theory-based indifferent threshold \(T_j\). The results are shown in Table 4.
| Attribute | \(T_j^{+}\) | \(T_j^{-}\) | \(T_j\) |
|---|---|---|---|
| \(C_1\) | 0.0796 | 0.1585 | 0.0225 |
| \(C_2\) | 0.0543 | 0.0724 | 0.0371 |
| \(C_3\) | 0.0569 | 0.0776 | 0.0376 |
| \(C_4\) | 0.0657 | 0.0571 | 0.0622 |
| \(C_5\) | 0.0824 | 0.1860 | 0.0113 |
| \(C_6\) | 0.0377 | 0.0380 | 0.0348 |
| \(C_7\) | 0.0587 | 0.0645 | 0.0479 |
| \(C_8\) | 0.1320 | 0.2300 | 0.0522 |
After sorting the combined evaluations in ascending order for each attribute, I calculate the ordered distances between adjacent alternatives. The significant differences are then derived by comparing those ordered distances with the final indifference thresholds. Table 5 shows the ordered-distance matrix.
| Ordered distance | \(C_1\) | \(C_2\) | \(C_3\) | \(C_4\) | \(C_5\) | \(C_6\) | \(C_7\) | \(C_8\) |
|---|---|---|---|---|---|---|---|---|
| \(o_{1j}\) | 0.1459 | 0.0781 | 0.0623 | 0.0285 | 0.2271 | 0.0193 | 0.0434 | 0.2229 |
| \(o_{2j}\) | 0.0577 | 0.0195 | 0.0335 | 0.0556 | 0.0265 | 0.0420 | 0.0478 | 0.0695 |
| \(o_{3j}\) | 0.0158 | 0.0309 | 0.0293 | 0.0288 | 0.0167 | 0.0420 | 0.0497 | 0.0192 |
| \(o_{4j}\) | 0.0305 | 0.0361 | 0.0299 | 0.0537 | 0.0463 | 0.0591 | 0.0213 | 0.0816 |
The meaningful gaps \(\phi_{ij}\) are shown in Table 6. If an ordered distance is not larger than the corresponding indifference threshold, its gap is set to zero.
| Gap | \(C_1\) | \(C_2\) | \(C_3\) | \(C_4\) | \(C_5\) | \(C_6\) | \(C_7\) | \(C_8\) |
|---|---|---|---|---|---|---|---|---|
| \(\phi_{1j}\) | 0.1234 | 0.0410 | 0.0247 | 0 | 0.2159 | 0 | 0 | 0.1707 |
| \(\phi_{2j}\) | 0.0352 | 0 | 0 | 0 | 0.0152 | 0.0073 | 0 | 0.0173 |
| \(\phi_{3j}\) | 0 | 0 | 0 | 0 | 0.0055 | 0.0073 | 0.0018 | 0 |
| \(\phi_{4j}\) | 0.0081 | 0 | 0 | 0 | 0.0350 | 0.0243 | 0 | 0.0294 |
Through Equation (18), the normalized attribute weights are calculated and listed in Table 7.
| \( \omega_1 \) | \( \omega_2 \) | \( \omega_3 \) | \( \omega_4 \) | \( \omega_5 \) | \( \omega_6 \) | \( \omega_7 \) | \( \omega_8 \) |
|---|---|---|---|---|---|---|---|
| 0.2087 | 0.0666 | 0.0402 | 0 | 0.3561 | 0.0429 | 0.0029 | 0.2826 |
It can be seen from Table 7 that the most important attribute in the dataset is \(C_5\), which corresponds to intelligent configuration. The second most important attribute is \(C_8\), driving comfort. Attribute \(C_1\), driving control, also receives a considerable weight. Attribute \(C_4\), interior decoration, receives zero weight, which indicates that the five electric cars are not sufficiently distinguished by the interior attribute in the collected reviews. This phenomenon may occur when consumers do not emphasize interior decoration in their text comments or when their sentiment scores are similar across different electric cars.
4.5 Ranking of electric cars
After obtaining the attribute weights, I aggregate the probabilistic linguistic evaluations through the PLWA operator. The aggregated decision vectors of the five alternatives and the two ideal solutions are shown in Table 8. Each cell contains a probabilistic linguistic term set expressed by its seven linguistic grades and probabilities.
| Alternative | Aggregated probabilistic linguistic vector |
|---|---|
| \(P_1\) | \(\{s_{-3}(0.1290),s_{-2}(0.0408),s_{-1}(0.0592),s_0(0.0248),s_1(0.0903),s_2(0.0469),s_3(0.6091)\}\) |
| \(P_2\) | \(\{s_{-3}(0.0363),s_{-2}(0.0121),s_{-1}(0.0214),s_0(0.0382),s_1(0.0458),s_2(0.0547),s_3(0.7915)\}\) |
| \(P_3\) | \(\{s_{-3}(0.0553),s_{-2}(0.0223),s_{-1}(0.0330),s_0(0.0435),s_1(0.0511),s_2(0.0530),s_3(0.7419)\}\) |
| \(P_4\) | \(\{s_{-3}(0.0399),s_{-2}(0.0207),s_{-1}(0.0222),s_0(0.0309),s_1(0.0509),s_2(0.0490),s_3(0.7864)\}\) |
| \(P_5\) | \(\{s_{-3}(0.0396),s_{-2}(0.0233),s_{-1}(0.0270),s_0(0.0346),s_1(0.0457),s_2(0.0598),s_3(0.7699)\}\) |
| \(AI\) | \(\{s_{-3}(0.0307),s_{-2}(0.0174),s_{-1}(0.0179),s_0(0.0329),s_1(0.0438),s_2(0.0454),s_3(0.8119)\}\) |
| \(AAI\) | \(\{s_{-3}(0.1277),s_{-2}(0.0420),s_{-1}(0.0609),s_0(0.0255),s_1(0.0899),s_2(0.0485),s_3(0.6055)\}\) |
Next, the score of each aggregated vector is calculated with the score function of probabilistic linguistic term sets. Table 9 presents the score values.
| Score | \(P_1\) | \(P_2\) | \(P_3\) | \(P_4\) | \(P_5\) | \(AI\) | \(AAI\) |
|---|---|---|---|---|---|---|---|
| \(S_i\) | 1.4833 | 2.3750 | 2.1393 | 2.3247 | 2.2827 | 2.4255 | 1.4755 |
Based on Equations (19) and (20), the utility degrees of each electric car with respect to the positive and negative ideal solutions are provided in Table 10.
| Utility degree | \(P_1\) | \(P_2\) | \(P_3\) | \(P_4\) | \(P_5\) |
|---|---|---|---|---|---|
| \(K_i^{+}\) | 0.6116 | 0.9792 | 0.8820 | 0.9585 | 0.9411 |
| \(K_i^{-}\) | 1.0053 | 1.6096 | 1.4499 | 1.5756 | 1.5471 |
Finally, the overall utility function values of the five candidate electric cars are calculated through Equation (21). Table 11 shows the ranking result.
| Alternative | \(f(K_i)\) | Rank |
|---|---|---|
| \(P_1\) | 0.7228 | 5 |
| \(P_2\) | 1.4605 | 1 |
| \(P_3\) | 1.2553 | 4 |
| \(P_4\) | 1.4163 | 2 |
| \(P_5\) | 1.3795 | 3 |
Therefore, the final preference order is \(P_2\succ P_4\succ P_5\succ P_3\succ P_1\). The AION Y receives the highest utility value, followed by the BYD Dolphin. The AION S is ranked last in the current dataset. This result suggests that the AION Y is regarded by online commenters as the most balanced electric car among the five alternatives, primarily because of its excellent intelligent configuration and comfort-related feedback.
5. Sensitivity and Comparative Analysis
In this section, I first examine how the prospect parameters influence attribute weights. Then, I compare the ranking result of the proposed model with the probabilistic linguistic TOPSIS and TODIM methods to verify the reliability of the result.
5.1 Sensitivity analysis of prospect parameters
The improved ITARA method contains three parameters: the loss aversion coefficient \(\theta\), the gain attitude coefficient \(\alpha\), and the loss attitude coefficient \(\beta\). Different consumers may have different risk attitudes, so it is important to check whether the ranking is sensitive to these parameters.
When the loss aversion coefficient \(\theta\) increases from 1 to 5, the weight of \(C_5\) gradually increases and eventually approaches 1. This means that a consumer with strong loss aversion relies mainly on intelligent configuration when choosing an electric car. The weights of \(C_1\) and \(C_8\) first increase and then decrease, while the weights of the other attributes generally decrease. This pattern confirms that the loss aversion coefficient has a substantial influence on attribute weight allocation.
When the gain attitude coefficient \(\alpha\) changes from 0 to 1, the weight of \(C_8\) is the most sensitive at first. At \(\alpha=0.2\), \(C_8\) receives a very large weight. As \(\alpha\) increases, the weight of \(C_5\) becomes more important. When \(\alpha\) approaches 1, the most important attributes are \(C_5\), \(C_8\), and \(C_1\). Consumers who focus more on gains put more emphasis on intelligent configuration and driving control.
When the loss attitude coefficient \(\beta\) varies from 0.5 to 1, the attribute \(C_5\) dominates at \(\beta=0.5\). Its weight then decreases gradually as \(\beta\) increases. Meanwhile, the weights of \(C_1\) and \(C_8\) rise. When \(\beta=1\), \(C_8\) has the largest weight. These results indicate that the proposed model is sufficiently sensitive to consumer risk preferences and can adapt to different psychological decision profiles.
5.2 Comparison with probabilistic linguistic TOPSIS
The TOPSIS method is widely used in multi-attribute decision-making. It ranks alternatives by their closeness to the positive ideal solution and their distance from the negative ideal solution. To verify my proposed method, I apply the probabilistic linguistic TOPSIS method to the same decision matrix and weights. The calculated distances from the positive and negative ideal solutions are shown in Table 12.
| Distance | \(P_1\) | \(P_2\) | \(P_3\) | \(P_4\) | \(P_5\) |
|---|---|---|---|---|---|
| \(d_i^{+}\) | 0.2576 | 0.0483 | 0.0855 | 0.0503 | 0.0586 |
| \(d_i^{-}\) | 0.0965 | 0.2394 | 0.1844 | 0.2318 | 0.2218 |
The TOPSIS ranking is \(P_2\succ P_4\succ P_3\succ P_5\succ P_1\), which is almost the same as the proposed MARCOS ranking. The only difference is that TOPSIS ranks \(P_3\) before \(P_5\), while my method ranks \(P_5\) before \(P_3\). Such a small discrepancy is reasonable because different ranking methods calculate closeness in different ways. The consistency between the two methods indicates that the proposed probabilistic linguistic MARCOS extension is reliable.
5.3 Comparison with probabilistic linguistic TODIM
The TODIM method is a behavioral multi-attribute decision-making method that explicitly accounts for decision-makers’ loss aversion. In the comparison, I use the same aggregated decision matrix and attribute weights, and I set the loss attenuation factor to one. The dominance degree of each alternative over the others is calculated under each attribute. After aggregating the dominance matrices, the overall prospect values are obtained. Table 13 reports the results.
| Prospect value | \(P_1\) | \(P_2\) | \(P_3\) | \(P_4\) | \(P_5\) |
|---|---|---|---|---|---|
| \(\xi(P_i)\) | 0 | 1.0000 | 0.1991 | 0.9420 | 0.5732 |
The TODIM ranking is \(P_2\succ P_4\succ P_5\succ P_3\succ P_1\), which is exactly the same as the ranking obtained by the proposed method. Both algorithms identify \(P_2\) as the best electric car and \(P_1\) as the least suitable option for the simulated consumer. Compared with TODIM, the proposed probabilistic linguistic MARCOS method is computationally simpler because it does not require exhaustive pairwise comparisons between alternatives. It also does not depend on a predefined loss attenuating parameter in the final ranking stage, which makes the result more objective and easier to reproduce.
6. Conclusion
This paper develops a purchase decision model for electric cars based on online reviews. The model combines probabilistic linguistic term sets with a modified ITARA weighting approach and a probabilistic linguistic MARCOS ranking framework. The entire process is data-driven and does not rely on subjective and arbitrary rating scale conversion.
The proposed model makes several contributions. First, it provides a systematic procedure for converting free-form textual review comments into quantitative probabilistic linguistic evaluations. Second, it handles multi-source website reviews by assigning reasonable website weights based on review quantity and reliability. Third, it improves the classic ITARA attribute weighting method by introducing prospect-theory-based indifference thresholds. Finally, it extends MARCOS to the probabilistic linguistic environment and obtains a stable ranking of electric car alternatives.
A case study with five popular electric cars and two automotive review websites is used to validate the model. The results show that AION Y is the most preferred electric car, while the AION S is ranked last. Sensitivity analysis demonstrates that the website weights and attribute weights can reflect different consumer risk attitudes. Comparative analysis with probabilistic linguistic TOPSIS and probabilistic linguistic TODIM confirms that the proposed method produces consistent and reasonable results.
There are also some limitations in this study. For newly launched electric cars with very few comments, the website data may not be sufficient to support reliable estimation of review probabilities. The model should therefore be enriched with expert knowledge or other information sources in such cold-start situations. In future studies, I plan to include online numerical ratings together with text reviews and to capture additional behavioral factors such as regret aversion, fairness concern, and trust in social networks. These extensions will make the decision model more comprehensive and more suitable for the complex electric car market environment.