In the era of digital economy, the rapid advancement of big data and artificial intelligence (AI) technologies is reshaping decision-making paradigms across various industries. Consumer decision-making patterns are gradually shifting from traditional experience reliance to data-driven intelligent decisions. Online reviews, as product information created by consumers based on usage experience, not only provide authentic experiences and opinions but also play a key role in the shopping process. However, the plethora of branded products on the market often leaves consumers facing information overload and choice difficulties during purchasing decisions. Although online reviews offer rich information references for decision-making, their voluminous and uneven quality content makes it challenging for consumers to efficiently extract key information, thereby reducing decision efficiency. Therefore, constructing an efficient and reliable decision support model to assist consumers in selecting the best options in complex information environments has become a critical research issue. This paper addresses the semantic conflicts and credibility calibration issues in online review data fusion, proposing an integrated decision support framework of “Conflict Resolution-Data-Driven-Risk Adaptation” to achieve comprehensive information utilization.
Existing research often focuses on single data sources, such as textual reviews, lacking integration and comprehensive use of data. This limitation leads to decision defects like missing information dimensions and misjudgment of credibility. Although some recent studies have addressed this issue, current methods have significant limitations. For instance, linear fusion methods with weighted coefficients force linear alignment between textual reviews and star ratings, ignoring nonlinear semantic associations between different data. In contrast, modular approaches that process attribute-specific evaluation information separately effectively avoid information loss during unified conversion of mixed information, a concept worth further exploration. Moreover, product ranking methods primarily rely on multi-attribute decision-making techniques, including TOPSIS, VIKOR, and their variants or combinations. However, these methods fail to incorporate consumer psychology into decisions, potentially leading to decision biases. To quantify consumers’ risk aversion psychology, this study introduces the TODIM model to assist in product ranking for EV cars.
Based on the above analysis, this study proposes a product ranking model based on the fusion of online review data, comprising four modules: data collection, data fusion, attribute weight determination, and product ranking. The data fusion module tackles the challenges of semantic conflicts and credibility calibration in online reviews.
Data Collection and Fusion
We collected online review data using web crawlers, targeting platforms like automotive vertical media sites. By inputting URLs to locate target interfaces, we obtained textual reviews, star ratings, and interactive behavior data. During data preprocessing, we enhanced data quality by removing special characters, deleting duplicate reviews, and other operations. For textual reviews, we employed Jieba for Chinese word segmentation and eliminated stop words to focus on key information.
The data fusion module is based on D-S evidence theory, implemented in three stages. In Stage 1, we used SnowNLP for sentiment analysis of textual reviews to obtain sentiment values. Both sentiment values and star ratings were converted into Basic Probability Assignments (BPA) according to rules where the main term has high probability, and adjacent term sets have low probability, as shown in Table 1.
| Rule | Sentiment Value \( s_i \) | Sentiment BPA | Star Rating \( v_i \) | Star BPA |
|---|---|---|---|---|
| Very negative, allowing minor negativity | [0, 0.2) | \( m_s(\{L_1\}) = 0.8, m_s(\{L_1, L_2\}) = 0.2 \) | 1 star | \( m_v(\{L_1\}) = 0.8, m_v(\{L_1, L_2\}) = 0.2 \) |
| Mainly negative, possibly near neutral or very negative | [0.2, 0.4) | \( m_s(\{L_2\}) = 0.7, m_s(\{L_1, L_2, L_3\}) = 0.3 \) | 2 stars | \( m_v(\{L_2\}) = 0.7, m_v(\{L_1, L_2, L_3\}) = 0.3 \) |
| Mainly neutral, allowing minor negative or positive | [0.4, 0.6) | \( m_s(\{L_3\}) = 0.8, m_s(\{L_2, L_3, L_4\}) = 0.2 \) | 3 stars | \( m_v(\{L_3\}) = 0.8, m_v(\{L_2, L_3, L_4\}) = 0.2 \) |
| Mainly positive, possibly near neutral or very positive | [0.6, 0.8) | \( m_s(\{L_4\}) = 0.7, m_s(\{L_3, L_4, L_5\}) = 0.3 \) | 4 stars | \( m_v(\{L_4\}) = 0.7, m_v(\{L_3, L_4, L_5\}) = 0.3 \) |
| Very positive, allowing minor possibility of positive | [0.8, 1] | \( m_s(\{L_5\}) = 0.9, m_s(\{L_4, L_5\}) = 0.1 \) | 5 stars | \( m_v(\{L_5\}) = 0.9, m_v(\{L_4, L_5\}) = 0.1 \) |
In Stage 2, we considered that user trust in reviews may be influenced by both interactive behaviors and content consistency. Thus, low interaction and high conflict were used as indicators of fake reviews. The Interactive Behavior Weight (IBW) was calculated by integrating likes, replies, page views, and other interactive behavior data, representing the degree of interaction. Conflict degree was computed using sentiment values and star ratings. A Softmax-based adaptive weighting mechanism was introduced, consistent with the dynamic weight allocation idea in attention mechanisms, avoiding the uncertainty of manual weighting. The page view function \( f(PV) \) was defined as:
$$ f(PV) = \begin{cases}
2 \times PV, & \text{if } PV \leq 0.5 \\
2 – 2 \times PV, & \text{if } PV > 0.5
\end{cases} $$
The suspicion index \( X_i \) was calculated, and the adjustment factor \( \alpha_i = 1 / (1 + X_i) \) was used to discount the textual and rating BPA, thereby down-weighting fake reviews. The discounted probability was assigned to global uncertainty \( \Theta \). The formulas are:
$$ IBW_i = (P(L_i) + P(R_i)) \times f(PV_i) $$
$$ X_i = \lambda_1 F_1 + \lambda_2 F_2, \quad F_1 = 1 – IBW_i, \quad F_2 = |s_i – v’_i|, \quad v’_i = v_i / 5 $$
where \( P(L_i) \) and \( P(R_i) \) are the percentages of likes and replies in the full dataset that are less than or equal to those of the i-th review, \( PV_i \) is the normalized page view rank, and \( \lambda_1, \lambda_2 \) are influence degrees of interaction and conflict, with \( \lambda_1 + \lambda_2 = 1 \).
In Stage 3, D-S evidence theory was used to fuse individual reviews. First, the conflict factor \( K \) was calculated, and the fused BPA was obtained. Then, all reviews for the same attribute were aggregated. Finally, the results were converted into a five-granularity probabilistic linguistic term set, where \( L_1, L_2, L_3, L_4, L_5 \) correspond to \( S_0, S_1, S_2, S_3, S_4 \), respectively. The probability of composite propositions was evenly distributed among sub-terms, while global uncertainty \( \Theta \) was retained. The fusion formula is:
$$ K = \sum_{B \cap C = \emptyset} m_1(B) m_2(C) $$
$$ m(A) = \frac{1}{1 – K} \sum_{B \cap C = A} m_1(B) m_2(C) $$
Data-Driven Decision Model Construction
Attribute Weight Determination Based on BWM
The Best-Worst Method (BWM) is commonly used to subjectively determine attribute weights. Traditional product ranking methods use subjective or objective weighting methods without fully exploiting the potential weight information in online reviews. Here, we utilized online reviews to define the best and worst indicators in the model, obtaining objective attribute weights. First, we calculated the word frequencies of satisfied and dissatisfied reviews, selected the top 100 high-frequency words for screening and synonym annotation, and mapped them to various attributes. Then, we calculated the cumulative relative frequency of each attribute in satisfied and dissatisfied reviews, taking the highest values as the most satisfied and most dissatisfied attributes. Next, based on the sum of star ratings and frequencies, we calculated the preference values of the most satisfied attribute relative to attribute j, \( r_{Bj}^i \), and attribute j relative to the most dissatisfied attribute, \( r_{jW}^i \). The preference values were scaled to the [1,9] interval to obtain preference matrices \( E_{Bj}^i \) and \( E_{jW}^i \). Finally, we used Python’s scipy.optimize.minimize function with the SLSQP algorithm to solve the following optimization model, obtaining the optimal weights for each attribute \( (\lambda_1^i, \lambda_2^i, \ldots, \lambda_m^i) \):
$$ r_{Bj}^i = \frac{ssr(f_B^i) + s_B^i \omega_B^i}{ssr(f_j^i) + s_j^i \omega_j^i}, \quad r_{jW}^i = \frac{ssr(f_j^i) + s_j^i \omega_j^i}{ssr(f_W^i) + s_W^i \omega_W^i} $$
$$ E_{Bj}^i = \frac{5 \times (r_{Bj}^i – 0.5)}{3}, \quad E_{jW}^i = \frac{5 \times (r_{jW}^i – 0.5)}{3} $$
$$ \min \delta $$
$$ \text{s.t.} \quad \left| \frac{\lambda_B^i}{\lambda_j^i} – E_{Bj}^i \right| \leq \delta, \quad \left| \frac{\lambda_j^i}{\lambda_W^i} – E_{jW}^i \right| \leq \delta, \quad \sum_{j=1}^m \lambda_j^i = 1, \quad \lambda_j^i \geq 0 $$
where \( ssr(f_B^i) \) and \( ssr(f_W^i) \) are the sums of star ratings for the most satisfied and most dissatisfied attributes of product i, \( ssr(f_j^i) \) is the sum of star ratings for attribute j of product i, and \( \omega_B^i, \omega_W^i \) are the cumulative relative frequencies of the most satisfied and most dissatisfied attributes of product i.
Product Ranking Based on PL-H-TODIM
We used Probabilistic Linguistic Term Sets (PLTS) to express user evaluation information. The TODIM method, an extension of prospect theory, quantifies decision-makers’ non-rational psychological preferences by introducing a risk aversion coefficient \( \theta \), capturing the decision behavior of risk-averse users. The proposed PL-H-TODIM model integrates PLTS with Hellinger distance measure, fully retaining probability distribution information. Hellinger distance is more sensitive to minor differences in probability distributions compared to Euclidean distance or KL divergence, effectively measuring semantic distribution differences. First, all weight values \( \omega_j^i \) for the same attribute were summed to get \( \omega_j \), then normalized to \( \overline{\omega}_j \). The relative weight value \( \omega_{jm} \) was calculated using the reference attribute \( \omega_{\max} \). Then, the gain and loss of product \( A_i \) relative to product \( A_k \) under attribute j were computed. The dominance degree of product \( A_i \) relative to product \( A_k \) was calculated, followed by the comprehensive dominance degree of product \( A_i \), and finally, the optimal ranking was obtained. The formulas are:
$$ \overline{\omega}_j = \frac{\omega_j}{\sum_{j=1}^m \omega_j}, \quad \omega_{jm} = \frac{\overline{\omega}_j}{\omega_{\max}} $$
$$ \phi_j(A_i, A_k) = \begin{cases}
\sqrt{\frac{\omega_{jm}}{\sum_{j=1}^m \omega_{jm}}} \cdot H_d(L(p)_{ij}, L(p)_{kj}), & \text{if } L(p)_{ij} > L(p)_{kj} \\
0, & \text{if } L(p)_{ij} \sim L(p)_{kj} \\
-\frac{1}{\theta} \sqrt{\frac{\sum_{j=1}^m \omega_{jm}}{\omega_{jm}}} \cdot H_d(L(p)_{ij}, L(p)_{kj}), & \text{if } L(p)_{ij} < L(p)_{kj}
\end{cases} $$
$$ \Phi(A_i, A_k) = \sum_{j=1}^m \phi_j(A_i, A_k) $$
$$ \delta(A_i) = \frac{\sum_{k=1}^n \Phi(A_i, A_k) – \min_i \sum_{k=1}^n \Phi(A_i, A_k)}{\max_i \sum_{k=1}^n \Phi(A_i, A_k) – \min_i \sum_{k=1}^n \Phi(A_i, A_k)} $$
where \( H_d(L(p)_{ij}, L(p)_{kj}) \) is the Hellinger distance between the PLTS of products i and k under attribute j.
Case Study: Ranking EV Cars
The rise of professional automotive evaluation platforms provides important references for consumers purchasing EV cars. We collected online review data from a leading domestic automotive vertical media platform. Suppose a consumer plans to purchase a new energy EV car priced between 250,000 and 350,000 yuan, with six alternative models A1 to A6. We crawled online reviews for these models from July 12, 2022, to March 8, 2025, classifying them into seven attributes: space (C1), driving experience (C2), range (C3), appearance (C4), interior (C5), cost-effectiveness (C6), and intelligence (C7). Given the consumer’s low loss sensitivity and adventurous decision characteristics, the risk aversion coefficient \( \theta \) was set to 0.2.
We collected 3454 online reviews, ensuring the sample size difference between models was less than 30%. After data cleaning, 29007 data points were obtained. Sentiment analysis was performed on the textual reviews for each attribute of the six alternative EV cars, and sentiment values were converted into textual and rating BPA. The suspicion index was calculated, and adjustment factors were used to adjust the BPA. The conflict factor was computed to fuse individual reviews, and BPA for multiple reviews under the same attribute were aggregated. Finally, the fused BPA for each model’s attributes were converted into probabilistic linguistic term sets, as shown in Table 5.
| Model | C1 | C2 | C3 | C4 | C5 | C6 | C7 |
|---|---|---|---|---|---|---|---|
| A1 | {(s0, 0.0503), (s1, 0.0268), (s2, 0.0316), (s3, 0.0682), (s4, 0.6840), (Θ, 0.1390)} | {(s0, 0.0239), (s1, 0.0098), (s2, 0.0205), (s3, 0.0489), (s4, 0.7848), (Θ, 0.1121)} | {(s0, 0.0557), (s1, 0.0276), (s2, 0.0250), (s3, 0.0665), (s4, 0.6817), (Θ, 0.1435)} | {(s0, 0.0059), (s1, 0.0048), (s2, 0.0061), (s3, 0.0283), (s4, 0.8622), (Θ, 0.0926)} | {(s0, 0.0131), (s1, 0.0056), (s2, 0.0191), (s3, 0.0677), (s4, 0.7927), (Θ, 0.1018)} | {(s0, 0.0260), (s1, 0.0118), (s2, 0.0317), (s3, 0.1514), (s4, 0.6546), (Θ, 0.1245)} | {(s0, 0.0219), (s1, 0.0127), (s2, 0.0186), (s3, 0.0476), (s4, 0.7918), (Θ, 0.1073)} |
| A2 | {(s0, 0.0843), (s1, 0.0350), (s2, 0.0499), (s3, 0.1566), (s4, 0.5772), (Θ, 0.0970)} | {(s0, 0.0272), (s1, 0.0236), (s2, 0.0241), (s3, 0.0509), (s4, 0.8182), (Θ, 0.0560)} | {(s0, 0.1627), (s1, 0.0468), (s2, 0.0758), (s3, 0.1903), (s4, 0.3782), (Θ, 0.1462)} | {(s0, 0.0139), (s1, 0.0068), (s2, 0.0108), (s3, 0.0250), (s4, 0.9095), (Θ, 0.0340)} | {(s0, 0.0302), (s1, 0.0163), (s2, 0.0197), (s3, 0.0555), (s4, 0.8269), (Θ, 0.0514)} | {(s0, 0.0532), (s1, 0.0217), (s2, 0.0361), (s3, 0.0890), (s4, 0.7273), (Θ, 0.0727)} | {(s0, 0.0313), (s1, 0.0184), (s2, 0.0328), (s3, 0.0592), (s4, 0.8005), (Θ, 0.0579)} |
| A3 | {(s0, 0.0935), (s1, 0.0351), (s2, 0.0416), (s3, 0.1534), (s4, 0.5661), (Θ, 0.1102)} | {(s0, 0.0327), (s1, 0.0185), (s2, 0.0201), (s3, 0.0722), (s4, 0.7919), (Θ, 0.0628)} | {(s0, 0.1526), (s1, 0.0443), (s2, 0.0646), (s3, 0.2261), (s4, 0.3692), (Θ, 0.1432)} | {(s0, 0.0157), (s1, 0.0125), (s2, 0.0110), (s3, 0.0320), (s4, 0.8819), (Θ, 0.0468)} | {(s0, 0.0249), (s1, 0.0121), (s2, 0.0139), (s3, 0.0644), (s4, 0.8314), (Θ, 0.0533)} | {(s0, 0.0451), (s1, 0.0192), (s2, 0.0341), (s3, 0.1115), (s4, 0.7158), (Θ, 0.0741)} | {(s0, 0.0276), (s1, 0.0185), (s2, 0.0203), (s3, 0.0720), (s4, 0.7983), (Θ, 0.0595)} |
| A4 | {(s0, 0.0893), (s1, 0.0264), (s2, 0.0404), (s3, 0.0953), (s4, 0.6558), (Θ, 0.0928)} | {(s0, 0.0459), (s1, 0.0202), (s2, 0.0512), (s3, 0.0913), (s4, 0.7302), (Θ, 0.0613)} | {(s0, 0.1033), (s1, 0.0425), (s2, 0.0596), (s3, 0.1272), (s4, 0.5586), (Θ, 0.1070)} | {(s0, 0.0097), (s1, 0.0073), (s2, 0.0246), (s3, 0.1038), (s4, 0.8240), (Θ, 0.0306)} | {(s0, 0.0274), (s1, 0.0128), (s2, 0.0382), (s3, 0.0959), (s4, 0.7807), (Θ, 0.0450)} | {(s0, 0.0456), (s1, 0.0233), (s2, 0.0374), (s3, 0.0967), (s4, 0.7369), (Θ, 0.0602)} | {(s0, 0.0479), (s1, 0.0271), (s2, 0.0324), (s3, 0.0589), (s4, 0.7744), (Θ, 0.0594)} |
| A5 | {(s0, 0.0882), (s1, 0.0387), (s2, 0.0364), (s3, 0.0824), (s4, 0.6600), (Θ, 0.0943)} | {(s0, 0.0472), (s1, 0.0315), (s2, 0.0442), (s3, 0.1208), (s4, 0.6917), (Θ, 0.0624)} | {(s0, 0.1247), (s1, 0.0371), (s2, 0.0317), (s3, 0.0798), (s4, 0.6087), (Θ, 0.1158)} | {(s0, 0.0272), (s1, 0.0175), (s2, 0.0490), (s3, 0.1648), (s4, 0.7013), (Θ, 0.0402)} | {(s0, 0.0670), (s1, 0.0279), (s2, 0.0473), (s3, 0.1558), (s4, 0.6346), (Θ, 0.0674)} | {(s0, 0.0374), (s1, 0.0221), (s2, 0.0278), (s3, 0.0509), (s4, 0.8129), (Θ, 0.0489)} | {(s0, 0.0260), (s1, 0.0171), (s2, 0.0327), (s3, 0.0521), (s4, 0.8314), (Θ, 0.0408)} |
| A6 | {(s0, 0.0601), (s1, 0.0238), (s2, 0.0263), (s3, 0.0732), (s4, 0.7268), (Θ, 0.0899)} | {(s0, 0.0270), (s1, 0.0177), (s2, 0.0290), (s3, 0.1328), (s4, 0.7311), (Θ, 0.0625)} | {(s0, 0.0776), (s1, 0.0344), (s2, 0.0507), (s3, 0.1151), (s4, 0.6141), (Θ, 0.1082)} | {(s0, 0.0049), (s1, 0.0031), (s2, 0.0096), (s3, 0.0689), (s4, 0.8748), (Θ, 0.0386)} | {(s0, 0.0260), (s1, 0.0138), (s2, 0.0260), (s3, 0.1539), (s4, 0.7196), (Θ, 0.0607)} | {(s0, 0.0378), (s1, 0.0151), (s2, 0.0333), (s3, 0.0946), (s4, 0.7520), (Θ, 0.0672)} | {(s0, 0.0336), (s1, 0.0173), (s2, 0.0274), (s3, 0.0740), (s4, 0.7830), (Θ, 0.0647)} |
Based on satisfied and dissatisfied reviews, word frequencies were calculated, and high-frequency words were screened and annotated with synonyms. After mapping high-frequency words to the seven attributes, the cumulative relative frequency of each attribute was calculated, with the highest values taken as the most satisfied and most dissatisfied attributes. The preference matrices \( E_{Bj} \) and \( E_{jW} \) were computed based on the sum of star ratings, and the BWM model was used to obtain the weights for each attribute, as shown in Table 8.
| Model | C1 | C2 | C3 | C4 | C5 | C6 | C7 |
|---|---|---|---|---|---|---|---|
| A1 | 0.1438 | 0.1452 | 0.1446 | 0.1519 | 0.1426 | 0.1286 | 0.1433 |
| A2 | 0.1349 | 0.1489 | 0.1254 | 0.1545 | 0.1471 | 0.1415 | 0.1476 |
| A3 | 0.1373 | 0.1481 | 0.1223 | 0.1557 | 0.1486 | 0.1406 | 0.1475 |
| A4 | 0.1461 | 0.1413 | 0.1393 | 0.1379 | 0.1426 | 0.1413 | 0.1514 |
| A5 | 0.1473 | 0.1394 | 0.1455 | 0.1316 | 0.1333 | 0.1479 | 0.1550 |
| A6 | 0.1527 | 0.1382 | 0.1429 | 0.1455 | 0.1337 | 0.1415 | 0.1455 |
The weights were normalized, and relative weight values were calculated. With \( \theta = 0.2 \), the dominance degree of model \( A_i \) relative to model \( A_k \) under each attribute was computed. Finally, the comprehensive dominance degrees for each model were: \( \delta(A1) = 1 \), \( \delta(A2) = 0.12 \), \( \delta(A3) = 0.14 \), \( \delta(A4) = 0.05 \), \( \delta(A5) = 0 \), \( \delta(A6) = 0.45 \), resulting in the ranking: A1 > A6 > A3 > A2 > A4 > A5. Thus, A1 is the best choice for the consumer.
Comparative Analysis
With \( \theta = 0.2 \), we compared the proposed method with other ranking methods, as shown in Table 10. In contrast, this study has advantages:
| Method | Alternative EV Car Ranking |
|---|---|
| 1: Three-stage review rating data fusion based on D-S evidence theory | A1 > A6 > A3 > A2 > A4 > A5 |
| 2: Based solely on textual review data | A2 > A1 > A3 > A5 > A4 > A6 |
| 3: Based solely on star rating data | A1 > A2 > A5 > A3 > A6 > A4 |
| 4: Linear fusion of review rating data based on PLWA | A1 > A2 > A5 > A3 > A4 > A6 |
a) Decision results are more robust. Method 1’s three-stage fusion mechanism integrates reviews and ratings. In contrast, Methods 2 and 3 rely on single data sources and are susceptible to data sparsity, while Method 4 does not fully explore the deep associations between text and ratings. In textual sentiment analysis, continuous probability values are often discretized into intervals, leading to information loss. A2 is overrated in Method 2, possibly due to subtle emotional differences not being captured; star ratings are coarse-grained quantifications, making it difficult to reflect隐含 emotions. A5 is highly ranked in Method 3, possibly because high ratings mask negative details in the text.
b) Optimized application of D-S evidence theory. Method 1’s three-stage fusion handles data uncertainty more effectively through evidence synthesis. In contrast, Method 4’s linear fusion assumes data independence and no conflict, which is unrealistic. In Method 4, A5 is highly ranked because it does not detect conflicts between data and forces fusion, while Method 1 corrects the ranking bias in Method 4 by allocating conflict weights through evidence theory, achieving a more reasonable ranking.
In real purchase decisions, adventurous consumers can tolerate larger expectation gaps, meaning the smaller \( \theta \), the lower the decision-maker’s sensitivity to loss. Conversely, the larger \( \theta \), the more risk-averse the decision-maker. This study conducted sensitivity analysis by changing \( \theta \) values (0.2, 1, 2, 5, 15, 50, 100) to explore product recommendation differences under different risk tolerance levels. The comprehensive dominance degrees are shown in Figure 1, and ranking results are in Table 11. Clearly, the product ranking support model is sensitive to parameter \( \theta \). Methodologically, the model can flexibly adapt to different decision scenarios by adjusting \( \theta \); in application, different \( \theta \) values reflect the heterogeneity of consumers’ risk attitudes.
| Parameter Value \( \theta \) | Alternative EV Car Ranking |
|---|---|
| 0.2 | A1 > A6 > A3 > A2 > A4 > A5 |
| 1 | A1 > A6 > A3 > A2 > A4 > A5 |
| 2 | A1 > A6 > A3 > A2 > A4 > A5 |
| 5 | A1 > A6 > A3 > A2 > A5 > A4 |
| 15 | A1 > A6 > A3 > A5 > A2 > A4 |
| 50 | A1 > A6 > A5 > A3 > A2 > A4 |
| 100 | A1 > A6 > A5 > A3 > A2 > A4 |
Conclusion
This study addresses the semantic conflicts and credibility calibration issues in online review data fusion, proposing an integrated decision support framework of “Conflict Resolution-Data-Driven-Risk Adaptation,” providing new methodological support for information fusion and consumer decision optimization. The framework introduces D-S evidence theory to construct BPA mapping rules, effectively fusing reviews and ratings; proposes an interactive-conflict dual-path suspicion model, designing an evidence credibility decay mechanism to achieve fake review identification and evidence credibility correction; finally, through a data-driven BWM attribute weight allocation model and a PL-H-TODIM risk perception ranking model, provides more realistic personalized recommendation schemes for consumers with differentiated risk aversion. Through case analysis of six alternative EV cars, the effectiveness of the online review data fusion method was verified; by comparing with other product ranking methods, the superiority of the proposed method was demonstrated; by examining the impact of \( \theta \) changes on ranking results, sensitivity analysis was completed. Future research can expand to multimodal data fusion (e.g., videos) and dynamic review update mechanisms to enhance the model’s universality and timeliness.