An online review usefulness identification method based on KAM theory and WQSSVM model

Abstract

A kind of online comment usefulness identification method based on KAM theory and WQSSVM model belongs to text information mining technical field, including the following steps: S1, automatically constructs corpus;S11, design corpus annotation strategy;S12, model training is carried out using ten-fold cross validation method;S2, constructs WQSSVM model.The present application uses KAM theory and model to establish the identification system of online comment usefulness, proposes new type non-nuclear weighted quadratic surface SVM (WQSSVM) model, the model directly uses a quadratic surface in original space to identify useful online comment, overcomes the above-mentioned shortcomings of kernel-based SVM model, and by the designed weight function, different weight coefficients are given to each sample, reduce the influence of abnormal sample or noise on optimal classification hyperplane.

Description

Technical Field

[0001] This invention belongs to the field of text information mining technology, specifically relating to an online comment usefulness identification method based on KAM theory and WQSSVM model. Background Technology

[0002] The 50th Statistical Report on Internet Development in China, published by the China Internet Network Information Center (CNNIC), indicates that as of June 2022, my country had 1.051 billion internet users. Simultaneously, the rapid development of mobile internet and social media platforms has led to the emergence and massive number of online reviews, a crucial means of information exchange between businesses, platforms, and consumers. By posting reviews, users can express their opinions on their experiences with purchased products or services. Online reviews are rich in consumer feelings and expectations after experiencing products / services, covering preferences, needs, and evaluations regarding product quality, design, price, and service, thus demonstrating increasing commercial value. Furthermore, online reviews have become an important component of the customer shopping experience and significantly influence consumer decisions. For example, a 2017 survey by Podium showed that 93% of consumers said online reviews influenced their shopping decisions, 58% browsed reviews at least weekly, 82% read product reviews before making a purchase, and 68% said they were willing to pay up to 15% more for the same product or service to achieve a better shopping experience.

[0003] However, due to cross-platform interaction and the diffusion effect of information, comments on certain topics, especially hot issues, have exploded. Faced with this massive amount of comments, people often find it difficult to effectively distinguish valuable information and accurately judge the true situation of products or services. Furthermore, due to user anonymity, differing motivations for evaluating products / services, and the lack of contact on online platforms, the quality and authenticity of comments vary greatly. Currently, some e-commerce platforms contain a large amount of useless and false information in their online reviews. Merchants hire numerous online shills to write fake reviews praising their own products or disparaging competitors' products. Data shows that 42% and 36% of online product reviews on Amazon and Walmart, respectively, are fake. In addition, many merchants on e-commerce platforms organize cashback programs for positive reviews to attract consumers to leave positive feedback. Media reports have exposed the chaos of fake reviews on Amazon, with some product reviews containing a lot of irrelevant content, leading to the phenomenon of "review merging." For example, nine out of the top ten headphone brands have reviews for irrelevant products, with one headphone brand having 1386 reviews about beach umbrellas. Therefore, the quality and authenticity of online reviews cannot be guaranteed, which will seriously affect consumers' purchasing decisions. This will not only reduce the consumer shopping experience and platform satisfaction, but also create unfair competition and severely damage the platform's reputation. Furthermore, given the information overload and inconsistent authenticity of online reviews, the cost of information filtering increases dramatically, and efficiency decreases significantly, making manual screening of useful information from online reviews impractical. Therefore, how to effectively and automatically identify useful customer online reviews has become an important research topic.

[0004] Predicting the usefulness of online reviews can be viewed as a binary classification problem: does the online review contain valuable or helpful information that other consumers have seen? Existing research has considered various factors in identifying the usefulness of online reviews. Some studies have considered internal factors, such as review star ratings, the number of sentences in the review, and the sentiment of the review; others have considered external factors, such as the reviewer's identity, expertise, and reputation.

[0005] To explore more factors influencing the usefulness of online reviews, numerous theories have been proposed. Existing literature models explore the factors affecting the usefulness of online reviews from aspects such as quantity and quality, and study how these factors influence the usefulness of online reviews. Another well-known theory is the Knowledge Adoption Model (KAM), which posits that the recipient's perception of information usefulness is a direct determinant of knowledge adoption. Theoretically, KAM further realizes the knowledge transfer process, constructing a relationship network between information publishers and recipients. For example, for a given review, the determinants of perceived usefulness include information quality and the credibility of the information source. Based on KAM theory, existing literature divides information quality into four components: relevance, comprehensiveness, timeliness, and accuracy, and information source credibility into the professionalism and reputation of the information publisher. However, this study limited the research to a student population and cannot explain the actual situation of all users. Building on this, existing literature has constructed datasets covering a more comprehensive user group, dividing the factors of review usefulness into qualitative and quantitative dimensions. The above research shows that KAM theory can be used to explore important factors influencing the prediction of online review usefulness, and has achieved significant results. Based on this, this paper constructs a feature set for predicting the usefulness of online comments by comprehensively considering two aspects: the quality of online comments (including the number of sentences in the comments, the percentage of stop words, the percentage of domain words, the number of comments replies, the depth of the comments, readability, the percentage of positive sentiment words, the percentage of negative sentiment words, and text sentiment analysis) and the credibility of the comment source (including the number of followers of the commenter and the number of likes the commenter received).

[0006] Currently, many researchers model the task of predicting the usefulness of online reviews as a binary classification problem of text. Solutions to this type of problem can employ machine learning-based classification models, primarily including Support Vector Machines (SVM), decision trees, neural networks (ANN), and Naive Bayes (NB). To effectively improve classification accuracy, existing literature has developed a method combining Expectation-Maximization (EM) and NB models for document classification. Existing literature proposes a method using Multilayer Perceptron Networks (MLP) for short text classification; experimental results show that this method outperforms SVM, K-Means, and other methods. Existing literature uses ANN, SVM, and Ordinary Least Squares (OLS) to predict the usefulness of online reviews for books; empirical results show that ANN and SVM algorithms outperform OLS. Existing literature demonstrates that applying NB or decision tree algorithms to identify useful online reviews on social media platforms is effective. However, the above research indicates that existing classification methods still have some shortcomings. For example, classification methods based on models such as NB and decision trees are prone to overfitting when dealing with large datasets, high-dimensional datasets, or datasets with nonlinear relationships between features, leading to poor classification performance in practical applications. While neural network-based classification models overcome some of the shortcomings of NB and decision trees, their classification performance mainly relies on the principle of minimizing empirical risk. Furthermore, neural networks require large-scale training corpora to train their network structures, and the parameter tuning process is extremely time-consuming. To address this, existing literature proposes a text classification algorithm based on SVM using various optimization strategies, and computational results show that this algorithm has higher classification accuracy than other traditional algorithms. Existing literature also proposes a new model using component weight allocation algorithms and support vector machines to predict the usefulness of reviews, showing that support vector machines are more effective than random forests. These studies demonstrate that support vector machines have good performance in text classification, especially in identifying the usefulness of reviews.

[0007] However, the kernel function and its corresponding parameters in the SVM model have a significant impact on the classification results. For a given dataset, there is a lack of a general rule to automatically select a suitable kernel, and the selection of the appropriate kernel and its parameters can consume considerable computational time and effort. Furthermore, the inverse of the singular kernel matrix of the selected kernel limits the efficiency and accuracy of kernel-based SVM models in classifying useful online reviews. To address the limitations of kernel-based support vector machines, existing literature proposes a kernelless quadric surface support vector machine (QSSVM) method, which offers highly competitive clustering performance. For the online credit card scoring problem, existing literature proposes a kernelless quadric surface support vector machine model based on QSSVM, which has been verified to have better efficiency than the baseline model. Numerical results from existing literature show that the proposed QSSVM-based model outperforms some state-of-the-art SVM models in terms of classification accuracy.

[0008] However, online comments often contain a lot of noise or outliers, which affects the performance of classification models. Summary of the Invention

[0009] To address the aforementioned problems, this invention proposes an online comment usefulness identification method based on KAM theory and the WQSSVM model, comprising the following steps:

[0010] S1. Automatically build corpus;

[0011] S11. Design a corpus annotation strategy;

[0012] S12. Use ten-fold cross-validation for model training;

[0013] S2. Construct the WQSSVM model.

[0014] Furthermore, step S11, designing the corpus annotation strategy, is as follows: The usefulness category of the comment text is automatically labeled based on the number of likes received.

[0015] First, 100 comments were randomly selected from the dataset, and their usefulness categories were manually labeled.

[0016] Then, by setting different thresholds for the number of likes, automatic category labeling results are obtained;

[0017] Finally, the accuracy rate is obtained by comparing and analyzing the multiple sets of automatic class labeling results with the manual class labeling results.

[0018] Furthermore, in step S12, the model is trained using ten-fold cross-validation:

[0019] First, the dataset is randomly divided into a training set and a test set in a 9:1 ratio. Then, the training set is randomly divided into 10 equal parts. One part is randomly selected as the validation set, and the remaining parts are used as the training set. The classifier is trained using these 9 parts. The trained classifier is then labeled with the class labels of the validation set. The experiment is repeated 10 times to obtain 10 sets of results. The average value of these results is used as the classifier result. Finally, the classifier is tested on the test set.

[0020] Furthermore, in step S2, which constructs the WQSSVM model, the following distance-based weight function is constructed: First, given a dataset {(x i ,y i ), i = 1, ..., n}, where Represents the m features constructed, y i = +1 or -1 represents x belonging to class C1 or class C2 i The tags represent whether the online comments are helpful or not. The number of elements in C1 and C2 are n1 and n2 respectively, where n1 + n2 = n. The initial classification hyperplane obtained from the QSSVM model is... The centers of the quadric surfaces of class C1 or class C2 are respectively: Training point x i The distance d(x) between the centers of the corresponding quadric surfaces i )for:

[0021]

[0022] definition δ is a pre-given small positive number. The distance-based weighting function is expressed as follows:

[0023]

[0024] The weight r of each sample is obtained based on the above weighting function. i .

[0025] The kernelless quadric surface support vector machine model is used, and the model formula is as follows:

[0026]

[0027]

[0028] ξ i ≥0, i=1,…,n,

[0029]

[0030] The objective function maximizes the weighted sum of the relative geometric intervals of each data point. And multiply by the weight r of each newly obtained sample.i Simultaneously minimize the weighted sum of errors for each training data point. Also multiplied by the weight r i This yields a robust WQSSVM model.

[0031] Furthermore, a novel distance-based weighting function is designed for kernelless quadric surface classification hyperplanes, and the weighting coefficients are incorporated into the classification model to obtain a novel WQSSVM model. The proposed WQSSVM model is as follows:

[0032]

[0033]

[0034] ξ i ≥0, i=1,…,n,

[0035]

[0036] Where, ξ i Represents x i Misclassification error, It is a penalty constant, and matrix Q is a symmetric matrix.

[0037] Because Q is a symmetric matrix, the WQSSVM model is simplified as follows: Let the coefficient vector... For each sample point structure The sample matrix M i Check in sequence For each element, for k = 1, 2, ..., m, if The p-th element is q jk or q kj Then let M i The element in the j-th row and p-th position is Otherwise, set it to 0; I represents the identity matrix, let the sample matrix be... Weighted Hessian matrix Define the i-th sample vector Unknown variables to be solved The simplified WQSSVM model is obtained as follows:

[0038]

[0039] sty i ((s i ) T z+c)≥1-ξ i i = 1, ..., n

[0040]

[0041] in, Given the parameters, the simplified model is a convex quadratic programming problem, which is easy to solve.

[0042] The beneficial effects of this invention are as follows: This invention uses KAM theory and model to establish an online comment usefulness identification system, and proposes a novel kernel-free weighted quadratic surface SVM (WQSSVM) model. This model directly uses a quadratic surface in the original space to identify useful online comments, overcoming the above-mentioned shortcomings of kernel-based SVM models. Furthermore, by assigning different weight coefficients to each sample through the designed weight function, the influence of abnormal samples or noise on the optimal classification hyperplane is reduced. Attached Figure Description

[0043] Figure 1 This is a comparison chart of the accuracy rates of the present invention for different like thresholds;

[0044] Figure 2 This is a diagram illustrating the overall model implementation framework of the present invention;

[0045] Figure 3 This is a comparison chart of the experimental results of WQSSVM with other classifiers under the optimal feature combination of the present invention;

[0046] Figure 4 This is a comparison chart of the experimental results of WQSSVM and other classifiers under different like thresholds of the present invention;

[0047] Figure 5 This is a sensitivity analysis test chart of the WQSSVM model when the parameters are changed according to the present invention. Detailed Implementation

[0048] To improve the identification of the usefulness of online reviews, this invention proposes a method based on KAM theory and the WQSSVM model. This invention improves the SVM model with a kernel function and constructs a novel weight function, proposing the WQSSVM model to enhance the algorithm's noise tolerance. Experimental results show that this method can effectively improve the identification of the usefulness of online reviews for traditional Chinese medicine, food delivery, and mobile phone products.

[0049] The present invention adopts the following technical solution:

[0050] S1. Corpus is automatically constructed.

[0051] S1.1 Designing a Corpus Annotation Strategy. Constructing the training corpus requires a reasonable strategy for automatically annotating the crawled raw corpus. Since the number of likes represents the reader's agreement with the author's viewpoint, a high number of likes indicates high reliability of the author's comments and viewpoints, suggesting a high probability that the comments are useful (Cai Shuqin, Qin Zhiyong, Li Cuiping, Yuan Qian, Zou Bipan. Research on the Influence of Emotional Intensity on Usefulness of Negative Online Comments [J]. Management Review, 2017, 29(02):79-86.), which to some extent reflects the usefulness of the comment text. Therefore, this paper automatically annotates the category labels of the usefulness of the comment text based on the number of likes. First, 100 comments from the dataset are randomly selected and manually annotated for usefulness; then, automatic category annotation results are obtained by setting different like thresholds (3, 5, 7, 9, and 11 respectively); finally, the accuracy is obtained by comparing and analyzing the multiple sets of automatic category annotation results with the manual category annotation results. Figure 1 As shown, the accuracy rate of the automatic annotation results reached 92% when the likes threshold was set to 3, 5, 7, 9, and 11 respectively. At this point, the likes threshold was 7. Therefore, during the dataset construction process, a threshold of 7 was determined; online comments with 7 or more likes were automatically labeled as useful, and those with less than 7 were labeled as useless. Simultaneously, comparative experiments were conducted to verify other thresholds.

[0052] S1.2, Ten-Fold Cross-Validation. In actual model training, a single test set has randomness and unpredictability. To obtain reliable and stable models and experimental results, this paper adopts the ten-fold cross-validation method. First, the dataset is randomly divided into training and test sets in a 9:1 ratio. Then, the training set is randomly divided into 10 equal parts. One part is randomly selected as the validation set, and the remaining parts are used as the training set. The classifier is trained using these 9 parts. The trained classifier is then labeled with the class labels of the validation set. The experiment is repeated ten times to obtain ten sets of results. The average value of these results is used as the classifier result. Finally, the classifier is tested on the test set.

[0053] S2 and WQSSVM model construction.

[0054] Construct the following distance-based weighting function.

[0055] First, given a dataset {(x i ,y i ), i = 1, ..., n}, where Represents the m features constructed, y i = +1 or -1 represents x belonging to class C1 or class C2 iThe tags represent whether the online comments are helpful or not. C1 and C2 have n1 and n2 elements respectively, where n1 + n2 = n. The initial classification hyperplane obtained from the QSSVM model is... The centers of the quadric surfaces of class C1 or class C2 are respectively: Training point x i The distance d(x) between the centers of the corresponding quadric surfaces i )for:

[0056]

[0057] definition δ is a pre-given small positive number. The distance-based weighting function expression constructed in this invention is as follows:

[0058]

[0059] The weight r of each sample is obtained based on the above weighting function. i .

[0060] The kernelless quadric surface support vector machine model is used, and the model formula is as follows:

[0061]

[0062]

[0063] ξ i ≥0, i=1,…,n,

[0064]

[0065] The objective function maximizes the weighted sum of the relative geometric intervals of each data point. And multiply by the weight r of each newly obtained sample. i Simultaneously minimize the weighted sum of errors for each training data point. Also multiplied by the weight r i This yields a robust WQSSVM model. When data point x... i When the distance from the center of the corresponding quadratic surface is large, the sample is more likely to be an outlier. In this case, r i The value is relatively small, so an r is assigned by weighting the relative geometric interval of each data point and the error weighting of each training data point. iBy assigning weights, the impact of outliers or noise on the optimal classification hyperplane can be reduced. This invention designs a novel distance-based weight function for kernel-free quadric surface classification hyperplanes and incorporates the weight coefficients into the classification model, resulting in a novel WQSSVM model. This model can handle not only accurate data points but also data containing outliers and noise, making it more suitable for real-world applications. In summary, the proposed WQSSVM model is as follows:

[0066]

[0067]

[0068] ξ i ≥0, i=1,…,n,

[0069]

[0070] Where, ξ i Represents x i Misclassification error, It is a penalty constant, and matrix Q is a symmetric matrix.

[0071] Since Q is a symmetric matrix, the WQSSVM model can be simplified as follows: Let the coefficient vector... For each sample point structure The sample matrix M i Check in sequence For each element, for k = 1, 2, ..., m, if The p-th element is q jk or q kj Then let M i The element in the j-th row and p-th position is Otherwise, set it to 0; I represents the identity matrix, let the sample matrix be... Weighted Hessian matrix Define the i-th sample vector

[0072] Unknown variables to be solved The simplified WQSSVM model is obtained as follows:

[0073]

[0074] sty i ((s i ) T z+c)≥1-ξ i i = 1, ..., n

[0075]

[0076] in, Given the parameters, the simplified model is a convex quadratic programming problem, which is easy to solve.

[0077] The implementation framework diagram of the present invention is as follows: Figure 2 As shown. Specifically:

[0078] The first step is data preparation and data preprocessing.

[0079] To verify the effectiveness of the proposed method in identifying the usefulness of online comments, online comments in the field of Traditional Chinese Medicine (TCM) generated by users on the Zhihu official website were used as the experimental dataset. This dataset, totaling 9350 comments, was obtained from the Zhihu official website using web scraping technology. Each comment includes the text content of the TCM-themed comment, the number of likes, the number of replies, and the number of likes and followers of the author who posted the comment. Word segmentation and word frequency counting were performed during preprocessing. To construct the training and validation sets, this invention designed a reasonable corpus annotation strategy. A sample set containing 100 randomly selected online comments was manually annotated. Based on the accuracy of the automatic labeling results, a threshold of 7 for the number of likes was determined, achieving a maximum accuracy of 92%. Therefore, when constructing the dataset, comments with 7 or more likes were labeled as useful, and those with less than 7 were labeled as useless. Ten-fold cross-validation was used to construct the training, validation, and test sets.

[0080] The second step is the feature engineering stage.

[0081] In the feature engineering phase, a feature extraction method based on KAM theory is proposed, which extracts features from two aspects. First, features are extracted based on the quality of online comments, including the percentage of domain-specific terms, the percentage of stop words, the number of sentences in a comment, the number of replies, comment depth, readability, the percentage of positive sentiment words, the percentage of negative sentiment words, and text sentiment analysis. Second, features are extracted based on the credibility of the online comment source, considering the number of the author's followers and the number of likes the author received. During feature combination, all features are recombinated according to the importance of features related to comment quality and the credibility of the comment source. Specifically:

[0082] 1) Determine the importance of features. This involves removing any feature from the combination of comment quality and comment source credibility, and observing the degree of decline in the recognition performance of the model trained with the new feature combination. The greater the decline in recognition performance, the more important that feature is for predicting the usefulness of the comment, thus determining the importance of each feature to the usefulness of the comment.

[0083] 2) Reconstruct the feature set. Features were added to the feature set one by one in descending order of importance, resulting in 20 different feature combinations. Different feature combinations were used to train comment usefulness prediction models. The feature combination used to train the model with the highest prediction accuracy was considered the optimal feature combination. The optimal feature combination obtained in the experiment was the number of sentences, the proportion of domain terms, the number of replies, readability, and comment depth, with an accuracy of 99.14%, an improvement of 0.21% compared to all other feature combinations, and an improvement of 0.1% to 0.96% compared to other feature combinations.

[0084] The third step is the classifier construction phase.

[0085] This invention proposes the WQSSVM model to identify the usefulness of online comments in the field of traditional Chinese medicine by improving the SVM model with a kernel function and designing a weight function to characterize the importance of samples. To verify the recognition effect of the proposed method, a comparative experiment was conducted. The dataset used in the experiment consisted of online comments related to traditional Chinese medicine posted by users on the Zhihu official website. The WQSSVM model obtained the optimal feature combination on this dataset, including the number of sentences, the proportion of domain terms, the number of replies, readability, and comment depth, achieving an accuracy of 99.14%, an improvement of 0.21% compared to the accuracy of all feature combinations. Simultaneously, the accuracy of NB, SVM, MLP, and RF classifiers trained with the optimal feature combination also improved, such as... Figure 3 As shown, the accuracy of the WQSSVM classifier is 17.41%, 2.38%, 2.33%, and 1.95% higher than the other four classifiers, respectively, further verifying the superiority of training the WQSSVM model based on the optimal feature combination proposed in this invention.

[0086] The fourth step is robustness and sensitivity analysis.

[0087] To verify the robustness of the results of this invention, experiments were also conducted on a traditional Chinese medicine dataset with all features for other like thresholds (2 to 10), and the F1 value was observed. Figure 4 The results show that after training the model at different thresholds, the F1 score of the WQSSVM classifier is significantly better than other classifiers. At the same threshold, WQSSVM can improve performance by 0.67% to 9.83% compared to other models. Therefore, the WQSSVM classifier of this invention is robust.

[0088] Furthermore, by varying the parameters in the proposed WQSSVM model, a set of sensitivity analysis tests were conducted based on a traditional Chinese medicine dataset with the optimal feature combination. The classification accuracy under different parameters is as follows: Figure 5As shown. In this invention, 10-fold cross-validation is used to select the optimal parameters, where the parameter η ranges from log₂η∈{25,26,…,55}. Figure 5 As shown, within a certain error range, the accuracy of the validation set increases with increasing η, eventually stabilizing. When log₂η = 25, the lowest accuracy of the validation set is 0.9559, while when log₂η = 47, the highest accuracy is 0.9667. Therefore, this invention selects log₂η = 47 as the optimal parameter.

[0089] The fifth step is validation based on two other online review datasets from different domains.

[0090] To further validate the performance of the proposed WQSSVM, this invention was tested using online review data from two different domains. Review datasets for food delivery and mobile phone products were constructed from Meituan and JD.com, respectively. Features predicting review usefulness were then developed from two perspectives: the quality of the online review content and the credibility of the review source. Due to differences in website interfaces, similar variables were used to replace some features of online reviews related to traditional Chinese medicine, such as whether the reviewer is a repeat customer and the review level, in different datasets. Experimental results on the two different datasets show that, for most metrics, WQSSVM still produces more accurate predictions of review usefulness than other test classifiers.

[0091] This paper presents an online comment usefulness identification method based on the Knowledge Adoption Model (KAM) theory and the Non-kernel Weighted Quadratic Surface Support Vector Machine (WQSSVM) model. The method comprises two parts: feature engineering and classification model construction. Specifically, based on KAM theory, a feature set for online comment usefulness identification is constructed comprehensively from two aspects: comment content quality and comment source credibility. Different feature combinations are used to train the WQSSVM model, continuously optimizing its classification performance. Compared with other existing classic and commonly used text classification models or methods, the online comment usefulness identification model based on KAM theory and WQSSVM exhibits superior classification performance.

[0092] To verify the recognition effect of the proposed method, multiple sets of comparative experiments were conducted using online comment data in the field of traditional Chinese medicine generated by users on Zhihu's official website, Meituan Waimai's comment data, and JD.com's mobile phone product comment data.

[0093] To improve the identification of the usefulness of online reviews, this invention improves the SVM model with a kernel function and constructs a novel weight function, proposing the WQSSVM model to enhance the algorithm's noise tolerance. Experimental results show that this method can effectively improve the identification of the usefulness of online reviews for traditional Chinese medicine, food delivery, and mobile phone products.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying the usefulness of online reviews based on KAM theory and the WQSSVM model, characterized in that, Includes the following steps: S1. Automatically build corpus; S11. Design a corpus annotation strategy; S12. Use ten-fold cross-validation for model training; S2. Construct the WQSSVM model; S3. Feature optimization and determination of the optimal feature combination; S4. Model training and parameter optimization; S5. Practical application of identifying the usefulness of online reviews; Step S11, designing the corpus annotation strategy, is as follows: Automatically label the usefulness of the comment text with category tags based on the number of likes. First, 100 comments were randomly selected from the dataset, and their usefulness categories were manually labeled. Then, by setting different thresholds for the number of likes, automatic category labeling results are obtained; Finally, the accuracy rate is obtained by comparing and analyzing the multiple sets of automatic class labeling results with the manual class labeling results. In step S2, which constructs the WQSSVM model, the following distance-based weight function is built: First, given a dataset... ,in Represents the m features constructed. or Represents belonging to class C1 or class C2 The tags represent whether the online comments are helpful or not. The number of elements in C1 and C2 are n1 and n2 respectively, where n1 + n2 = n. The initial classification hyperplane obtained from the QSSVM model is... Then the centers of the quadric surfaces of class C1 or class C2 are respectively: , Training points and the center distance of the corresponding quadric surface for: definition , Given a positive number, the distance-based weight function expression is constructed as follows: The weight of each sample is obtained based on the above weighting function. ; The kernelless quadric surface support vector machine model is used, and the model formula is as follows: The objective function maximizes the weighted sum of the relative geometric intervals of each data point. And multiply by the weight of each newly obtained sample. Simultaneously minimize the weighted sum of errors for each training data point. Also multiplied by the weight This yields a robust WQSSVM model; We design a novel distance-based weighting function for kernelless quadric surface classification hyperplanes, and incorporate the weighting coefficients into the classification model to obtain a novel WQSSVM model. The proposed WQSSVM model is as follows: in, represent Misclassification error, It is a penalty constant, a matrix It is a symmetric matrix. ; Because Q is a symmetric matrix, the WQSSVM model is simplified as follows: Let the coefficient vector... For each sample point ,structure sample matrix Check in sequence For each element, for ,if The The element is or Then let The Line 1 The elements are Otherwise, set it to 0; Let the identity matrix be the sample matrix. Weighted Hessian matrix ; Define the first sample vectors Unknown variables to be solved The simplified WQSSVM model is obtained as follows: in, Given the parameters, the simplified model is a convex quadratic programming problem, which is easy to solve.

2. The online review usefulness identification method based on KAM theory and WQSSVM model as described in claim 1, characterized in that, Step S12 involves training the model using ten-fold cross-validation. First, the dataset is randomly divided into a training set and a test set in a 9:1 ratio. Then, the training set is randomly divided into 10 equal parts. One part is randomly selected as the validation set, and the remaining parts are used as the training set. The classifier is trained using these 9 parts. The trained classifier is then labeled with the class labels of the validation set. The experiment is repeated 10 times to obtain 10 sets of results. The average value of these results is used as the classifier result. Finally, the classifier is tested on the test set.

Images