A method for detecting fake reviews based on big data

Abstract

This invention proposes a method for detecting fake reviews based on big data, comprising: acquiring a review dataset, wherein the data in the review dataset includes review texts from multiple users and corresponding user behavior data, and performing data preprocessing; extracting review text features from the review dataset, and extracting user behavior features and review behavior features from the user behavior data; defining relationship categories based on the review text and user behavior features, and constructing a multidimensional relationship enhancement graph (MREGC) based on multiple relationships; using a dynamic adaptive feature enhancement graph neural network to learn features on the MREGC relationship graph, obtaining the embedding representation of each node in the multidimensional relationship enhancement graph, and aggregating the embedding representations; inputting the aggregated features into a classifier to determine whether a review is a fake review, and outputting a fake review label if the review is classified as a fake review; this invention, by fusing residual networks and multi-relationship review graphs, fully considers the multidimensional relationships and deep features between reviews, enabling the model to more accurately identify fake reviews, and significantly improving the accuracy and robustness of fake review detection.

Description

Technical Field

[0001] This invention relates to the fields of big data processing technology and fake review detection technology, and in particular to a method for detecting fake reviews based on big data. Background Technology

[0002] With the rapid development of e-commerce and short video platforms, more and more users tend to rely on online reviews as a crucial basis for their purchasing decisions. These reviews not only influence users' shopping choices but also shape brand image to some extent. However, the proliferation of fake reviews has become a serious problem in the current internet environment. Some users or businesses publish fake reviews with the intention of misleading potential consumers, increasing the attractiveness of their products or services, or discrediting competitors, thereby negatively impacting the decisions of other consumers.

[0003] Fake reviews take many forms, including but not limited to fabricated positive reviews, malicious negative reviews, and fake order reviews. These reviews not only make it difficult for consumers to distinguish genuine information but also damage the platform's credibility and affect its normal operation. Over time, the number of fake reviews has been increasing, becoming one of the major problems for major e-commerce platforms, social media, and short video communities.

[0004] Currently, although many platforms and companies have adopted various methods to identify and remove fake reviews, traditional detection methods still have certain limitations due to the complexity and diversity of review data. Review data often exhibits imbalanced labeling, with fake reviews being few in number and unevenly distributed, posing challenges to the training and generalization capabilities of detection algorithms. Traditional keyword- or rule-based detection methods often cannot effectively handle this type of data, and with the surge in review volume, manual review struggles to meet the demands of real-time performance and scalability. Therefore, how to effectively identify fake reviews from large amounts of review data, especially how to handle imbalanced samples, complex user behavior data, and complex relationships between reviews, has become a current research hotspot in the field of data mining. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method for detecting fake reviews based on big data. The method includes: acquiring a review dataset, wherein the data in the review dataset includes review texts from multiple users and corresponding user behavior data, and performing data preprocessing; extracting review text features from the review dataset, and extracting user behavior features and review behavior features from the user behavior data; defining relationship categories based on the review text and user behavior features, and constructing a multidimensional relationship enhancement graph (MREGC) based on multiple relationships; using a dynamic adaptive feature enhancement graph neural network to learn features on the MREGC relationship graph, obtaining the embedding representation of each node in the multidimensional relationship enhancement graph, and aggregating the embedding representations; inputting the aggregated features into a classifier to determine whether a review is a fake review, and if the review is classified as a fake review, outputting a fake review label.

[0006] The beneficial effects of this invention are:

[0007] This invention, by fusing residual networks and multi-relationship comment graphs, fully considers the multidimensional relationships and deep features between comments, enabling the model to more accurately identify fake comments and significantly improving the accuracy and robustness of fake comment detection. By introducing residual networks and combining them with graph neural networks and graph attention mechanisms, this invention can adaptively assign different weights to different nodes, effectively solving the problem of imbalanced fake comment samples and improving classification accuracy and the model's generalization ability. Furthermore, by introducing the residual network structure, this invention makes deep networks easier to train and accelerates the convergence process, thereby significantly improving the model's training efficiency and optimization performance. Attached Figure Description

[0008] Figure 1 This is a flowchart of the big data-based fake review detection method of the present invention;

[0009] Figure 2 This is a model structure diagram of the fake review detection method based on big data of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] This invention provides a method for detecting fake reviews based on big data, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0012] S1 acquires a comment dataset, which includes comment texts from multiple users and corresponding user behavior data, and performs data preprocessing.

[0013] S2 extracts comment text features from the comment dataset and user behavior features and comment behavior features from the user behavior data;

[0014] Further data processing steps include:

[0015] Step 1: Extract text features. Using NLP technology and TF-IDF embedding, extract keywords from the comment text and introduce a keyword weighting mechanism to calculate the text keyword weights and generate a vector representation of the comment.

[0016] Step 2: Extract user behavior features from user behavior data.

[0017] Step 3: Extract comment behavior pattern features and define the comment behavior anomaly index. Used to measure users Is the commenting behavior abnormal?

[0018] Step 4: Extracted comment text features, user behavior features, and comment behavior features are fused to obtain more expressive feature representations.

[0019] Furthermore, the extraction of user behavior features includes the following steps:

[0020] First, user activity features are extracted to reflect user engagement on the platform. User activity is modeled based on dimensions such as comment frequency and participation in rating.

[0021] Secondly, historical rating characteristics are extracted; high rating volatility and frequent extreme rating behaviors may indicate fake reviews. The instability of ratings is represented by calculating the standard deviation of user rating history.

[0022] Furthermore, the standard deviation of the user rating history is used to represent the instability of the ratings as follows:

[0023]

[0024] In the formula: yes In time The rating; User The mean of all ratings.

[0025] Furthermore, the commenting behavior is considered an anomaly index. The rules are as follows:

[0026]

[0027] In the formula: :user In time The emotional score; :user In the time window The average sentiment score within; This measures the fluctuations in the sentiment trend of user comments. Large fluctuations indicate that the user's sentiment is unstable and may be abnormal behavior. :user In the time window The frequency of comments within the app can be a factor; if comments are too frequent, they may be fake. : Standard deviation of the time interval between user comments, which measures the fluctuation of the comment interval. If the fluctuation is too small, it may be a behavior of mass commenting; α, β, γ: Hyperparameters, which control the importance of sentiment trend, comment frequency and comment interval respectively.

[0028] S3: Based on the text and user behavior features of comments, define multiple relationship categories and construct a multidimensional relationship enhancement graph (MREGC) based on multiple relationships;

[0029] Furthermore, these multiple relationship categories include: user-comment relationship, rating relationship, time period relationship, sentiment relationship, content similarity relationship, and social interaction relationship.

[0030] Furthermore, based on the textual and user behavior features of the comments, multiple relationship categories are defined, and a multidimensional relationship enhancement graph (MREGC) is constructed according to these multiple relationships, including the following steps:

[0031] Step 1: Employ a relationship weighting mechanism, assigning different weights to each type of relationship to highlight the influence of each relationship on the detection of fake comments.

[0032] Step 2: Combining temporal features and multi-relationship graph construction, further enhance the temporal dynamic capability of the comment graph. Introduce a temporal decay function and an emotional attitude weighting factor to dynamically adjust the relation weights, thereby making the edges of the graph structure more timely and emotionally relevant.

[0033] Step 3: In the graph convolutional layer, MREGC performs feature fusion through multi-dimensional embedding, effectively integrating information from different relationships.

[0034] Furthermore, the weight adjustment rules are as follows:

[0035]

[0036] In the formula: It is time Time relationship The weights; It is an initial weight based on the relation type; and These are comments and the current time's timestamp; It is the time decay factor; It is based on the weight of comment sentiment, reflecting the impact of the comment sentiment on the relationship.

[0037] Furthermore, the multi-dimensional embedding update rules are as follows:

[0038]

[0039] in: It is the first The node feature matrix of the layer; It is the first Layer weight matrix; It is an activation function.

[0040] This approach enables each comment node to be dynamically updated under the influence of different relationships, integrating various types of information and providing a rich graph structure.

[0041] S4: Based on the MREGC relationship graph, a Dynamic Adaptive Feature Augmentation Graph Neural Network (DAFE-GNN) is used for feature learning. This involves multi-level fusion of temporal weighting and residual connections to model the temporal information and multiple relationships in the graph, obtaining the embedding representation of each node, and then performing feature aggregation. This includes:

[0042] Step 1: Use Dynamic Adaptive Feature Augmentation Graph Neural Network (DAFE-GNN) to learn features from the relation graph.

[0043] Step 2: Perform multi-scale feature fusion (MSF). To more intelligently combine short-term local patterns (GRU) and long-term dependency information (Transformer), an adaptive gating fusion mechanism is designed to adjust the fusion ratio. It can dynamically change based on the features of the current input, rather than using fixed hyperparameters, and introduce a gating unit. The fusion weights are adjusted based on adaptation.

[0044] Step 3: Perform Graph Signal Filtering (GCN). GCN mainly relies on topological structure for feature propagation but does not explicitly consider frequency domain information. To enhance feature representation, Graph Fourier Transform (GFT) is introduced to enhance information representation in the frequency domain.

[0045] Step 4: Introduce an Adaptive Residual Network (AR-ResNet) to further uncover deeper patterns in features. Define a Composite Residual Block (ARB) that adds adaptive weights and dynamic channel attention to the traditional ResNet structure.

[0046] Step 5: Integrate dynamic neighborhood attention, temporal modeling, multi-scale feature fusion, graph signal filtering, and adaptive residual network to update the feature formula.

[0047] Furthermore, the aforementioned dynamic adaptive feature-enhanced graph neural network introduces adaptive attention weights, allowing the influence of neighboring nodes to be dynamically adjusted over time, based on relationship type, and sentiment information. The calculation rules are as follows:

[0048]

[0049] in: It is the first Layer node characteristics; It is a trainable transformation matrix; These are attention weights, calculated using dynamic domain attention. The formula for calculating dynamic attention is:

[0050]

[0051] in: It is an attention scoring function, defined as follows:

[0052]

[0053] Furthermore, the aforementioned Multi-Scale Feature Fusion (MSF) employs an adaptive gating mechanism to more intelligently combine Short-Term Local Patterns (GRU) and Long-Term Dependency Information (Transformer), thereby increasing the fusion ratio. It can dynamically change based on the features of the current input, rather than using fixed hyperparameters, and introduce a gating unit. The fusion weights are adjusted based on adaptation.

[0054] Furthermore, the formula for calculating the Local Short-Term Model (GRU) is as follows:

[0055]

[0056] The global long-range mode (Sparse Transformer) calculation formula is as follows:

[0057]

[0058] Furthermore, adaptive fusion weights are calculated using gating units. To control the contributions of local and global modes, the gating unit... The calculation formula is:

[0059]

[0060] in: It is a trainable parameter matrix. It is the input feature of the current time step. It is a bias term. It is a sigmoid activation function, ensuring that in the gated unit Between (0,1).

[0061] In this embodiment, to enhance the information flow, weighted residual fusion is performed, which allows the fused representation to retain an additional portion of the original input. The calculation formula is as follows:

[0062]

[0063] Where: β is a learnable parameter (initialized to 0) that controls the proportion of residuals in the input information.

[0064] Furthermore, the calculation rules for the graph signal filtering (GFT) operation are as follows:

[0065]

[0066] Where: U is the normalized Laplace matrix We obtain it by eigenvalue decomposition, i.e. ; High-order information representing frequency characteristics, and These are trainable weights and biases; It is an adaptive frequency filter that dynamically adjusts the weights of different frequency components; It is a nonlinear transformation (such as ReLU or GELU) that enhances the ability to represent features; Perform a graphical Fourier transform to project the features into the frequency domain; Returning to the time domain via inverse transform (iGFT).

[0067] This method can realize the entire process of GFT transform → frequency domain filtering → inverse iGFT transform → nonlinear enhancement, and can simultaneously consider local details and global patterns, avoid over-smoothing, and enhance local dynamics.

[0068] Furthermore, the computation rules for the Adaptive Residual Network (AR-ResNet) are as follows:

[0069] Define the input feature as X, and the feature extraction of each layer follows the calculation as follows:

[0070]

[0071] in: Two-dimensional convolutional layers extract spatial features; Batch normalization is used to achieve stable training. ReLU or GELU activation functions enhance nonlinear expressive power. Trainable transformation matrix; Sigmoid activation ensures that the adaptive residual is between (0,1); Dynamic channel attention:

[0072]

[0073] Furthermore, the final feature update formula is as follows:

[0074]

[0075] S5: The embedded features obtained from DAFE-GNN are used as input and passed to the MLP classifier to determine whether the comment is a fake comment. If the comment is classified as a fake comment, a fake comment label is output.

[0076] Furthermore, the classifier is a multilayer perceptron (MLP), and the support vector machine performs binary classification by calculating the feature vectors of the comment nodes to determine whether the comment is a fake comment.

[0077] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting fake reviews based on big data, characterized in that, include: Obtain the comment dataset, which includes comment texts from multiple users and corresponding user behavior data, and perform data preprocessing. Extract comment text features from the comment dataset, and extract user behavior features and comment behavior features from the user behavior data; Relationship categories are defined based on textual and user behavior features of comments. A multidimensional relationship enhancement graph (MREGC) is constructed based on multiple relationships. On the basis of the MREGC relationship graph, a dynamic adaptive feature enhancement graph neural network is used for feature learning to obtain the embedding representation of each node in the multidimensional relationship enhancement graph. The embedding representations are then used for feature aggregation. The aggregated features are input into a classifier to determine whether a comment is a fake comment. If a comment is classified as a fake comment, a fake comment label is output.

2. The method for detecting fake reviews based on big data according to claim 1, characterized in that, Extracting features from comment text involves using NLP techniques to extract keywords from the comment text using TF-IDF embedding, and introducing a keyword weighting mechanism to calculate the text keyword weights and generate a vector representation of the comment. Extracting user behavior features and comment behavior features includes: extracting user activity features; extracting rating history features, representing rating instability by calculating the standard deviation of user rating history; and extracting comment behavior pattern features, defining a comment behavior anomaly index. Used to measure users Is the commenting behavior abnormal? 3. The method for detecting fake reviews based on big data according to claim 2, characterized in that, The standard deviation of user rating history is calculated as follows: ； in, yes In time The rating; User The mean of all ratings.

4. The method for detecting fake reviews based on big data according to claim 2, characterized in that, The abnormal comment behavior index is: ； in, For users In time The emotional score; For users In the time window The average sentiment score within; To measure the fluctuations in user sentiment trends in comments; For users In the time window Comment frequency within; denoted as the standard deviation of the time interval between user comments; α, β, and γ are hyperparameters that control the importance of sentiment trend, comment frequency, and comment interval, respectively.

5. The method for detecting fake reviews based on big data according to claim 1, characterized in that, The construction of the Multidimensional Relationship Enhancement Graph (MREGC) includes: assigning different weights to each type of relationship using a relationship weighting mechanism; constructing a multi-relationship graph based on weights and temporal features, and introducing a temporal decay function and an emotional attitude weighting factor to dynamically adjust the relationship weights; and weighting each type of relationship in the graph convolutional layer and performing feature fusion through multi-dimensional embedding.

6. The method for detecting fake reviews based on big data according to claim 5, characterized in that, Adjusting relation weights includes: ； in, It is time Time relationship The weights; It is an initial weight based on the relation type; and These are comments and the current time's timestamp; It is the time decay factor; It is based on the weight of the sentiment in the comments.

7. The method for detecting fake reviews based on big data according to claim 1, characterized in that, Feature learning using a dynamically adaptive feature-enhanced graph neural network includes: introducing adaptive attention weights to dynamically adjust the influence of neighboring nodes over time, based on relationship type, and sentiment information; extracting features from the adjusted information and performing multi-scale feature fusion; performing graph signal filtering on the fused feature map; inputting the filtered feature map into an adaptive residual network to obtain the deep pattern of the features; and updating the features by integrating dynamic neighborhood attention, temporal modeling, multi-scale feature fusion, graph signal filtering, and the adaptive residual network.

8. The method for detecting fake reviews based on big data according to claim 7, characterized in that, Adaptive residual networks process input features in the following ways: ； in, It is a two-dimensional convolutional layer used to extract spatial features; For batch normalization layer; Use ReLU or GELU activation functions; All are trainable transformation matrices; Activated for Sigmoid; For dynamic channel attention.

9. The method for detecting fake reviews based on big data according to claim 1, characterized in that, The classifier is a multilayer perceptron, and the support vector machine performs binary classification by calculating the feature vectors of the comment nodes to determine whether the comment is fake.

Images