Multi-field marketing topic robot detection method based on feature fusion

Abstract

The invention relates to a multi-field marketing topic robot detection method based on feature fusion, and belongs to the technical field of social network data analysis and machine learning model detection. Constructing weighted metadata feature representation, vectorizing metadata such as a user introduction, a concerned number and a concerned number through an embedded layer, and introducing an attention mechanism to calculate an attribute weight through multi-layer perceptron nonlinear transformation to obtain weighted metadata features; performing domain division on the text topics based on a potential Dirichlet allocation model, constructing a multi-domain relation network diagram by taking likes, collection, forwarding and comments of users in different domain topics as edges, and fusing weighted metadata features into the diagram; extracting multi-domain social relation features in the domain relation network by using fine granularity of a graph neural network; a long-short-term memory network model based on multiple fields is constructed, user multi-time slice behavior features are extracted to form a time sequence feature matrix, and the marketing topic robot is effectively detected and recognized in combination with social relation features.

Description

Technical Field

[0001] This invention belongs to the field of social network data analysis and machine learning model detection technology, and relates to a method for detecting multi-domain marketing topic robots based on feature fusion. Background Technology

[0002] Marketing bots primarily refer to fake accounts that post marketing messages on social networks. With the rapid development of online social networks, various social networking platforms not only provide a bridge for communication between people but also serve as an important medium for information dissemination. However, the openness and widespread nature of social networks have also made them a breeding ground for malicious actors to exploit bots. Some businesses use marketing bot accounts to post fake positive reviews and five-star ratings, and to smear competitors, leading to a loss of trust on social networking platforms and chaos in market order. Therefore, timely and accurate detection and prevention of these marketing bots has become an urgent problem for researchers and social network platform operators.

[0003] In recent years, experts and scholars have conducted a series of studies on social bot detection. Based on different research focuses, these research methods can be broadly divided into two categories: one is to establish a bot detection model based on a deep neural network model, thereby detecting bots in social networks through improvements to the deep neural network model; the other is to focus on developing more effective methods for extracting social bot features, improving the effectiveness of feature extraction.

[0004] In studying the behavior of marketing bots, some unscrupulous companies or marketers have used them to exaggerate their products on social networks, misleading consumers. For example, a luxury music festival held in the Bahamas in 2017 used a large number of social media influencers and bots to attract attention during its promotion, but the event collapsed due to poor execution and deceptive advertising, and a large amount of negative news and angry reactions on social media severely damaged the brand's reputation.

[0005] In summary, to improve the accuracy of marketing robot detection, it is necessary to combine multiple user characteristics and comprehensively consider their differences across multiple dimensions. Analysis reveals the following problems with existing marketing robot detection methods: 1. The complexity of measuring the contribution of user metadata attributes: User metadata in social networks contains valuable information, but how to effectively assign contribution to these attributes to accurately distinguish between normal users and bot users is a problem that existing technologies cannot solve.

[0006] 2. Diversity of social relationships in different social fields: Users' social relationships vary significantly across different social fields. How to accurately capture and utilize these differences to improve the accuracy and efficiency of robot detection is a major challenge we currently face.

[0007] 3. The dynamic and diverse nature of marketing robot behavior: The behavioral characteristics of marketing robots change over time, and their strategies are diverse and complex, which increases the difficulty of detection.

[0008] In recent years, scholars at home and abroad have studied the behavioral patterns of robots in social networks from multiple perspectives in response to the above problems, and have achieved fruitful research results, providing a theoretical basis for robot detection.

[0009] To address the issue of measuring the contribution of user metadata attributes, existing technologies propose a framework using minimal account metadata, strategically selecting a subset of training data to improve model accuracy and generalization. Other technologies propose a detection model that integrates metadata and multiple user information modalities to combat social bot spoofing. Some technologies employ a novel detection approach, extracting user profiles from topic texts and using the similarity of user profiles between two topics as a detection advantage. Still others use metadata as input to a Long Short-Term Memory (LSTM) network to capture the semantic information of user profiles. While these research papers focus on incorporating user metadata as a whole into the detection model, it's also worth considering the varying importance of each metadata attribute.

[0010] To address the diversity of social relationships across different social domains, existing technologies propose a heterogeneous graph approach to integrate various users and activities in social networks. By minimizing high-dimensional structural entropy, a hierarchical community structure of the social network is generated. Other technologies propose a domain-aware detection method based on a multi-relationship graph neural network (GNN), which constructs a multi-relationship graph using user features and relationships, and obtains user representation through graph embedding. Still other technologies utilize non-negative matrix factorization community detection algorithms to partition the social graph, control intra- and inter-community walks by designing community walk rules, and learn user node embeddings using graph representation learning. Finally, some technologies propose inferring social relationships from heterogeneous user movement graphs. While these studies can provide vector representations of social relationships, further research is needed to simplify the differentiation of social domains.

[0011] To address the dynamic and diverse nature of user behavior, existing technologies model personal data, time, and activity information as sequences; they design individual social behavior profiles by combining users' individual behavioral characteristic indicators; they calculate attribute values ​​for each user by acquiring user comments and network structure information, and then concatenate these values ​​as user behavioral features; and they utilize time correlation functions from statistical physics and evolutionary methods from machine learning to discover behavioral characteristics in social networks. While these studies have shown significant effectiveness in addressing the diversity of user behavior, further research is needed to address its dynamic aspects. Summary of the Invention

[0012] This invention aims to address the problem of identifying marketing bots within the marketing topic network space. Therefore, the purpose of this invention is to leverage the differences in user metadata, social relationships, and behavioral characteristics to analyze user metadata information and relationship networks, and integrate temporal behavioral characteristics to propose a multi-domain marketing topic bot detection model based on feature fusion.

[0013] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting multi-domain marketing topic bots based on feature fusion includes the following steps: S1. Constructing a weighted metadata feature representation: To address the problem of measuring the contribution of attributes in user metadata, embedding representation learning is used to construct a user metadata feature vector. A multilayer perceptron-attention mechanism (MLP-Attention) model is constructed to quantify the contribution of metadata attributes through the attention mechanism, and the user metadata is represented by a weighted metadata feature representation based on the contribution. S2. Constructing a multi-domain social relationship feature representation: In view of the differences in the social relationship features of users in different domains, the text topic domain is analyzed based on the Latent Dirichlet Allocation (LDA) model. The likes, collections, forwards and comments of users on topics in different domains are used as edges in the corresponding domain relationship network graph to construct multiple domain relationship network graphs, and the weighted metadata features are integrated into the relationship network. S3. Extracting multi-domain social relationship features: Using a graph neural network (GNN) in the domain relationship network graph, fine-grained features of the user's social relationships in various domains are extracted to construct a multi-domain feature representation of the user; S4. Detect Marketing Topic Bots: Construct a multi-domain Long Short-Term Memory (BMD-LSTM) network model, extract user behavior features in multiple time slices and construct a temporal feature matrix. By capturing the feature change patterns in each time slice and utilizing the social relationship features, detect marketing bots in the marketing topic network.

[0014] Furthermore, S1 also includes the following steps: The user's various metadata attributes are vectorized. Discrete metadata attributes such as profile, number of followers and number of followers are mapped to a continuous low-dimensional vector space through an embedding layer. The corresponding embedding vector is obtained by looking up the embedding matrix, and the embedding vectors are concatenated to obtain the metadata feature vector. The metadata feature vector is input into a multilayer perceptron and subjected to a nonlinear transformation to obtain new features; An attention mechanism is introduced to calculate an attention score for each embedding vector; Based on the attention score, the user metadata features are weighted to obtain a weighted feature vector, which is the weighted metadata feature representation.

[0015] Furthermore, the extraction of social relationship features in S2 includes constructing a multi-domain relationship network, specifically comprising the following steps: The text topic content is preprocessed, and the preprocessed text is converted into a document-word matrix using text vectorization methods, so that each row of the matrix represents a document, each column represents a word, and the matrix elements represent the frequency of word occurrence in the document; By setting hyperparameters in the latent Dirichlet distribution model and training the document-word matrix, the document-topic distribution is obtained. The domain to which a topic belongs is determined based on the topic with the highest probability in the document-topic distribution. In each topic domain, users are treated as nodes, and the likes, favorites, reposts, and comments between users on that topic are treated as edges. The number of such actions is used as the weight of the edges to construct a domain relationship network graph. The weighted metadata features are then input into the relationship network graph as node features.

[0016] Furthermore, step S3 involves extracting the user's social relationship features in the relationship network graph using the graph neural network, specifically including the following steps: Linear transformation and activation function are used to perform preliminary transformation on the features of each node; the features of the node are updated by aggregating the features of neighboring nodes, and the aggregated information is combined with the node's own features; multi-layer graph convolution operation is used to capture the deeper dependencies and feature representations of nodes in the relationship graph, so that the node features in the relationship graph include information from direct neighbors and more distant nodes.

[0017] Furthermore, the combination of the aggregated information with the node's own features is expressed as follows:

[0018]

[0019]

[0020] in, Is the user in the The characteristic state of the layer, It is a non-linear activation function. It is the first The weight matrix of the layer, It's a bias. Represents a node The set of neighboring nodes, It is the normalization coefficient. It is a node Neighborhood context features after aggregating adjacent nodes.

[0021] Furthermore, S3 also includes the following steps: The system acquires user behavior data such as likes, favorites, reposts, and comments over a period of time in a social network, extracts temporal behavior feature vectors for users across multiple time slices, and constructs a temporal feature matrix. It also extracts user social relationship feature vectors, which are obtained by aggregating information from user nodes and their neighboring nodes. The aggregation process includes preliminary transformation of the node's own features and aggregation of features from neighboring nodes. Finally, it fuses the temporal behavior feature vectors of each time slice with the corresponding social relationship feature vectors to obtain fused feature vectors for each time slice, and combines them to form a fused feature matrix.

[0022] Furthermore, the fused feature matrix is ​​represented as:

[0023]

[0024] in, In time slice The fusion feature vector at that time It is a temporal behavior feature vector. It is a feature vector of social relationships. It is a temporal feature matrix composed of the fused feature vectors of all time slices.

[0025] Furthermore, S4 also includes the following steps: The fused feature matrix is ​​used as input to the long short-term memory network and passed step by step over time. At each time step, the feature matrix first passes through a forget gate to determine the information to be retained or forgotten, then through an input gate to determine the amount of new information to be written and to generate candidate cell states. The results are combined with the cell states of the previous step to update the current cell state. The output gate generates hidden states based on the current cell state. Each layer of the fused feature matrix model dynamically adjusts the flow of information through a gating mechanism to ensure that effective information is transmitted and new hidden states and outputs are generated. Finally, the last hidden state is output at the output gate.

[0026] Furthermore, the last hidden state , represented as:

[0027]

[0028] in, It is the output of the output gate. It is the weight matrix of the output gate. It's a bias. It is the hidden state at the current moment, combined with the current cell state. Generate a new hidden state, and use element-wise multiplication to output the gate. Activation value relative to the current cell state Multiply element by element.

[0029] Furthermore, S4 also includes the following steps: The last hidden state of the Long Short-Term Memory (LSTM) network is input into a fully connected layer, and the Softmax function is used to predict the class probability of the LSM network. When the model output is at its maximum value, the current user is determined to be a marketing bot; when the model output is at its minimum value, the current user is determined to be a real user.

[0030] The beneficial effects of this invention are as follows: It proposes an MLP-Attention model to address the problem of measuring the contribution of user metadata attributes. This model utilizes an attention mechanism to assign contribution values ​​to various attributes in user metadata, and enhances the model's accuracy in recognizing user metadata features through a weighted feature representation method. It uses LDA to distinguish user behaviors across different domains and GNN to extract multi-domain social relationship features. By extracting user relationship features in different social domains through graph neural networks, the model can more accurately capture user behavioral patterns in a multi-dimensional social space, thereby improving detection performance.

[0031] A BMD-LSTM model is proposed to address the dynamic nature of marketing bot behavior. By combining it with a Long Short-Term Memory (LSTM) network to capture and analyze the temporal changes in user behavior characteristics, this model can more comprehensively understand and identify user behavior patterns at different time points, improving the accuracy and reliability of marketing bot detection. This model can not only effectively identify marketing bots but also differentiate between marketing bots from different domains, thereby improving detection accuracy.

[0032] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of a multi-domain marketing topic robot detection model based on feature fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the MLP-Attention model according to an embodiment of the present invention. Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0035] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0036] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0037] With the development of online social networks, timely detection of marketing bots on the internet can effectively prevent the malicious manipulation of online information. Considering the diversity of marketing bots, this invention proposes a multi-domain marketing topic bot detection model based on feature fusion.

[0038] First, to address the issue of measuring the contribution of attributes in user metadata, and considering the advantages of attention mechanisms in capturing key attributes, a Multilayer Perceptron-Attention (MLP-Attention) model is constructed. This model introduces embedding representation learning to construct feature vectors for user metadata, while simultaneously integrating the attention mechanism to quantify the contribution of metadata attributes and perform weighted feature representation of user metadata based on attribute contributions.

[0039] Secondly, considering the differences in social relationship characteristics among users in different domains, and taking into account the advantages of Latent Dirichlet Allocation (LDA) in analyzing text topic domains, we use users' likes, favorites, reposts, and comments on topics in different domains as edges of nodes in the relationship graph of that domain, thereby constructing relationship network graphs for different domains. At the same time, we integrate weighted metadata features into the relationship network and use Graph Neural Network (GNN) to extract fine-grained features of users' social relationships in various domains, thus constructing feature representations for multiple domains.

[0040] Finally, considering the advantages of Long Short-Term Memory (LSTM) networks in processing dynamic data, a Multi-Domain LSTM (BMD-LSTM) model is constructed. This model first extracts user behavioral features in each time slice and constructs a temporal feature matrix. By capturing the patterns of feature changes in each time slice and comprehensively considering the influence of social relationships, marketing robots can be detected from multiple domains.

[0041] In the process of building the marketing robot detection model, the following features are first defined: 1. Weighted metadata features

[0042] The fundamental attributes in user metadata have varying degrees of importance in distinguishing between bot users and real users, and studying the importance of these fundamental attributes is valuable for detecting marketing bots. Therefore, the weighted metadata features of users are expressed by the following formula: (1) (2) in, Indicates the current user. Indicates the first Weighted feature representation of each basic attribute, Indicates the first Feature representation of each basic attribute Indicates the first The weights of the basic attributes.

[0043] 2. Relationship Network Diagram

[0044] A relationship network diagram illustrates the interactions between users in a specific domain, and it is represented by the following formula: (3) in, Indicates the first A relational network diagram for each domain. This represents the set of users in the relationship graph. This represents the set of edges between users in the graph. This represents the set of edge weights in the graph.

[0045] 3. Topic Areas

[0046] Describing the social sphere to which a marketing topic belongs, studying the sphere of a marketing topic can effectively distinguish users' social spheres, and it can be expressed by the following formula: (4) in, This refers to a document containing word vectors generated based on this marketing topic. It is a document Chinese theme The weight, Indicates to make The biggest theme .

[0047] 4. Social Relationship Characteristics

[0048] This is used to describe a user's social relationships on social networks. Because marketing bots often exhibit low centrality and low clustering coefficients on social networks, the social relationship feature is represented by the following formula: (5) in, Represents a node In GNN Layer feature representation, It is an activation function. It is a node Neighborhood Day point, Represents a node The set of neighboring nodes, It is a normalization function. and It is GNN's Layer weights and biases It is GNN's Layer bias terms.

[0049] 5. Characteristics of periodic behavior

[0050] The formula used to describe the patterns of change in user behavior characteristics is as follows: (6) in, Indicates that the user is Behavioral feature matrix in each time slice Indicates the user's time slice The behavioral feature vector in.

[0051] The symbols and feature summaries for the above features are shown in Table 1: Table 1

[0052] This invention aims to identify marketing bots in the marketing topic domain by studying multi-dimensional differences in user characteristics. It is based on the metadata characteristics of users in social networks. Weighted metadata features are obtained by fusing attention mechanisms using a multilayer perceptron. Combined with domain relationship networks Extract users Social relationship characteristics within the domain Then, integrate the user's periodic behavioral characteristics. The user's behavioral tendencies are obtained using a Long Short-Term Memory (LSTM) network model. The problem can be expressed as the following formula: (7) Based on the above definitions and descriptions, the input to the model of this invention is determined as follows: 1. Metadata Feature Matrix: ,in This indicates the first [item] in the corresponding user metadata. Feature vectors of each attribute.

[0053] 2. Topic Area: ,in This represents a document containing word vectors generated based on the corresponding marketing topic. Indicates to make The biggest theme .

[0054] 3. Relationship network diagram: ,in This represents the set of users in the relationship graph. This represents the set of edges between users in the graph. This represents the set of edge weights in the graph.

[0055] 4. Periodic behavioral characteristics: ,in Indicates the user's time slice The behavioral feature vector in.

[0056] Based on the above inputs, the output of the model of this invention is: 1. Weighted metadata features By constructing a metadata feature matrix and using a multilayer perceptron to fuse attention mechanisms, the importance of metadata attributes is obtained, and the metadata feature matrix is ​​weighted according to the importance of the attributes.

[0057] 2. Characteristics of social relationships By detecting marketing topics, we construct user social relationship networks across different domains by using users' likes, favorites, reposts, and comments on topics in different domains as edges. We then use weighted metadata features as attribute features of network nodes and extract users' social relationship features in the relationship network graph using a graph neural network model.

[0058] 3. Marketing robot testing To accurately detect marketing bots, a multi-domain marketing bot detection model, BMD-LSTM, based on feature fusion, is proposed. It integrates social relationship features and temporal behavior features, and utilizes the differences in multi-dimensional features to detect marketing bots in marketing topic networks.

[0059] Please see Figure 1 This is a schematic diagram of a multi-domain marketing topic robot detection model based on feature fusion, according to an embodiment of the present invention.

[0060] First, an attention mechanism is introduced to obtain the importance of metadata attributes, and then weighted metadata features are extracted. Second, LDA is used to identify marketing topic domains, thereby constructing a relationship network graph of different domains. GNN is used to extract the social relationship features of users in different domains in the relationship network graph. Finally, the temporal feature representation of user behavior is extracted and combined with social relationship features to propose a feature fusion-based detection model BMD-LSTM, thereby detecting marketing topic bots in the marketing topic network.

[0061] 1. Topic User Weighted Metadata Feature Representation User metadata attributes in social networks exhibit high heterogeneity, and the degree to which users value attributes determines the varying contributions of each attribute to distinguishing user metadata. Therefore, directly using user metadata may overlook the influence of some key factors, affecting the accuracy of model detection. Weighted metadata feature representation, however, can better quantify and utilize these heterogeneous attributes. (See [link to relevant documentation]). Figure 2 This is a schematic diagram of the MLP-Attention model in an embodiment of the present invention.

[0062] First, it is necessary to vectorize various metadata attributes of users, including profile information, number of followers, and number of followed users. To better represent this metadata, this invention maps discrete metadata attributes to a continuous low-dimensional vector space through an embedding layer, by looking up the embedding matrix. Obtain the corresponding embedding vector The embedded vectors are concatenated to obtain the feature vector of the metadata. The feature is represented by the following formula: (8) in, Indicates the first in the metadata A vector representation of each attribute.

[0063] To further capture the non-linear relationships of these metadata features, a multilayer perceptron is used to process the feature vectors. Perform a nonlinear transformation to obtain new features. It can be expressed as the following formula: (9) in, yes Feature representation after transformation by two layers of neural network and The first layer consists of the weight matrix and the bias vector. and These are the weight matrix and bias vector of the second layer, respectively. It is a non-linear activation function .

[0064] To quantify the importance of each metadata attribute, an attention mechanism is introduced, followed by an embedding vector for each attribute. Its attention score is calculated through an attention mechanism. It can be expressed as the following formula: (10) (11) in, It is a scoring function. It is a learnable parameter matrix. This is achieved by analyzing the attention score. After normalization, the attention weights for each attribute are obtained. .

[0065] Based on attention weight The user's metadata feature vector is weighted to obtain the weighted feature representation, which is expressed by the following formula: (12) in, It is a weighted feature vector that combines the contributions of all metadata attributes.

[0066] 2. Constructing multi-domain social relationship representations In social relationship networks, marketing bots differ significantly from real users in network topology, often exhibiting characteristics such as uniform degree distribution and low clustering coefficients. Therefore, research on user characteristics within social relationship networks is extremely important. This invention proposes a method for extracting social relationship features, which consists of two stages: the first stage involves classifying the domains of marketing topics based on the LDA model and constructing a multi-domain relationship network; the second stage involves extracting fine-grained representations of social relationship features based on graph neural networks.

[0067] 1) Constructing a multi-domain relationship network Considering that topics serve as a medium for interconnection among users in social networks, identifying the domain to which a topic belongs can effectively segment users' social domains. First, the text topic content is preprocessed, and then text vectorization methods are used to transform the preprocessed text into a document-word matrix. X It can be expressed as the following formula: (13) In this matrix, each row represents a document, each column represents a word, and the matrix elements... Document Chinese words Frequency of occurrence.

[0068] Setting the hyperparameters of the Dirichlet distribution in the LDA model The document-word matrix is ​​trained using a model to obtain the document-topic distribution. It can be expressed as the following formula: (14) in, This represents the document-topic distribution, that is, within the document... Chinese topic The frequency of occurrence Document Chinese topic Number of times it appears It is a document The topic was assigned to Examples of words.

[0069] Based on document-topic distribution Identify the main topic areas for each document. The topic with the highest probability is represented by the following formula: (15) Through the above process, the present invention can classify users' text topics into different topic areas.

[0070] This invention will construct an independent relationship graph for each topic domain. Each relationship graph consists of a finite non-empty set of nodes. The set of edges between vertices and the set of edge weights The graph is composed of several parts. Users are treated as nodes in a relationship graph, and the actions users take to like, favorite, share, and comment on a topic are treated as edges. The number of likes, favorites, shares, and comments is used as the weight of each edge. This can be expressed as the following formula: (16) in, Indicates the first A relational network diagram for each domain.

[0071] Furthermore, the extracted weighted metadata features are used as node features in the relation graph, providing rich feature representations for each node. It can be expressed as the following formula: (17) in, It is the first in the user metadata The weighted feature values ​​of each attribute. It refers to the dimension and shape of the feature vector.

[0072] 2) Extracting social relationship features Graph neural networks can capture dependencies between nodes and extract more accurate social relationship features by aggregating information from user nodes and their neighboring nodes. This invention performs an initial transformation on the features of each node using linear transformations and activation functions, then updates the node's features by aggregating the features of neighboring nodes, and finally combines the aggregated information with the node's own features, expressed as the following formula: (18) (19) (20) in, Is the user in the The characteristic state of the layer, It is a non-linear activation function. It is the first The weight matrix of the layer, It's a bias. Represents a node The set of neighboring nodes, These are the normalization coefficients, whose values ​​are calculated using the max-min normalization formula. It is a node Neighborhood context features after aggregating adjacent nodes.

[0073] To capture deeper dependencies and feature representations of nodes in a relational graph, multi-layer graph convolution operations are used. This allows the node features in the relational graph to include not only information from direct neighbors but also information from more distant nodes, as expressed in the following formula: (twenty one) (twenty two) in, It is the adjacency matrix after adding self-loops. yes The degree matrix. Through multi-layer graph convolution operations, the feature vectors of a node gradually fuse information from its neighboring nodes, ultimately encoding the node's position and relationship information within the entire social network.

[0074] 3. Marketing robot testing Marketing bots on social networks are influenced by manipulators and exhibit stronger regularity in their behavior patterns compared to real users, providing a theoretical basis for identifying marketing bots. This invention considers the historical behavioral patterns of users as a key factor in identifying marketing bots. It constructs a time-series feature matrix by extracting user behavior data such as likes, favorites, shares, and comments over a period of time. This can be expressed as the following formula: (twenty three) in, Is the user in the Behavioral feature vectors in each time slice.

[0075] Considering the significant differences between real users and marketing bots across multiple dimensions, single-dimensional detection is susceptible to outliers, affecting the model's robustness. Therefore, this invention fuses extracted temporal behavioral features with social relationship features to obtain a fused feature matrix, expressed as follows: (twenty four) (25) in, In time slice The fusion feature vector at that time It is a temporal behavior feature vector. It is a feature vector of social relationships. It is a temporal feature matrix composed of the fused feature vectors of all time slices.

[0076] Subsequently, a Long Short-Term Memory (LSTM) network is used to capture the dynamic changes in the data from the temporal perspective of the fused features. The fused feature matrix serves as the model input, progressively passed across time steps. At each time step, the feature matrix first passes through a forgetting gate, determining which information needs to be retained or forgotten. Next, the input gate determines the amount of new information to be written, while candidate cell states generate new potential information content. These results are combined with the cell states from the previous step to update the current cell state. Finally, the output gate generates hidden states based on the current cell state, determining the final output. Each layer of the model dynamically adjusts the flow of information through a gating mechanism, ensuring that effective information is passed and new hidden states and outputs are generated. The model ultimately outputs the last hidden state at the output gate. It can be expressed as the following formula: (26) (27) in, It is the output of the output gate. It is the weight matrix of the output gate, which contains the weights of the input vector and the previous hidden state. It's a bias. It is the hidden state at the current moment, combined with the current cell state. Generate a new hidden state, and use element-wise multiplication to output the gate. Activation value relative to the current cell state Multiply element by element.

[0077] The last hidden state of the model The input is passed to a fully connected layer. The function predicts the probability value of the Long Short-Term Memory network to detect whether it is a marketing bot. The model output can be expressed as the following formula: (28) in, This represents the class probability output by the model. It is the weight matrix of the fully connected layer. It's a bias. It is the model at the last time step The output hidden state vector.

[0078] Since the marketing topic robot detection problem studied in this invention is a binary classification problem, let the model output be represented as: Specifically, it is defined as follows: (29) in, This indicates that the current user has been identified as a marketing bot, and This indicates that the current user has been identified as a real user. It is a probability label for the category "robot". It is a probability label for the category of "real users".

[0079] 4. Model Algorithm In summary, this paper proposes a BMD-LSTM model to address the problem of marketing bot detection. This model utilizes metadata features and combines them with an attention mechanism to construct a weighted metadata feature representation. Then, it leverages topic domains to build a multi-domain user relationship network, fusing the weighted metadata features to extract social relationship feature representations. Finally, it integrates users' temporal behavior features with social relationship features to detect marketing bots. The experiments consist of model training and model testing. The specific model algorithm is shown in Table 2. Table 2

[0080] The algorithm complexity analysis of this model is as follows: The time complexity of using a multilayer perceptron to fuse attention mechanisms for weighted representation of metadata is expressed as follows: The time complexity of constructing a multi-domain social relationship network graph is: The time complexity of using a graph neural network to extract social relationship features is: The time complexity of using a Long Short-Term Memory (LSTM) network to detect patterns in user behavior is: The time complexity of the model algorithm is: .

[0081] In summary, this invention proposes a multi-domain marketing topic bot detection model based on feature fusion. First, it uses a multilayer perceptron model with an attention mechanism to perform weighted feature representation of metadata. Second, it proposes using LDA to segment the user's social relationship network into different domains, integrating the weighted representation of metadata into the relationship graph, and using GNN to extract the user's social feature representation in the relationship graph. Finally, it extracts the user's behavioral features and performs temporal representation, integrating social features into the temporal feature representation, and employs a Long Short-Term Memory network to detect bots among users. The effectiveness of the model is verified using a real dataset. The beneficial technical effects of this invention are as follows: 1. An MLP-Attention model is proposed to address the problem of measuring the contribution of user metadata attributes. This model utilizes an attention mechanism to assign contribution values ​​to each attribute in the user metadata, and enhances the model's accuracy in recognizing user metadata features through a weighted feature representation method.

[0082] 2. Use LDA to distinguish user behaviors across different domains, and use GNN to extract multi-domain social relationship features. By extracting user relationship features in different social domains through graph neural networks, the model can more accurately capture user behavioral patterns in multi-dimensional social spaces, thereby improving detection performance.

[0083] 3. A BMD-LSTM model is proposed to address the dynamic nature of marketing robot behavior. By combining a long short-term memory network to capture and analyze the temporal changes in user behavior characteristics, a more comprehensive understanding and identification of user behavior patterns at different time points can be achieved, improving the accuracy and reliability of marketing robot detection.

[0084] This model can not only effectively identify marketing topic bots, but also differentiate between marketing topic bots in different fields, thereby improving the accuracy of detection.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting multi-domain marketing topic robots based on feature fusion, characterized in that, Includes the following steps: S1. Constructing a weighted metadata feature representation: To address the problem of measuring the contribution of attributes in user metadata, embedding representation learning is used to construct a user metadata feature vector. A multilayer perceptron-attention mechanism (MLP-Attention) model is constructed to quantify the contribution of metadata attributes through the attention mechanism, and the user metadata is represented by a weighted metadata feature representation based on the contribution. S2. Constructing a multi-domain social relationship feature representation: In view of the differences in the social relationship features of users in different domains, the text topic domain is analyzed based on the Latent Dirichlet Allocation (LDA) model. The likes, collections, forwards and comments of users on topics in different domains are used as edges in the corresponding domain relationship network graph to construct multiple domain relationship network graphs, and the weighted metadata features are integrated into the relationship network. S3. Extracting multi-domain social relationship features: Using a graph neural network (GNN) in the domain relationship network graph, fine-grained features of the user's social relationships in various domains are extracted to construct a multi-domain feature representation of the user; S4. Detect marketing topic bots: Construct a multi-domain long short-term memory network (BMD-LSTM) model, extract user behavior features in multiple time slices and construct a temporal feature matrix. By capturing the feature change patterns in each time slice and utilizing the social relationship features, detect marketing bots in the marketing topic network.

2. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 1, characterized in that, S1 further includes the following steps: The user's various metadata attributes are vectorized. Discrete metadata attributes such as profile, number of followers and number of followers are mapped to a continuous low-dimensional vector space through an embedding layer. The corresponding embedding vector is obtained by looking up the embedding matrix, and the embedding vectors are concatenated to obtain the metadata feature vector. The metadata feature vector is input into a multilayer perceptron and subjected to a nonlinear transformation to obtain new features; An attention mechanism is introduced to calculate an attention score for each embedding vector; Based on the attention score, the user metadata features are weighted to obtain a weighted feature vector, which is the weighted metadata feature representation.

3. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 1, characterized in that, The extraction of social relationship features in S2 includes constructing a multi-domain relationship network, specifically comprising the following steps: The text topic content is preprocessed, and the preprocessed text is converted into a document-word matrix using text vectorization methods, so that each row of the matrix represents a document, each column represents a word, and the matrix elements represent the frequency of word occurrence in the document; By setting hyperparameters in the latent Dirichlet distribution model and training the document-word matrix, the document-topic distribution is obtained. The domain to which a topic belongs is determined based on the topic with the highest probability in the document-topic distribution. In each topic domain, users are treated as nodes, and the likes, favorites, reposts, and comments between users on that topic are treated as edges. The number of such actions is used as the weight of the edges to construct a domain relationship network graph. The weighted metadata features are then input into the relationship network graph as node features.

4. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 1, characterized in that, In step S3, the social relationship features of users in the relationship network graph are extracted through the graph neural network, specifically including the following steps: Linear transformation and activation function are used to perform preliminary transformation on the features of each node; the features of the node are updated by aggregating the features of neighboring nodes, and the aggregated information is combined with the node's own features; multi-layer graph convolution operation is used to capture the deeper dependencies and feature representations of nodes in the relationship graph, so that the node features in the relationship graph include information from direct neighbors and more distant nodes.

5. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 4, characterized in that, The combination of the aggregated information with the node's own features is represented as follows: in, Is the user in the The characteristic state of the layer, It is a non-linear activation function. It is the first The weight matrix of the layer, It's a bias. Represents a node The set of neighboring nodes, It is the normalization coefficient. It is a node Neighborhood context features after aggregating adjacent nodes.

6. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 1, characterized in that, S3 further includes the following steps: The system acquires user behavior data such as likes, favorites, reposts, and comments over a period of time in a social network, extracts temporal behavior feature vectors for users across multiple time slices, and constructs a temporal feature matrix. It also extracts user social relationship feature vectors, which are obtained by aggregating information from user nodes and their neighboring nodes. The aggregation process includes preliminary transformation of the node's own features and aggregation of features from neighboring nodes. Finally, it fuses the temporal behavior feature vectors of each time slice with the corresponding social relationship feature vectors to obtain fused feature vectors for each time slice, and combines them to form a fused feature matrix.

7. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 6, characterized in that, The fused feature matrix is ​​represented as: in, In time slice The fusion feature vector at that time It is a temporal behavior feature vector. It is a social relationship feature vector. It is a temporal feature matrix composed of the fused feature vectors of all time slices.

8. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 6, characterized in that, S4 further includes the following steps: The fused feature matrix is ​​used as input to the long short-term memory network and passed step by step over time. At each time step, the feature matrix first passes through a forget gate to determine the information to be retained or forgotten, then through an input gate to determine the amount of new information to be written and to generate candidate cell states. The results are combined with the cell states of the previous step to update the current cell state. The output gate generates hidden states based on the current cell state. Each layer of the fused feature matrix model dynamically adjusts the flow of information through a gating mechanism to ensure that effective information is transmitted and new hidden states and outputs are generated. Finally, the last hidden state is output at the output gate.

9. The method for detecting multi-domain marketing topic robots based on feature fusion according to claim 8, characterized in that, The last hidden state , is represented as: in, It is the output of the output gate. It is the weight matrix of the output gate. It's a bias. It is the hidden state at the current moment, combined with the current cell state. Generate a new hidden state, and use element-wise multiplication to output the gate. Activation value relative to the current cell state Multiply element by element.

10. A method for detecting multi-domain marketing topic robots based on feature fusion according to claim 8, characterized in that, S4 further includes the following steps: The last hidden state of the Long Short-Term Memory (LSTM) network is input into a fully connected layer, and the Softmax function is used to predict the class probability of the LSM network. When the model output is at its maximum value, the current user is determined to be a marketing bot; when the model output is at its minimum value, the current user is determined to be a real user.

Images