Advertisement risk control method and system based on multi-dimensional behavior characteristics

Abstract

The application relates to the technical field of advertisement risk control, and discloses an advertisement risk control method and system based on multi-dimensional behavior characteristics. In view of the problems that the feature fusion of the existing scheme is inefficient, and modeling is difficult to adapt to the time sequence and correlation characteristics of cheating behaviors, the method first collects multi-source behavior data of advertisement interaction, carries out normalization and missing value filling preprocessing to obtain a standardized data set; then extracts features from four dimensions of basis, time sequence, correlation and semantics to construct a multi-dimensional feature set; the features are dynamically weighted and fused through an attention mechanism, and time sequence graph neural network is inputted to complete risk modeling; and finally, a dynamic risk threshold is used to output a risk control decision. The application realizes accurate identification of complex advertisement cheating behaviors, improves the generalization ability and scene adaptability of risk control, reduces the waste of marketing fees of advertisers, and maintains the fairness and stability of the digital advertisement ecology.

Description

Technical Field

[0001] This invention relates to the field of advertising risk control technology, and in particular to an advertising risk control method and system based on multi-dimensional behavioral characteristics. Background Technology

[0002] With the rapid development of the digital advertising industry, advertising fraud, such as click fraud, impression fraud, malicious traffic diversion, and fake conversions, has become increasingly rampant. This not only wastes advertisers' marketing budgets but also undermines the fairness of the digital advertising ecosystem. Existing advertising risk control designs have key flaws in feature fusion and risk modeling, directly affecting the accuracy and generalization ability of fraud detection.

[0003] Existing solutions often employ fixed weighting, simple concatenation, or linear combination to process multi-dimensional behavioral features. They fail to consider the varying contributions of different features across different advertising scenarios and fraud types, and cannot dynamically adjust feature weights. This results in redundant information masking crucial, effective features for fraud identification, leading to inefficient feature fusion. Furthermore, most existing models focus solely on modeling single temporal features, neglecting the interrelationships between users, devices, and ads, or focusing only on node-related characteristics. They lack in-depth analysis of the temporal patterns of fraudulent behavior, making it difficult to adapt to the combined temporal and network correlations inherent in fraudulent behavior. This results in weak model generalization and poor adaptability to diverse fraud methods.

[0004] These shortcomings make it difficult for existing advertising risk control solutions to accurately identify complex and ever-changing advertising fraud behaviors, and they cannot meet the real-time and reliability requirements of advertising risk control. Summary of the Invention

[0005] The main purpose of this application is to provide an advertising risk control method and system based on multi-dimensional behavioral characteristics, which aims to solve the problems of one-sided data perception, rough signal processing, weak prediction and evaluation capabilities, single fault tolerance mechanism and lack of optimization closed loop in the existing technology.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an advertising risk control method based on multi-dimensional behavioral characteristics, comprising the following steps: S1. Collect multi-source behavioral data during the advertising interaction process, and perform normalization and missing value filling on the collected data in sequence to obtain a standardized multidimensional behavioral dataset. S2. Based on the standardized multidimensional behavior dataset, extract features from four dimensions: basic behavior, temporal behavior, related behavior, and semantic behavior, and construct a multidimensional behavior feature set; S3. The multidimensional behavioral feature set is dynamically weighted and fused using an attention mechanism to obtain a fused feature vector, and the fused feature vector is input into a time-series graph neural network to complete risk state modeling. S4. Calculate the advertising risk score based on the output of the time-series graph neural network, set a dynamic risk threshold, and output the corresponding risk control decision based on the comparison result of the risk score and the dynamic risk threshold.

[0007] Optionally, step S1 specifically includes: S11. Collect basic behavioral data, time-series behavioral data, related behavioral data, and semantic behavioral data during the advertising interaction process to form a multi-source behavioral data set; S12. An improved min-max normalization algorithm is used to unify the dimensions of the multi-source behavioral data set to obtain a normalized dataset. S13. Use a categorical imputation strategy to fill in the missing values ​​in the normalized dataset and output the standardized multidimensional behavior dataset.

[0008] Optionally, step S2 specifically includes: S21. Based on the standardized basic behavioral data, extract the effective exposure click-through rate, click dwell time ratio, conversion completion rate, and click location concentration features. S22. Based on the standardized time-series behavioral data, extract the trend coefficient, periodic coefficient, and mutation coefficient features; S23. Construct an undirected association graph of user device advertisements based on standardized association behavior data, and extract the degree centrality, betweenness centrality and clustering coefficient features of the nodes in the graph; S24. Based on the standardized semantic behavior data, extract the semantic vector mean, semantic vector variance, and sentiment tendency features using the BERT semantic coding model. S25. Based on all the features extracted in steps S21 to S24, the multidimensional behavioral feature set is constructed.

[0009] Optionally, step S21 specifically includes: S211. Calculate the effective exposure click-through rate by counting the number of effective exposures and the number of effective clicks, and calculate the effective exposure click-through rate based on the ratio between the two. S212. Calculate the average click-and-dwell time and combine it with the number of valid clicks to calculate the click-and-dwell time ratio; S213. Assign corresponding weights according to conversion type, and calculate the conversion completion rate by combining the ratio of the number of conversions of different types to the number of valid clicks. S214. Calculate the dispersion of the click location coordinates and then calculate the concentration of the click location based on the dispersion.

[0010] Optionally, step S3 specifically includes: S31. Employ a multi-head self-attention mechanism to dynamically weight and fuse a multi-dimensional behavioral feature set, and output a fused feature vector. S32. Use the fused feature vector as the initial feature of the node in the user device advertising association graph, and input it into the temporal graph neural network; S33. Update node temporal information through gated recurrent units and aggregate neighbor node features through graph convolutional layers; S34. Perform global average pooling on all node features and output the risk state vector.

[0011] Optionally, step S31 specifically includes: S311. Map the four dimensions of behavioral features to the same dimensional space through linear transformations to obtain the mapped feature vectors for each dimension. S312. Concatenate the feature vectors of each dimension to obtain the feature matrix, and map the feature matrix to generate the query matrix, key matrix and value matrix respectively. S313. Split the query matrix, key matrix, and value matrix according to the set number of attention heads, and calculate the output result for each attention head; S314. Concatenate the outputs of all attention heads and perform a linear transformation. Combine residual connections and layer normalization to obtain the fused feature vector.

[0012] Optionally, step S32 specifically includes: S321. Based on the user device advertising undirected association graph constructed in step S23, clarify the identification information of all nodes in the graph and their association with their neighbors; S322. Distribute the fused feature vectors output in step S31 according to node type, and use them as the initial features for user nodes, device nodes, and advertising nodes respectively. S323. Perform dimension adaptation processing on the initial features of each node to ensure that they match the input dimension of the temporal graph neural network. After completing the node feature initialization, input the node features into the temporal graph neural network.

[0013] Optionally, step S33 specifically includes: S331. The node's historical features and fused feature vector are processed through the reset gate and update gate of the gated loop unit to generate candidate hidden states, and then the node's temporal feature vector at the current time is updated. S332. The node's own historical features and the features of its neighboring nodes are weighted and aggregated through a graph convolutional layer, and then normalized by combining the node degree to obtain the aggregated node feature vector.

[0014] Optionally, step S4 specifically includes: S41. The risk state vector is mapped to a set numerical range using the sigmoid activation function to obtain the advertising risk score. S42. Based on the statistical parameters of historical advertising data and the real-time fraud rate of the current scenario, a dynamic risk threshold is calculated, wherein the dynamic risk threshold is adapted to different scenarios through adaptive adjustment coefficients. S43. Set the fault tolerance coefficient and output risk control decisions such as normal deployment, further review, or suspension of deployment based on the relationship between the risk score and the dynamic risk threshold.

[0015] Secondly, this application provides an advertising risk control system based on multi-dimensional behavioral characteristics, including: The data acquisition and preprocessing module is used to collect multi-source behavioral data during the advertising interaction process, perform normalization and missing value imputation in sequence, and output a standardized multidimensional behavioral dataset. The multidimensional feature extraction module is used to extract features from four dimensions—basic behavior, temporal behavior, related behavior, and semantic behavior—based on the standardized multidimensional behavior dataset, and to construct a multidimensional behavior feature set. The dynamic fusion modeling module is used to dynamically weight and fuse the multi-dimensional behavioral feature set using an attention mechanism to obtain a fusion feature vector. The fusion feature vector is then input into a time-series graph neural network to complete risk state modeling and output a risk state vector. The risk assessment module is used to calculate the advertising risk score based on the risk state vector, set a dynamic risk threshold, and output the corresponding risk control decision by comparing the magnitude of the risk score and the dynamic risk threshold.

[0016] Through the above technical solutions, the beneficial effects of this invention are as follows: This application, by constructing a full-dimensional feature system covering basic behavior, temporal behavior, related behavior, and semantic behavior, breaks through the limitation of single feature dimensions in existing solutions, and can deeply capture the core features of covert advertising fraud behavior; by using an attention mechanism to achieve dynamic weighted fusion of multi-dimensional features, it can adapt to the differences in the importance of each feature under different scenarios, effectively avoid the information redundancy problem caused by fixed fusion modes, and improve feature utilization efficiency; by using a temporal graph neural network to organically integrate the temporal pattern of behavior with the correlation characteristics of users, devices, and advertisements, it significantly enhances the model's adaptability to various fraud methods and improves the real-time performance and robustness of risk control identification; by combining historical data statistical patterns with the dynamic risk threshold set by the real-time fraud rate in the current scenario, it can flexibly adapt to the risk level of different advertising scenarios, effectively reduce the probability of misjudgment and missed judgment, and finally, through the synergistic effect of various technical links, it achieves accurate identification and efficient control of complex and varied advertising fraud behavior, which not only reduces the waste of advertisers' marketing expenses, but also maintains the fairness and stability of the digital advertising ecosystem. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention, and the embodiments in the accompanying drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an advertising risk control method based on multi-dimensional behavioral characteristics provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of an advertising risk control system based on multi-dimensional behavioral features provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0020] The foregoing and other technical contents, features and effects of the present invention are described in conjunction with the appendix below. Figure 1-3 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings.

[0021] Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] In one exemplary embodiment, such as Figure 1 As shown, an advertising risk control method based on multi-dimensional behavioral characteristics is provided. This method achieves accurate identification and control of advertising fraud behavior through the implementation of steps S1-S4. The specific workflow and detailed operation of each step are as follows: Step S1 involves collecting multi-source behavioral data during the advertising interaction process. The collected data is then normalized and imputed for missing values ​​to obtain a standardized multidimensional behavioral dataset. This step specifically includes three sub-steps: S11. Collect basic behavioral data, time-series behavioral data, related behavioral data, and semantic behavioral data during the advertising interaction process to form a multi-source behavioral data set; S12. An improved min-max normalization algorithm is used to unify the dimensions of the multi-source behavioral data set to obtain a normalized dataset. S13. Use a categorical imputation strategy to fill in the missing values ​​in the normalized dataset and output the standardized multidimensional behavior dataset.

[0023] In one specific embodiment, firstly, the terminal SDK, server log storage system, and user behavior tracking module of the advertising platform are used to comprehensively capture four types of core behavioral data throughout the entire advertising interaction process, ultimately forming a multi-source behavioral data set. Basic behavioral data refers to the core operational data generated when a user directly interacts with an advertisement, including exposure duration, click location coordinates, click dwell time, conversion operation type, and click frequency. Exposure duration is the duration from when the advertisement finishes loading until the user closes or redirects to the next page. Click location coordinates are the screen pixel positions corresponding to when a user clicks the advertisement; for example, if a user clicks the upper right corner of an information feed advertisement, its coordinates might be (1080, 450). Temporal behavioral data focuses on the temporal distribution characteristics of the behavior, including behavior timestamps, behavior interval sequences, and behavior sequence codes. The behavior interval sequence is a sequence composed of the time difference between two consecutive user interactions with an advertisement; for example, if a user clicks an advertisement at 9:00 and views it again at 9:05, the corresponding behavior interval is 300 seconds. Association-based behavioral data is used to characterize the relationships between users, devices, and advertisements. This includes user device fingerprints, the number of times a user is associated with an advertisement, the density of associations between devices and advertisements, and the duration of user-device binding. User device fingerprints are unique identifier strings generated based on multi-dimensional information such as device hardware information, operating system version, and network IP, used to accurately distinguish different interactive devices. Semantic behavioral data covers text and path information generated during user interactions, including user comments or sharing text, and browsing path node information. Browsing path node information refers to the sequence of pages a user is redirected to after clicking an advertisement; for example, after clicking an e-commerce advertisement, a user is redirected to the product details page, shopping cart page, and checkout page. Through these multi-channel collection methods, the comprehensiveness and completeness of behavioral data are ensured, providing a rich data source for subsequent analysis.

[0024] Secondly, since different types of behavioral data have significantly different units of measurement—for example, click frequency is measured in "times / hour" while click duration is measured in "seconds"—directly using them for subsequent calculations can lead to an imbalance in feature weights. Therefore, normalization is needed to unify the units of measurement. The improved min-max normalization algorithm introduces a minimum parameter on top of the traditional algorithm to avoid the abnormal case of a zero denominator, while also optimizing the data mapping interval. Its standardized mathematical expression is: in, These are the normalized eigenvalues. These are the original behavioral data values. This represents the theoretical minimum value for this type of behavioral data. This represents the theoretical maximum value for this type of behavioral data. =10 -5 This is the minimum parameter. The algorithm maps all original data to the interval [0,1], preserving the relative distribution characteristics of the data while eliminating the influence of dimensional differences. For example, if the original value of the click-through duration of an advertisement is 20 seconds, and the theoretical minimum value for this type of data is 0 seconds and the maximum value is 120 seconds, substituting these values ​​into the formula yields a normalized value of (20 - 0 + 10). -5 ) / (120-0+2×10 -5 The value is approximately 0.167. This calculation enables a unified scale conversion of data with different dimensions, resulting in a normalized dataset.

[0025] Finally, to address the issue of missing data due to network interruptions, equipment malfunctions, and other factors during data collection, a data type-specific imputation strategy was adopted to ensure the integrity of the dataset. For time-series behavioral data, a time-weighted interpolation imputation method was used. This method assigns weights based on the time distance between the missing value and the valid data before and after it, with closer values ​​receiving greater weights. Its mathematical expression is: in, For missing data values, , These are the first and second missing values, respectively. One valid data value, =1 / These are the weighting coefficients. =3 represents the size of the selected effective data window.

[0026] For example, if a user's five consecutive ad clicks occur within the time interval sequence [120, 150, missing, 180, 210], substituting this into the formula yields a missing value of [(1×150+0.5×120)+(1×180+0.5×210)] / [2×(1+0.5)]=165. For related behavioral data, the cluster mean imputation method is used. First, K-means clustering is used to group users or devices with similar behavioral patterns into the same cluster. Then, the mean of the same data within each cluster is calculated as the imputation value for missing values. For example, if 1000 users are clustered into 5 clusters based on device usage habits, and the average number of times users in a certain cluster are associated with ads is 8, then users in that cluster with missing association counts are uniformly imputed with 8. For basic behavioral data, due to its numerical characteristics, zero-value imputation is suitable, so missing values ​​are directly filled with 0. For semantic behavioral data, because it involves text information, empty strings are used as imputation values ​​to avoid formatting errors during text processing. After filling in all missing values, output a standardized multidimensional behavioral dataset with a complete structure and uniform format.

[0027] Step S2: Based on the standardized multidimensional behavior dataset, features are extracted from four dimensions: basic behavior, temporal behavior, related behavior, and semantic behavior, to construct a multidimensional behavior feature set. The specific implementation process of this step is as follows: S21. Based on the standardized basic behavioral data, extract the effective exposure click-through rate, click dwell time ratio, conversion completion rate, and click location concentration features. S22. Based on the standardized time-series behavioral data, extract the trend coefficient, periodic coefficient, and mutation coefficient features; S23. Construct an undirected association graph of user device advertisements based on standardized association behavior data, and extract the degree centrality, betweenness centrality and clustering coefficient features of the nodes in the graph; S24. Based on the standardized semantic behavior data, extract the semantic vector mean, semantic vector variance, and sentiment tendency features using the BERT semantic coding model. S25. Based on all the features extracted in steps S21 to S24, the multidimensional behavioral feature set is constructed.

[0028] In one specific embodiment, firstly, four key features are extracted based on standardized basic behavioral data. The calculation of each feature combines data characteristics and business logic design. Effective exposure click-through rate (EPSR) measures the conversion efficiency from ad exposure to effective clicks. Effective exposure refers to ad display duration of at least 1 second without being quickly closed by the user. Effective clicks refer to genuine clicks excluding accidental clicks (click duration less than 0.5 seconds). The calculation formula is as follows: in, To effectively increase exposure and click-through rate, For valid clicks, This refers to the number of effective impressions. For example, if a news feed ad has 2000 effective impressions and 80 effective clicks, the effective impression-click-through rate (EPS) is 0.04. A value that is too high or too low may indicate abnormal behavior.

[0029] Click-through time ratio is used to assess a user's level of interest after clicking on an ad. It is calculated as the ratio of average click-through time to the number of valid clicks. The formula is as follows: in, The ratio of click-through time to dwell time. The average dwell time for all valid clicks (in seconds).

[0030] Conversion completion rate measures the probability that a user will complete a target action after clicking. It is weighted differently based on conversion type (e.g., download, registration, purchase), and the calculation formula is as follows: in, To improve the conversion completion rate, For the first The weight of class transformation (the sum of the weights is 1). Let i be the number of transformations of type i. This represents the total number of conversion types. For example, if the conversion types include download, registration, and purchase, with corresponding conversion counts of 20, 15, and 10 respectively, and the number of valid clicks is 80, the conversion completion rate calculation result can intuitively reflect the user's willingness to convert after clicking.

[0031] Click location concentration is used to characterize the concentration of user click locations. Cheating behavior often exhibits a high concentration of click locations, calculated by the dispersion of click coordinates. The formula is as follows: in, Concentration of click locations Click coordinates The variance of the axis, Click coordinates The variance of the axis indicates that the smaller the variance, the more concentrated the click locations. For example, 10 valid clicks on an advertisement. The variance of the axial coordinates is 25. If the variance of the axis coordinates is 36, then the concentration of click positions is 30.5, reflecting that the click positions are relatively concentrated.

[0032] Secondly, focusing on the temporal distribution patterns of these behaviors, we explore the temporal evolution characteristics of user behavior and extract three core features: trend coefficient, periodicity coefficient, and mutation coefficient. The trend coefficient describes the overall trend of the behavioral interval sequence and is calculated using the slope of a linear regression. The formula is as follows: in, This is the trend coefficient. The length of the time series sequence. For the first The behavior interval (unit: seconds) is calculated. A positive trend coefficient indicates that the behavior interval is gradually increasing and the user interaction frequency is decreasing; a negative trend coefficient indicates that the behavior interval is gradually decreasing and the interaction frequency is increasing.

[0033] The periodicity coefficient is used to identify whether a behavior has a fixed time period. It is calculated by transforming the time-domain sequence to the frequency domain using a Fourier transform, and taking the maximum amplitude value in the frequency domain as the periodicity coefficient. The calculation formula is as follows: in, For periodic coefficients, The result is the Fourier transform. The unit is imaginary. For example, clicks triggered by cheat scripts often have a fixed cycle, and their periodicity coefficient is significantly higher than that of random interactions by normal users.

[0034] The mutation coefficient is used to capture anomalous mutations in behavioral interval sequences, and the calculation formula is as follows: in, The mutation coefficient is... The mean of the interval sequence is denoted as . The value range is 2 to A higher mutation coefficient indicates a more significant and sudden change in the behavioral interval, which may be an abnormally triggered behavior.

[0035] Next, an undirected association graph of user device advertising is constructed based on the standardized related behavior data. This graph is a network model depicting the relationships among users, devices, and advertisements. The nodes in the graph are divided into three categories: user nodes (uniquely identifying a single user), device nodes (uniquely identifying a single interactive device), and advertisement nodes (uniquely identifying a single advertisement). The edges in the graph represent the associations between nodes. For example, an edge between a user node and a device node indicates that the user has used the device to interact with the advertisement, and an edge between a device node and an advertisement node indicates that the device has triggered an advertisement exposure or click. Based on this association graph, three core association features of the nodes are extracted: degree centrality, betweenness centrality, and clustering coefficient. Degree centrality reflects the tightness of the association between nodes in the network, and the calculation formula is: in, For nodes Degree centrality, For nodes The degree (number of associated edges). This represents the total number of nodes in the association graph. For example, if a device node is associated with 10 user nodes and 8 ad nodes, and the total number of nodes is 100, then the degree centrality of this device node is approximately 0.182. An excessively high degree centrality may indicate that the device is engaging in fraudulent behavior by maliciously associating with multiple users or ads.

[0036] Betweenness centrality reflects the mediating role of a node in a network, that is, the probability that a node acts as a mediator in the shortest path between other nodes. The formula for calculating it is: in, For nodes betweenness centrality, For the set of all nodes, For nodes To the node The total number of shortest paths, For the nodes The number of shortest paths.

[0037] The clustering coefficient reflects the degree of clustering of neighboring nodes, and the calculation formula is: in, For nodes Clustering coefficient, For nodes The actual number of edges between neighboring nodes. A higher clustering coefficient indicates that the nodes are more closely related to each other, forming a dense local network.

[0038] Then, the BERT semantic encoding model is used to process the standardized semantic behavior data. The BERT model is a pre-trained language model based on a bidirectional Transformer architecture, capable of deeply capturing the contextual semantic information of text, avoiding the limitations of traditional text processing methods that only focus on local words. First, user comments or sharing texts and browsing path node descriptions are converted into token sequences that the model can process. After inputting into the BERT model, the hidden layer vector of each token is obtained. Based on this vector sequence, three core semantic features are extracted: semantic vector mean, semantic vector variance, and sentiment value. The semantic vector mean is used to represent the overall semantic information of the text, and its calculation formula is: in, The mean of the semantic vector. The length of the token sequence. For the first The hidden layer vector of each token (768 dimensions).

[0039] Semantic vector variance is used to measure the degree of dispersion of text semantics, and the calculation formula is as follows: in, The variance of the semantic vector.

[0040] Sentiment score is used to determine a user's attitude towards an advertisement (positive, negative, neutral). It is calculated by inputting the mean of the semantic vector into a fully connected layer and applying a sigmoid activation function. The calculation formula is as follows: in, This is the sentiment tendency value (the value range is [0,1]). This is the weight matrix of the fully connected layer. For bias terms, This is a sigmoid activation function. For example, if a user comments that an advertisement is false and completely inconsistent with the description, the processed sentiment score might be 0.15, reflecting a negative attitude; while if a user comments that a product is useful and the advertisement description is truthful, the sentiment score might be 0.88, reflecting a positive attitude.

[0041] Finally, all the extracted features are summarized and integrated to form a multi-dimensional behavioral feature set. This set includes 13 core features: basic behavioral features (4), temporal behavioral features (3), related behavioral features (3), and semantic behavioral features (3). Each feature has been standardized with a uniform value range, comprehensively depicting the interaction characteristics between users and advertisements from four dimensions: basic behavioral operations, temporal patterns, correlations, and semantic attitudes. Through this step, scattered behavioral data is transformed into feature vectors with clear physical meaning that can support risk identification, providing a solid feature foundation for subsequent dynamic fusion modeling and risk assessment.

[0042] Step S3 involves using an attention mechanism to dynamically weight and fuse the multidimensional behavioral feature set to obtain a fused feature vector. This fused feature vector is then input into a time-series graph neural network to complete risk state modeling. The specific implementation process of this step is as follows: First, a multi-head self-attention mechanism is used to dynamically weight and fuse a multi-dimensional behavioral feature set. This mechanism can adaptively identify the importance of different features in risk identification, avoid redundancy of effective information caused by fixed weighting, and finally output a fused feature vector. The specific process consists of four steps: The first step is feature dimension mapping. The feature vectors of the four dimensions—basic behavior, temporal behavior, relational behavior, and semantic behavior—are each mapped to the same dimensional space through independent linear transformations, ensuring feature fusion. The mathematical expression is: in, For the first The mapping feature vector of each dimension ( =1, 2, 3, 4 correspond to the four feature dimensions respectively. For the first A linear transformation weight matrix in one dimension. This is a bias term. For example, if the original dimension of the basic behavioral feature vector is 128, it is mapped to 256 dimensions through a linear transformation, maintaining consistency with other dimensional features.

[0043] The second step is feature matrix construction and mapping, which involves concatenating the four-dimensional mapped feature vectors in sequence to form a unified feature matrix. Then, the query matrix Q, key matrix K, and value matrix V are generated through three different linear transformations, which are used for subsequent attention calculations.

[0044] The third step is multi-head attention computation. Q, K, and V are equally divided according to the set number of attention heads (e.g., 8). Each attention head independently calculates the feature association weights and outputs the results. The mathematical expression for single-head attention output is: in, For the first The output of each attention head Characteristic matrix Dimensions To focus on the number of heads, This is a scaling factor to prevent gradient vanishing due to excessively large calculation results.

[0045] The fourth step is fusion and normalization. The outputs of all attention heads are concatenated and integrated into a unified vector through a linear transformation. This vector is then concatenated with the residual vector of the feature mean to alleviate the vanishing gradient problem during model training. Finally, after layer normalization, the fused feature vector is obtained. The mathematical expression is: in, For the final fused feature vector, This is the concatenated linear transformation weight matrix. The feature mean residual vector, This is a layer normalization operation to ensure stable feature distribution.

[0046] Secondly, the fused feature vector is used as the initial feature of the nodes in the undirected association graph of user device advertisements, and after initialization, it is input into the temporal graph neural network.

[0047] Based on the undirected association graph of user device advertisements constructed in step S23, the unique identifier information and neighbor relationships of all nodes in the graph are first clarified. For example, the neighbor nodes of user node U2 include its bound device nodes D3 and D5, and the advertising nodes A2 and A4 that it has interacted with. Then, fusion feature vectors are assigned according to node type: fusion features focusing on user behavior patterns are assigned to user nodes, fusion features emphasizing device association characteristics are assigned to device nodes, and fusion features highlighting advertising interaction patterns are assigned to advertising nodes, which serve as the initial features for the three types of nodes respectively. Finally, the initial features of each node are subjected to dimensionality adaptation processing. Through zero padding or feature truncation, it is ensured that they completely match the input dimensions of the temporal graph neural network to avoid model input dimension incompatibility. After the node feature initialization is completed, the association graph structure and the initial node features are input into the temporal graph neural network.

[0048] Next, through the synergistic effect of the gated recurrent unit (GRU) and the graph convolutional layer, node temporal information updates and neighbor node feature aggregation are achieved respectively, deeply mining the temporal correlation and network correlation of behaviors. On the one hand, the gated recurrent unit (GRU) is used to update the node temporal information. The GRU dynamically adjusts the weights of the node's historical features and the current fused features through reset gates and update gates, effectively capturing long-term temporal behavioral dependencies. The mathematical expression of its update process is: in, For the first Class Node The door to reset time. To update the door, For the first Class Node Historical feature vector at any given moment For candidate feature vectors, For element-wise multiplication, , , This is the weight matrix. , , This is an offset term. For example, device node D3 in... The temporal characteristics of a moment will be combined with its The historical interaction characteristics at any given moment are updated with the current fusion characteristics to capture the temporal changes in device behavior.

[0049] On the other hand, graph convolutional layers are used to aggregate features of neighboring nodes. Based on the relationships between nodes, the temporal features of neighboring nodes are weighted and aggregated to the current node. At the same time, node degree is used for normalization to avoid excessively high feature proportions for nodes with high degrees. The mathematical expression is: in, For the first Class Node The aggregated feature vector at each time step, For the first The set of neighboring nodes of a class node. , The first Class node and neighbor node The degree, This is the weight matrix representing the node's own features. This is the weight matrix of the features of neighboring nodes. The activation function is used to introduce nonlinear feature representation capabilities.

[0050] Then, global average pooling is performed on the feature vectors of all nodes after time-series updates and association aggregation to eliminate the influence of differences in the number of nodes on the feature output, resulting in a risk state vector that comprehensively reflects the overall risk status of the advertisement. The mathematical expression for global average pooling is: in, Let be the risk state vector. This represents the total number of nodes in the undirected graph of user device advertising. For the first Nodes The feature vector after time-series aggregation. For example, the association graph contains 100 user nodes, 50 device nodes, and 30 advertising nodes, with a total of 180 nodes. By calculating the average of the feature vectors of the 180 nodes, a risk state vector with unified dimensions is obtained. This vector comprehensively integrates the temporal and association features of users, devices, and advertisements, thus completing the risk state modeling.

[0051] By implementing step S3, the dispersed multidimensional behavioral features are transformed into a risk state vector with strong expressive power. This not only retains the core information of each dimension feature, but also integrates the temporal patterns and correlation characteristics of behavior, providing an accurate and comprehensive modeling foundation for subsequent advertising risk score calculation and risk control decision-making.

[0052] Step S4: Calculate the advertising risk score based on the output of the temporal graph neural network, set a dynamic risk threshold, and output the corresponding risk control decision based on the comparison between the risk score and the dynamic risk threshold, thereby achieving effective control over advertising fraud behavior. The specific implementation process is as follows: First, the risk state vector is mapped to a set numerical range using the sigmoid activation function, resulting in an ad risk score that intuitively reflects the probability of ad fraud. The sigmoid activation function is a commonly used non-linear mapping function that can compress vector values ​​from any real number range into the [0,1] interval, where the values ​​directly represent the risk probability. The closer the score is to 1, the higher the probability of ad fraud; the closer it is to 0, the higher the probability of ad fraudulent behavior. Its mathematical expression is: in, Score the risk of advertising. The risk state vector output in step S3 This is the output layer weight matrix. This is the output layer bias term. For example, after the risk state vector of an advertisement is calculated by the output layer, substituting it into the formula yields a risk score of 0.93, indicating that the probability of the advertisement engaging in fraudulent behavior is extremely high; while another advertisement has a risk score of 0.12, which is likely a normal advertisement.

[0053] Secondly, based on historical advertising data statistics and the real-time fraud rate in the current scenario, a dynamic risk threshold is calculated. This threshold can adapt to the risk level of different advertising scenarios, avoiding misjudgments or missed judgments caused by fixed thresholds. The calculation of the dynamic risk threshold combines the statistical regularity of historical data with the real-time status of the current scenario, and its mathematical expression is: in, For dynamic risk thresholds, This represents the average of historical normal advertising risk scores. The standard deviation of historical normal advertising risk scores. For adaptive adjustment coefficients, and , The real-time fraud rate for the current advertising scenario is calculated using the following formula: , This represents the number of fraudulent ads identified in the current scenario. This represents the total number of ads in the current scenario. For example, the historical average risk score of normal ads in a certain e-commerce promotion scenario. =0.3, standard deviation =0.1, current real-time cheating rate =0.08, then the adaptive adjustment coefficient =1.1 + 0.4 × 0.08 = 1.132, Dynamic Risk Threshold =0.3 + 1.132 × 0.1 = 0.4132. This calculation allows for a higher threshold in scenarios with high fraud rates to reduce false positives for legitimate ads, and a lower threshold in scenarios with low fraud rates to prevent missed detections of fraudulent ads.

[0054] Then, a tolerance coefficient is set to balance the rigor and flexibility of risk control decisions. Based on the relationship between risk scores and dynamic risk thresholds, three types of risk control decisions are output in a tiered manner. (Tolerance coefficient) =0.05, its function is to reserve a buffer for advertisements with risk scores near the threshold, avoiding decision-making bias caused by small fluctuations. The specific decision-making logic is as follows: When advertising risk score If the risk of the advertisement is significantly lower than the risk level of the current scenario, it is judged as a normal advertisement, and a "normal placement" decision is output, allowing the advertisement to continue to be displayed on the corresponding channel; when If the advertisement is suspicious and cannot be directly judged as normal or fraudulent, a "further review" decision will be output, and a second verification will be carried out by a human review team or a more sophisticated special algorithm. when If the ad's risk level is significantly higher than the current scenario's risk level, it is judged as fraudulent advertising. A "pause" decision is output, immediately stopping ad display and adding the associated user or device to a risk control blacklist, restricting their subsequent interaction permissions. For example, the risk score of an ad... =0.47, dynamic risk threshold =0.4132, fault tolerance coefficient =0.05, then + =0.4632. Since 0.47 > 0.4632, the output is "pause delivery".

[0055] Finally, through the aforementioned quantitative calculations and hierarchical decision-making logic, precise control of advertising risks is achieved. This step not only uses the sigmoid function to probabilistically represent risk, but also adapts to the risk differences in different scenarios through dynamic thresholds. Furthermore, it balances control precision and flexibility through fault tolerance coefficients and three-level decision-making, effectively solving the problems of misjudgment and missed judgment caused by fixed thresholds in existing solutions. This ensures that risk control decisions can accurately block fraudulent advertisements while protecting the legitimate rights of normal advertisements, thus guaranteeing the fairness of the digital advertising ecosystem.

[0056] Based on the same inventive concept, this application also provides an advertising risk control system based on multi-dimensional behavioral features for implementing the advertising risk control method based on multi-dimensional behavioral features described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the advertising risk control system based on multi-dimensional behavioral features provided below can be found in the limitations of the advertising risk control method based on multi-dimensional behavioral features described above, and will not be repeated here.

[0057] In one exemplary embodiment, such as Figure 2 As shown, an advertising risk control system based on multi-dimensional behavioral features is provided, including a data collection and preprocessing module, a multi-dimensional feature extraction module, a dynamic fusion modeling module, and a risk assessment module, wherein: The data acquisition and preprocessing module collects multi-source behavioral data during the advertising interaction process, performs normalization and missing value imputation sequentially, and outputs a standardized multidimensional behavioral dataset. Through the terminal acquisition components, server log system, and behavior tracking module of the advertising platform, it comprehensively collects basic behavioral data, temporal behavioral data, correlated behavioral data, and semantic behavioral data during the advertising interaction process, forming a complete multi-source dataset. Subsequently, an improved min-max normalization algorithm is used to unify the scale of data with different dimensions, eliminating the impact of data differences on subsequent analysis. Finally, for missing data that appears during the acquisition process, appropriate imputation strategies are used according to data type to complete the information, ultimately outputting a standardized multidimensional behavioral dataset with a complete structure and uniform format.

[0058] The multidimensional feature extraction module is used to extract features from four dimensions—basic behavior, temporal behavior, related behavior, and semantic behavior—based on the standardized multidimensional behavior dataset, constructing a multidimensional behavior feature set. Specifically, this module extracts key features from each of the four dimensions: basic behavior, temporal behavior, related behavior, and semantic behavior. The basic behavior dimension focuses on extracting core indicators from direct user interaction patterns; the temporal behavior dimension mines the temporal evolution characteristics of behavior; the related behavior dimension extracts node association features by constructing an association graph; and the semantic behavior dimension extracts textual semantic features using a professional coding model. Finally, the features extracted from all dimensions are summarized and integrated to construct a multidimensional behavior feature set covering all attributes of behavior. This module transforms the raw data into feature vectors with clear risk identification significance, uncovering the essential differences between normal behavior and cheating behavior, and providing core analytical basis for risk modeling.

[0059] The dynamic fusion modeling module employs an attention mechanism to dynamically weight and fuse the multi-dimensional behavioral feature set to obtain a fused feature vector. This fused feature vector is then input into a temporal graph neural network to complete risk state modeling and output a risk state vector. First, an attention mechanism is used to dynamically weight and fuse the multi-dimensional behavioral feature set, adaptively adjusting the importance weights of different features to avoid information redundancy, and outputting a fused feature vector. Then, this fused feature vector is assigned to various nodes in the user device advertising association graph according to node type, serving as the initial node features input to the temporal graph neural network. Gated recurrent units update the temporal information of the nodes, capturing the temporal dependence patterns of behavior. Simultaneously, graph convolutional layers aggregate the features of neighboring nodes to mine the correlation characteristics between nodes. Finally, global average pooling is performed on all node features to output a risk state vector that comprehensively reflects the advertising risk state. The function of this module is to fuse the temporal patterns and correlation characteristics of behavior, construct a risk model with strong generalization ability, and accurately characterize the risk level of advertising.

[0060] The risk assessment module calculates an advertising risk score based on the risk state vector, sets a dynamic risk threshold, and outputs a corresponding risk control decision by comparing the risk score with the dynamic risk threshold. This module maps the risk state vector to a fixed numerical range using an activation function to obtain a risk score that intuitively reflects the probability of advertising fraud. It combines statistical parameters from historical advertising data with the real-time fraud rate of the current campaign scenario to calculate a dynamic risk threshold that adapts to different risk levels in different scenarios. It sets a tolerance coefficient to balance the strictness and flexibility of the decision-making process, and outputs risk control decisions based on the relationship between the risk score and the dynamic risk threshold, such as further review of normal campaigns or suspension of campaigns. The purpose of this module is to achieve quantitative assessment and hierarchical control of advertising risks, accurately blocking fraudulent ads while avoiding misjudgments of normal ads, thus ensuring the fairness and effectiveness of the advertising ecosystem.

[0061] In summary, this invention overcomes the limitations of existing solutions with their single feature dimension by constructing a comprehensive feature system covering basic behavior, temporal behavior, related behavior, and semantic behavior. This system can deeply capture the core features of covert advertising fraud. By leveraging an attention mechanism to achieve dynamic weighted fusion of multi-dimensional features, it can adapt to the differences in the importance of each feature in different scenarios, effectively avoiding information redundancy caused by fixed fusion patterns and improving feature utilization efficiency. Through a temporal graph neural network, it organically integrates the temporal patterns of behavior with the correlation characteristics of users, devices, and advertisements, significantly enhancing the model's adaptability to various fraud methods and improving the real-time performance and robustness of risk control identification. By combining historical data statistical patterns with a dynamic risk threshold set based on the real-time fraud rate in the current scenario, it can flexibly adapt to the risk levels of different advertising scenarios, effectively reducing the probability of false positives and false negatives. Finally, through the synergistic effect of various technical aspects, it achieves accurate identification and efficient control of complex and varied advertising fraud behaviors, reducing advertisers' marketing expenses while maintaining the fairness and stability of the digital advertising ecosystem.

[0062] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores advertising risk control data based on multi-dimensional behavioral characteristics. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an advertising risk control method based on multi-dimensional behavioral characteristics.

[0063] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0064] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0065] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0066] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0067] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Memory includes ROM, magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0069] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An advertising risk control method based on multi-dimensional behavioral characteristics, characterized in that, Includes the following steps: S1. Collect multi-source behavioral data during the advertising interaction process, and perform normalization and missing value filling on the collected data in sequence to obtain a standardized multidimensional behavioral dataset. S2. Based on the standardized multidimensional behavior dataset, extract features from four dimensions: basic behavior, temporal behavior, related behavior, and semantic behavior, and construct a multidimensional behavior feature set; S3. The multidimensional behavioral feature set is dynamically weighted and fused using an attention mechanism to obtain a fused feature vector, and the fused feature vector is input into a time-series graph neural network to complete risk state modeling. S4. Calculate the advertising risk score based on the output of the time-series graph neural network, set a dynamic risk threshold, and output the corresponding risk control decision based on the comparison result of the risk score and the dynamic risk threshold.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Collect basic behavioral data, time-series behavioral data, related behavioral data, and semantic behavioral data during the advertising interaction process to form a multi-source behavioral data set; S12. An improved min-max normalization algorithm is used to unify the dimensions of the multi-source behavioral data set to obtain a normalized dataset. S13. Use a categorical imputation strategy to fill in the missing values ​​in the normalized dataset and output the standardized multidimensional behavior dataset.

3. The method according to claim 2, characterized in that, Step S2 specifically includes: S21. Based on the standardized basic behavioral data, extract the effective exposure click-through rate, click dwell time ratio, conversion completion rate, and click location concentration features. S22. Based on the standardized time-series behavioral data, extract the trend coefficient, periodic coefficient, and mutation coefficient features; S23. Construct an undirected association graph of user device advertisements based on standardized association behavior data, and extract the degree centrality, betweenness centrality and clustering coefficient features of the nodes in the graph; S24. Based on the standardized semantic behavior data, extract the semantic vector mean, semantic vector variance, and sentiment tendency features using the BERT semantic coding model. S25. Based on all the features extracted in steps S21 to S24, the multidimensional behavioral feature set is constructed.

4. The advertising risk control method based on multi-dimensional behavioral characteristics according to claim 3, characterized in that, Step S21 specifically includes: S211. Calculate the effective exposure click-through rate by counting the number of effective exposures and the number of effective clicks, and calculate the effective exposure click-through rate based on the ratio between the two. S212. Calculate the average click-and-dwell time and combine it with the number of valid clicks to calculate the click-and-dwell time ratio; S213. Assign corresponding weights according to conversion type, and calculate the conversion completion rate by combining the ratio of the number of conversions of different types to the number of valid clicks. S214. Calculate the dispersion of the click location coordinates and then calculate the concentration of the click location based on the dispersion.

5. The advertising risk control method based on multi-dimensional behavioral characteristics according to claim 1, characterized in that, Step S3 specifically includes: S31. Employ a multi-head self-attention mechanism to dynamically weight and fuse a multi-dimensional behavioral feature set, and output a fused feature vector. S32. Use the fused feature vector as the initial feature of the node in the user device advertising association graph, and input it into the temporal graph neural network; S33. Update node temporal information through gated recurrent units and aggregate neighbor node features through graph convolutional layers; S34. Perform global average pooling on all node features and output the risk state vector.

6. The advertising risk control method based on multi-dimensional behavioral characteristics according to claim 5, characterized in that, Step S31 specifically includes: S311. Map the four dimensions of behavioral features to the same dimensional space through linear transformations to obtain the mapped feature vectors for each dimension. S312. Concatenate the feature vectors of each dimension to obtain the feature matrix, and map the feature matrix to generate the query matrix, key matrix and value matrix respectively. S313. Split the query matrix, key matrix, and value matrix according to the set number of attention heads, and calculate the output result for each attention head; S314. Concatenate the outputs of all attention heads and perform a linear transformation. Combine residual connections and layer normalization to obtain the fused feature vector.

7. The advertising risk control method based on multi-dimensional behavioral characteristics according to claim 5, characterized in that, Step S32 specifically includes: S321. Based on the user device advertising undirected association graph constructed in step S23, clarify the identification information of all nodes in the graph and their association with their neighbors; S322. Distribute the fused feature vectors output in step S31 according to node type, and use them as the initial features for user nodes, device nodes, and advertising nodes respectively. S323. Perform dimension adaptation processing on the initial features of each node to ensure that they match the input dimension of the temporal graph neural network. After completing the node feature initialization, input the node features into the temporal graph neural network.

8. The advertising risk control method based on multi-dimensional behavioral features according to claim 5, characterized in that, Step S33 specifically includes: S331. The node's historical features and fused feature vector are processed through the reset gate and update gate of the gated loop unit to generate candidate hidden states, and then the node's temporal feature vector at the current time is updated. S332. The node's own historical features and the features of its neighboring nodes are weighted and aggregated through a graph convolutional layer, and then normalized by combining the node degree to obtain the aggregated node feature vector.

9. The advertising risk control method based on multi-dimensional behavioral characteristics according to claim 1, characterized in that, Step S4 specifically includes: S41. The risk state vector is mapped to a set numerical range using the sigmoid activation function to obtain the advertising risk score. S42. Based on the statistical parameters of historical advertising data and the real-time fraud rate of the current scenario, a dynamic risk threshold is calculated, wherein the dynamic risk threshold is adapted to different scenarios through adaptive adjustment coefficients. S43. Set the fault tolerance coefficient and output risk control decisions such as normal deployment, further review, or suspension of deployment based on the relationship between the risk score and the dynamic risk threshold.

10. An advertising risk control system based on multi-dimensional behavioral characteristics, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source behavioral data during the advertising interaction process, perform normalization and missing value imputation in sequence, and output a standardized multidimensional behavioral dataset. The multidimensional feature extraction module is used to extract features from four dimensions—basic behavior, temporal behavior, related behavior, and semantic behavior—based on the standardized multidimensional behavior dataset, and to construct a multidimensional behavior feature set. The dynamic fusion modeling module is used to dynamically weight and fuse the multi-dimensional behavioral feature set using an attention mechanism to obtain a fusion feature vector. The fusion feature vector is then input into a time-series graph neural network to complete risk state modeling and output a risk state vector. The risk assessment module is used to calculate the advertising risk score based on the risk state vector, set a dynamic risk threshold, and output the corresponding risk control decision by comparing the magnitude of the risk score and the dynamic risk threshold.

Images