Information fraud identification method based on big data and artificial intelligence

Abstract

The invention discloses an information fraud identification method based on big data and artificial intelligence, and relates to the technical field of information security, and the method comprises the steps: S1, multi-source data collection and processing, S2, dynamic heterogeneous graph construction, S3, feature encoder design, S4, heterogeneous space-time attention mechanism, S5, multi-modal fusion, and S6, dynamic fraud detection model. According to the method, by introducing a heterogeneous space-time diagram attention mechanism, the fraud detection precision is improved, specifically, by designing a space-time incidence matrix and dynamically calculating the space-time attention weight between nodes, abnormal behaviors can be effectively recognized, and the fraud detection accuracy is improved. The space-time correlation matrix combines node features and time difference information, the dynamic evolution law of user behaviors can be captured, the complex fraud mode with space-time correlation can be more accurately recognized through the design, and the accuracy and reliability of fraud detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of information security, and particularly relates to a method for identifying information fraud based on big data and artificial intelligence. Background Art

[0002] In the fields of financial payment, e-commerce, etc., the complexity and concealment of fraud behaviors are increasing day by day, and many limitations of traditional fraud detection methods have gradually emerged. Existing fraud detection technologies mainly rely on static feature analysis and rule engines. Although these methods can identify known fraud patterns to a certain extent, they often perform poorly in the face of new and complex fraud behaviors. Specifically, traditional methods usually analyze based on static features and cannot capture the temporal evolution law of user behaviors. For example, features such as the transaction frequency and amount distribution of users will change over time, and static methods are difficult to effectively model these dynamic patterns. In view of the above problems, the following solutions are proposed. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for identifying information fraud based on big data and artificial intelligence. By introducing a heterogeneous spatio-temporal graph attention mechanism, the accuracy of fraud detection is improved, and the problem that existing methods usually only rely on static features or single-modal data and are difficult to capture complex spatio-temporal correlation patterns is solved.

[0004] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0005] The present invention is a method for identifying information fraud based on big data and artificial intelligence, including:

[0006] Step S1, multi-source data collection and processing: Collect user attributes, transaction records, device fingerprints, text descriptions, and relational network data, construct a time series matrix and a heterogeneous graph structure, and extract text features;

[0007] Step S2, dynamic heterogeneous graph construction: Define a heterogeneous graph containing multiple node types and edge relationships, and divide it by time slices;

[0008] Step S3, feature encoder design: The time series encoder is based on temporal dilation convolution to extract time series features, and the graph structure encoder is based on the graph attention network to extract graph structure features;

[0009] Step S4, heterogeneous spatio-temporal attention mechanism: Design a spatio-temporal correlation matrix, calculate the spatio-temporal correlation degree between nodes, generate spatio-temporal attention weights, and fuse the time series features and graph structure features by calculating the final node representation;

[0010] Step S5, Multimodal Fusion: Define a multimodal fusion function, and combine the text semantic information extracted by BERT to fuse the graph structure features and text features through adaptive weights;

[0011] Step S6, Dynamic Fraud Detection Model: Construct an end-to-end training framework, and combine cross-entropy loss, regularization loss, and contrastive loss to optimize the model parameters;

[0012] Step S7, Online Incremental Learning: Through the sliding window mechanism and momentum update, achieve dynamic update of the model parameters to adapt to new fraud patterns;

[0013] Step S8, Visualization Explanation System: Generate a report to display the decision path and feature importance scores, and analyze the spatio-temporal propagation pattern of fraud behavior;

[0014] In the said step S4, the spatio-temporal correlation matrix in the heterogeneous spatio-temporal attention mechanism is:

[0015]

[0016] In the formula, M ij is the spatio-temporal correlation degree between nodes i and j (used to measure the correlation strength between two nodes in time and space, and is used for subsequent attention weight calculation), Φ(·) is the spatio-temporal correlation function (used to map node features and time difference to the correlation degree), are the temporal feature representations of nodes i and j respectively (representing the behavioral patterns of nodes in the time dimension), Δt ij is the time difference between nodes i and j (representing the time interval between the occurrences of the behaviors of two nodes, and is used to capture the temporal correlation), W φ is a learnable weight matrix (used to map the non-linearly transformed features to the correlation degree space), tanh(·) is the hyperbolic tangent activation function, W t , W h are both learnable weight matrices (used to transform the features of nodes i and j respectively), w Δt is a learnable weight vector (used to weight the time difference feature), h i , h j are the original feature representations of nodes i and j respectively;

[0017] In the said step S4, the calculation formula of the spatio-temporal attention weight in the heterogeneous spatio-temporal attention mechanism is:

[0018]

[0019] In the formula, β ij is the spatio-temporal attention weight between nodes i and j, M ijis the spatio-temporal correlation degree between node i and node j (calculated through the spatio-temporal correlation matrix), exp(·) is the exponential function, and N i is the set of neighbor nodes of node i, and ε is an extremely small positive number (used to prevent the denominator from being zero and ensure numerical stability);

[0020] In step S4, the formula for calculating the final node representation in the heterogeneous spatio-temporal attention mechanism is:

[0021]

[0022] In the formula, is the final feature representation of node i, and β ij is the spatio-temporal attention weight between node i and node j, and W g is a learnable weight matrix (used to transform the temporal features), is the temporal feature representation of node j, and W s is a learnable weight matrix (used to transform the graph structure features), is the graph structure feature representation of node j, and ⊕ is the element-wise product.

[0023] Further, in step S3, the formula of the temporal encoder in the feature encoder design is:

[0024]

[0025] In the formula, TDC(x t ) is the output result of the temporal encoder at time step t after processing the input features based on time dilation convolution, and x t is the input feature at time step t, is the weight of the τ-th convolution kernel with dilation coefficient d, is the convolution operation, b is the bias term, σ is the activation function, and L is the length of the convolution kernel;

[0026] In step S3, the formula of the graph structure encoder in the feature encoder design is:

[0027]

[0028] In the formula, is the feature representation of node i at the l+1-th layer, σ(·) is the non-linear activation function, K is the total number of relationship types (there are multiple edge types in the heterogeneous graph), is the set of neighbor nodes of node i under relationship type k, is the attention weight between node i and node j under relationship type k, and W k is the learnable weight matrix corresponding to relationship type k, is the feature representation of node j at the l-th layer;

[0029] The attention weight The calculation formula is:

[0030]

[0031] Where a k is the learnable attention vector corresponding to relation type k, W is the shared learnable weight matrix, h i ,h j are the original feature representations of node i and node j respectively, || is the vector concatenation operation, LeakyReLU(·) is the leaky linear rectification function, and softmax(·) is the normalized exponential function.

[0032] Furthermore, in step S5, the multimodal fusion function formula in the multimodal fusion is:

[0033] F fusion =λ×F graph +(1-λ)×F text ;

[0034] Where, F fusion is the fused multimodal feature representation, λ is the adaptive weight (used to control the fusion ratio of graph structure features and text features), F graph is the graph structure feature representation, F text It is the text feature representation;

[0035] The calculation formula of the adaptive weight λ is:

[0036] λ=σ(W λ [F graph ||F text ]);

[0037] Where σ(·) is the Sigmoid activation function, W λ is the learnable weight matrix, [F graph ||F text ] is the concatenation operation of graph structure features and text features.

[0038] Furthermore, in step S6, the end-to-end training framework constructed in the dynamic fraud detection model is:

[0039] L=L ce +γL reg +ηL contrast ;

[0040] Where L is the total loss function, L ce is the cross entropy loss, γ is the weight coefficient of regularization loss, L reg is the regularization loss, η is the weight coefficient of contrast loss, Lcontrast is the contrast loss;

[0041] The contrast loss term L contrast has the following calculation formula:

[0042]

[0043] In the formula, sim(·,·) is the similarity calculation function, z i , z j and z

[0044] Further, in step S7, the formula for the sliding window update mechanism in online incremental learning is:

[0045]

[0046] In the formula, W t+1 is the model parameter for the (t + 1)-th iteration, W t is the model parameter for the t-th iteration, W t-1 is the model parameter for the (t - 1)-th iteration, η is the learning rate, is the gradient of the loss function L with respect to the model parameter W t , and μ is the momentum coefficient.

[0047] Further, in step S8, the formula for the importance score of generating a report in the visualization interpretation system is:

[0048]

[0049] In the formula, Importance_score(i) is the feature importance score of node i, Path is the weight vector on the model decision path, ω j is the weight of feature j, Θ is the element-wise multiplication, and x i is the feature value of node i.

[0050] The present invention has the following beneficial effects:

[0051] 1) By introducing the heterogeneous spatio-temporal graph attention mechanism, the present invention improves the accuracy of fraud detection. Specifically, by designing the spatio-temporal correlation matrix and dynamically calculating the spatio-temporal attention weights between nodes, it can effectively identify abnormal behaviors. The spatio-temporal correlation matrix combines node features and time difference information, and can capture the dynamic evolution law of user behaviors. Such a design can more accurately identify complex fraud patterns with spatio-temporal correlations, and improve the accuracy and reliability of fraud detection.

[0052] 2. The present invention can improve the real-time performance of the model by designing an online incremental learning module. Specifically, by adopting a sliding window mechanism and a momentum update strategy, it can dynamically adjust the model parameters in the data stream to adapt to new types of fraud behaviors. The incremental learning module accelerates model convergence and reduces oscillations by combining historical gradient information and the current gradient, ensuring the model remains stable in a dynamic environment. This design enables the model to quickly respond to new fraud patterns and meet the real-time risk control requirements.

[0053] 3. The present invention improves the credibility and practicality of the model by generating an interpretable report. Specifically, by calculating the feature importance scores and generating a visual decision path report, it helps users understand the logic of the model's decisions. The feature importance scores combine the weights and feature values on the decision path, clearly showing the contribution degree of each feature to the model's decisions. In addition, the model supports the visual backtracking of the fraud propagation path to help analyze the spatio-temporal propagation patterns of fraud behaviors. This design improves the transparency of the model and provides support for risk control decisions.

[0054] 4. The present invention realizes the deep fusion of multi-source heterogeneous data through multi-modal adaptive fusion. Specifically, by adopting a gating mechanism, it dynamically adjusts the fusion ratio of graph structure features and text features to ensure that the contribution degrees of the two features can be adaptively adjusted according to the specific scenario. By combining global graph structure information and local semantic information, it can better capture the features of fraud behaviors. This design can enhance the generalization ability and detection effect of the model.

[0055] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic flowchart of a method for identifying information fraud based on big data and artificial intelligence of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] Please refer to Figure 1 As shown, the present invention is an information fraud recognition method based on big data and artificial intelligence, including:

[0060] Step S1, multi-source data collection and processing: Collect user attributes, transaction records, device fingerprints, text descriptions, and relationship network data, construct a time series matrix and a heterogeneous graph structure, and extract text features;

[0061] Step S2, dynamic heterogeneous graph construction: Define a heterogeneous graph containing multiple node types and edge relationships, and divide it by time slices;

[0062] Step S3, feature encoder design: The time series encoder is based on temporal dilation convolution to extract time series features, and the graph structure encoder is based on the graph attention network to extract graph structure features;

[0063] Step S4, heterogeneous spatio-temporal attention mechanism: Design a spatio-temporal correlation matrix, calculate the spatio-temporal correlation degree between nodes, generate spatio-temporal attention weights, and fuse time series features and graph structure features by calculating the final node representation;

[0064] Step S5, multi-modal fusion: Define a multi-modal fusion function, combine the text semantic information extracted by BERT, and fuse the graph structure features and text features through adaptive weights;

[0065] Step S6, dynamic fraud detection model: Construct an end-to-end training framework, combine cross-entropy loss, regularization loss, and contrast loss to optimize model parameters;

[0066] Step S7, online incremental learning: Through the sliding window mechanism and momentum update, realize the dynamic update of model parameters to adapt to new fraud patterns;

[0067] Step S8, visualization interpretation system: Generate a report, display the decision path and feature importance scores, and analyze the spatio-temporal propagation pattern of fraud behavior;

[0068] In step S4, the spatio-temporal correlation matrix in the heterogeneous spatio-temporal attention mechanism is:

[0069]

[0070] In the formula, M ij is the spatio-temporal correlation degree between nodes i and j (used to measure the correlation strength between two nodes in time and space, and is used for subsequent attention weight calculation), Φ(·) is the spatio-temporal correlation function (used to map node features and time difference to correlation degree), are the time series feature representations of node i and node j respectively (representing the behavior patterns of nodes in the time dimension), Δt ijis the time difference between node i and node j (representing the time interval between the occurrences of the behaviors of the two nodes, used to capture the temporal correlation), W φ is a learnable weight matrix (used to map the non-linearly transformed features to the correlation space), tanh(·) is the hyperbolic tangent activation function, W t ,W h are all learnable weight matrices (used to transform the features of node i and node j respectively), w Δt is a learnable weight vector (used to weight the time difference feature), h i ,h j are the original feature representations of node i and node j respectively;

[0071] Step S4, the formula for calculating the spatio-temporal attention weight in the heterogeneous spatio-temporal attention mechanism is:

[0072]

[0073] In the formula, β ij is the spatio-temporal attention weight between node i and node j, M ij is the spatio-temporal correlation degree between node i and node j (obtained through the spatio-temporal correlation matrix), exp(·) is the exponential function, N i is the set of neighbor nodes of node i, and ε is an extremely small positive number (used to prevent the denominator from being zero and ensure numerical stability);

[0074] Step S4, the formula for calculating the final node representation in the heterogeneous spatio-temporal attention mechanism is:

[0075]

[0076] In the formula, is the final feature representation of node i, β ij is the spatio-temporal attention weight between node i and node j, W g is a learnable weight matrix (used to transform the temporal features), is the temporal feature representation of node j, W s is a learnable weight matrix (used to transform the graph structure features), is the graph structure feature representation of node j, and ⊕ is the element-wise product.

[0077] Step S3, the formula for the temporal encoder in the feature encoder design is:

[0078]

[0079] In the formula, TDC(x t ) is the output result of the temporal encoder at time step t after processing the input features based on time dilation convolution, x tis the input feature at time step t, is the weight of the τ-th convolutional kernel with dilation factor d, is the convolution operation, b is the bias term, σ is the activation function, and L is the length of the convolutional kernel;

[0080] Step S3, the formula for the graph structure encoder in the feature encoder design is:

[0081]

[0082] In the formula, is the feature representation of node i at the l+1 layer, σ(·) is the non-linear activation function, K is the total number of relationship types (there are multiple edge types in the heterogeneous graph), is the set of neighbor nodes of node i under relationship type k, is the attention weight between node i and node j under relationship type k, W k is the learnable weight matrix corresponding to relationship type k, is the feature representation of node j at the l layer;

[0083] Attention weight The calculation formula is:

[0084]

[0085] In the formula, a k is the learnable attention vector corresponding to relationship type k, W is the shared learnable weight matrix, h i , h j are the original feature representations of node i and node j respectively, || is the vector concatenation operation, LeakyReLU(·) is the leaky rectified linear unit function, and softmax(·) is the normalized exponential function.

[0086] Step S5, the formula for the multimodal fusion function in multimodal fusion is:

[0087] F fusion = λ × F graph + (1 - λ) × F text ;

[0088] In the formula, F fusion is the fused multimodal feature representation, λ is the adaptive weight (used to control the fusion ratio of the graph structure feature and the text feature), F graph is the graph structure feature representation, F text is the text feature representation;

[0089] The calculation formula for the adaptive weight λ is:

[0090] λ = σ(W λ [Fgraph ||F text );

[0091] Wherein, σ(·) is the Sigmoid activation function, and W λ is the learnable weight matrix, and [F graph ||F text is the concatenation operation of the graph structure feature and the text feature.

[0092] Step S6, the end-to-end training framework constructed in the dynamic fraud detection model is:

[0093] L = L ce + γL reg + ηL contrast ;

[0094] Wherein, L is the total loss function, L ce is the cross-entropy loss, γ is the weight coefficient of the regularization loss, and L reg is the regularization loss, η is the weight coefficient of the contrastive loss, and L contrast is the contrastive loss;

[0095] The calculation formula of the contrastive loss term L contrast is:

[0096]

[0097] Wherein, sim(·,·) is the similarity calculation function, and z i , z j are the feature representations of node i and node j respectively, and τ is the temperature coefficient.

[0098] Step S7, the formula of the sliding window update mechanism in online incremental learning is:

[0099]

[0100] Wherein, W t+1 is the model parameter of the (t + 1)-th iteration, W t is the model parameter of the t-th iteration, W t-1 is the model parameter of the (t - 1)-th iteration, η is the learning rate, is the gradient of the loss function L with respect to the model parameter W t , and μ is the momentum coefficient.

[0101] Step S8, the importance score formula for generating a report in the visualization explanation system is:

[0102]

[0103] Where Importance_score(i) is the feature importance score of node i, Path is the weight vector on the model decision path, ω j is the weight of feature j, Θ is the element-wise multiplication, and x i is the feature value of node i.

[0104] A specific application of this embodiment is as follows:

[0105] 1. Data collection and preprocessing

[0106] Input data: User attributes: 10-dimensional features (age, authentication level, historical risk score, etc.); Transaction records: timestamp, amount, currency, GPS coordinates (accuracy 100m); Device fingerprint: device ID, SIM card serial number, jailbreak status; Text data: transaction description (mixed Chinese and English, average length 15 words); Relationship network: strong association edges with the number of device sharing accounts ≥ 3;

[0107] Preprocessing: Time series construction: Convert the user's behavior in the last 30 days into a time series matrix (time window = 1 hour, feature dimension = 8); Heterogeneous graph construction: Define 4 types of edge relationships (same device, same IP segment, fund receipt and payment, geographical location proximity < 1km); Text processing: Use BERT-wwm to extract transaction description features (dimension = 128);

[0108] 2. Model deployment

[0109] Temporal encoder: Apply dilated convolution (d = 3, L = 5) to the transaction frequency sequence to capture hourly, daily, and weekly periodic patterns:

[0110]

[0111] It is detected that the trading volume of a certain account has increased by 300% in 3 consecutive time windows (interval 3 hours);

[0112] Heterogeneous spatio-temporal attention:

[0113] When it is detected that 5 transactions occur on the same device within 10 minutes, calculate the spatio-temporal correlation matrix:

[0114] M ij = W φ [tanh(0.6h i + 0.3h j + 1.2Δt ij )];

[0115] Generate attention weight β = 0.92 (β < 0.4 for normal transactions), triggering a high-risk warning;

[0116] Multimodal fusion:

[0117] The text feature detected the keyword "purchase", but the amount distribution was abnormal (95% of the transactions < $10), and the adaptive weight λ = 0.73 favored the graphic feature;

[0118] 3. Online detection

[0119] Dynamic update: Incremental learning is performed hourly. When a new fraud pattern appears (such as a new "cryptocurrency" transaction type), the model reaches a 90% detection accuracy within 2 update cycles (μ = 0.9);

[0120] Early warning case: It was detected that a certain device cluster (3 mobile phones) logged in alternately in Paris, Singapore, and New York. Through spatio-temporal path analysis, it was found that the physical movement speed was supersonic (practically infeasible), and it was confirmed as VPN fraud.

[0121] In the description of this specification, the descriptions referring to terms such as "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0122] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not elaborate on all the details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art in the relevant technical field can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An information fraud recognition method based on big data and artificial intelligence, characterized in that, It includes the following steps: Step S1, Multi-source data collection and processing: Collect user attributes, transaction records, device fingerprints, text descriptions, and relational network data, construct a time series matrix and a heterogeneous graph structure, and extract text features; Step S2, Dynamic heterogeneous graph construction: Define a heterogeneous graph containing multiple node types and edge relationships, and divide it by time slices; Step S3, Feature encoder design: The time series encoder is based on temporal dilated convolution to extract time series features, and the graph structure encoder is based on the graph attention network to extract graph structure features; Step S4, Heterogeneous spatio-temporal attention mechanism: Design a spatio-temporal correlation matrix, calculate the spatio-temporal correlation degree between nodes, generate spatio-temporal attention weights, and fuse time series features and graph structure features by calculating the final node representation; Step S5, Multimodal fusion: Define a multimodal fusion function, combine the text semantic information extracted by BERT, and fuse the graph structure features and text features with adaptive weights; Step S6, Dynamic fraud detection model: Construct an end-to-end training framework, combine cross-entropy loss, regularization loss, and contrastive loss to optimize the model parameters; Step S7, Online incremental learning: Through the sliding window mechanism and momentum update, achieve dynamic update of model parameters to adapt to new fraud patterns; Step S8, Visualization explanation system: Generate a report, display the decision path and feature importance scores, and analyze the spatio-temporal propagation pattern of fraud behavior; In step S4, the spatio-temporal correlation matrix in the heterogeneous spatio-temporal attention mechanism is: Where, M ij is the spatio-temporal correlation degree between nodes i and j, and Φ(·) is the spatio-temporal correlation function, are the temporal feature representations of nodes i and j respectively, and Δt ij is the time difference between nodes i and j, W φ is a learnable weight matrix, tanh(·) is the hyperbolic tangent activation function, W t , W h are both learnable weight matrices, w Δt is a learnable weight vector, h i , h j are the original feature representations of nodes i and j respectively; In step S4, the calculation formula for the spatio-temporal attention weights in the heterogeneous spatio-temporal attention mechanism is: where β ij is the spatio-temporal attention weight between node i and node j, M ij is the spatio-temporal correlation degree between node i and node j, exp(·) is the exponential function, N i is the set of neighbor nodes of node i, and ε is a positive number; In step S4, the calculation formula for the final node representation in the heterogeneous spatio-temporal attention mechanism is: wherein, is the final feature representation of node i, β ij is the spatio-temporal attention weight between node i and node j, W g is a learnable weight matrix, is the temporal feature representation of node j, W s is a learnable weight matrix, is the graph structure feature representation of node j, is the element-wise product.

2. The information fraud recognition method based on big data and artificial intelligence according to claim 1, characterized in that, In step S3, the formula for the time series encoder in the feature encoder design is: where TDC(x t ) is the output result of the temporal encoder at time step t after processing the input features based on dilated convolution, x t is the input feature at time step t, is the weight of the τ-th convolutional kernel with a dilation factor of d, is the convolution operation, b is the bias term, σ is the activation function, and L is the length of the convolutional kernel; In step S3, the formula for the graph structure encoder in the feature encoder design is: wherein is the feature representation of node i at the (l + 1)-th layer, σ(·) is a non-linear activation function, K is the total number of relationship types, is the set of neighbor nodes of node i under relationship type k, is the attention weight between node i and node j under relationship type k, W k is the learnable weight matrix corresponding to relationship type k, is the feature representation of node j at the l-th layer; The attention weight The calculation formula is as follows: where a k is the learnable attention vector corresponding to the relationship type k, W is the shared learnable weight matrix, h i , h j are the original feature representations of nodes i and j respectively, || is the vector concatenation operation, LeakyReLU(·) is the leaky rectified linear unit function, and softmax(·) is the normalized exponential function.

3. A method for identifying information fraud based on big data and artificial intelligence according to claim 1, characterized in that, In step S5, the formula for the multimodal fusion function in multimodal fusion is: F fusion = λ × F graph + (1 - λ) × F text ; where F fusion is the fused multi-modal feature representation, λ is the adaptive weight, F graph is the graph structure feature representation, and F text is the text feature representation; The calculation formula for the adaptive weight λ is: λ = σ(W λ [F graph ||F text ); where σ(·) is the Sigmoid activation function, and W λ is the learnable weight matrix, and [F graph ||F text is the concatenation operation of the graph structure feature and the text feature.

4. The information fraud recognition method based on big data and artificial intelligence according to claim 1, wherein In step S6, the end-to-end training framework constructed in the dynamic fraud detection model is: L = L ce + γL reg + ηL contrast ; Where L is the total loss function, L ce is the cross-entropy loss, γ is the weight coefficient of the regularization loss, L reg is the regularization loss, η is the weight coefficient of the contrastive loss, L contrast is the contrastive loss; The contrastive loss term L contrast is calculated as follows: where sim(·,·) is the similarity calculation function, z i , z j are the feature representations of node i and node j respectively, and τ is the temperature coefficient.

5. An information fraud recognition method based on big data and artificial intelligence according to claim 1, characterized in that, In step S7, the formula for the sliding window update mechanism in online incremental learning is: Where, W t+1 is the model parameter for the (t + 1)-th iteration, W t is the model parameter for the t-th iteration, W t-1 is the model parameter for the (t - 1)-th iteration, η is the learning rate, is the gradient of the loss function L with respect to the model parameter W t , and μ is the momentum coefficient.

6. An information fraud recognition method based on big data and artificial intelligence according to claim 1, characterized in that, In step S8, the formula for the importance score of generating a report in the visualization explanation system is: Where Importance_score(i) is the feature importance score of node i, Path is the weight vector on the model decision path, ω j is the weight of feature j, Θ is the element-wise multiplication, and x i is the feature value of node i.

Images