A financial behavior pattern analysis and prediction method based on time series chart representation learning

Abstract

The application belongs to the technical field of financial transaction security, and particularly relates to a financial behavior mode analysis and prediction method based on time sequence chart representation learning, which comprises the following steps: obtaining transaction flow data; obtaining flow embedding vectors based on an event encoder and a time encoder; learning the flow embedding vectors based on an autoregressive model, and obtaining a representation matrix through self-supervised contrast learning training; constructing an interaction graph, modeling node interaction on graph structure features using a graph neural network to form a graph enhancement module, and enhancing the representation matrix; and constructing a machine learning classification model to realize financial behavior mode analysis and prediction. Compared with the prior art, the application can automatically capture more complex and abstract information from massive information, mine the time sequence, interaction and business correlation between account flow and labels, and realize rapid and wide application of the model in the absence of business expert experience analysis.

Description

Technical Field

[0001] This invention relates to the field of financial transaction security technology, and in particular to a method for analyzing and predicting financial behavior patterns based on time series graph representation learning. Background Technology

[0002] With the development of information technology in the financial sector, financial transactions are recorded, generating massive amounts of transaction data. This transaction data contains a wealth of information and is indispensable for improving financial services and preventing financial risks. By modeling the historical behavior of accounts through transaction data, it is possible to analyze and predict financial behavior patterns, which can be applied to areas such as fraud prevention, risk identification and prevention, and abnormal behavior pattern detection.

[0003] Traditional feature extraction methods only make simple changes to the original data, failing to fully capture more complex and abstract concepts within the information. In practical applications, summarizing business characteristics through expert analysis often cannot automatically and promptly capture new behavioral models in transaction log data, exhibiting certain lag and limitations.

[0004] CN 111797177 A discloses a financial time series classification method and its application for detecting abnormal financial accounts. This method constructs and expands a financial time series dataset of financial accounts from transaction flow data of both abnormal and normal financial accounts. It uses a neural network model with multiple stacked blocks (each block containing LocalBiLSTM, Self-Attention, residual connections, LayerNormalization, and Position-wise Feed-Forward Networks) to simultaneously extract local and global pattern features from the financial time series. Finally, a softmax classification layer is used to classify the financial time series, ultimately achieving the function of detecting abnormal financial accounts. However, this method still lacks sufficient feature representation ability for transaction flow data and does not extract and utilize account interaction information from transaction flow data, thus limiting the model's prediction performance. Summary of the Invention

[0005] The purpose of this invention is to provide a financial behavior pattern analysis and prediction method based on time series graph representation learning, which does not rely on human feature engineering and can effectively capture the time series, interaction and business information in the transaction flow, thereby improving the prediction effect.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for analyzing and predicting financial behavior patterns based on time-series graph representation learning includes the following steps:

[0008] Obtain transaction log data;

[0009] The business characteristics and time-series characteristics of each transaction data are input into the event encoder and time encoder respectively to obtain the transaction embedding vector;

[0010] The flow embedding vector is learned based on an autoregressive model and trained through self-supervised contrastive learning to obtain the representation matrix;

[0011] Using accounts and counterparties in the transaction data as nodes and transaction flow as edges, an interactive information graph is constructed.

[0012] Graph neural networks are used to model the node interactions on graph structure features, forming a graph augmentation module. The graph augmentation module can enhance the representation matrix's ability to represent account interaction information, resulting in the weight matrix of the graph neural network.

[0013] A machine learning classification model is constructed based on the enhanced representation matrix and account labels, and the weight matrix of the machine learning classification model is obtained.

[0014] By updating the weight matrix of the machine learning classification model and the weight matrix of the graph neural network through backpropagation, the model is trained to achieve financial behavior pattern analysis and prediction.

[0015] The transaction log data is represented as follows:

[0016] S u =(e u1 ,e u2 ,…,e un )

[0017] Among them, S u This represents the transaction log data for account u, where n is the number of transactions and e is the transaction record. ui For the i-th transaction record, e ui =(u,v) ui ,t ui ,attr ui ), which is a quadruple, v ui Indicates the counterparty in the transaction, t ui Indicates timestamp, attr ui Represents an attribute.

[0018] The event encoder encodes the attribute information in the transaction events of account u:

[0019]

[0020] Where, num i For all numerical features in the i-th transaction record, cat ij Let W,W be the j-th categorical feature in the i-th transaction record.j is the feature transformation matrix, used to map data into a d-dimensional vector, and Emb(·) represents the embedding mapping of categorical attributes.

[0021] The time encoder encodes the timestamp information of the transaction events of account u: for time-series encoding, the purpose is to construct a continuous function mapping from the time domain to the vector space. It also possesses translation invariance, meaning there exists a ψ that satisfies... The time encoder is:

[0022]

[0023] Time encoding is obtained through a time encoder.

[0024] The encoding of each event in the pipeline embedding vector is the sum of the encoding results of the event encoder and the encoding results of the time encoder:

[0025]

[0026] The process of extracting features from the transaction embedding vector of account u based on an autoregressive model and constructing the hidden state is as follows:

[0027] The encoded pipeline embedding vector is represented as (x1,…,x). n The input is fed into an autoregressive model, which is then used to extract the latent states of the flowing embedding vectors.

[0028] z1,…,z i =AR(x1,…,x i ), i = 1, ..., n

[0029] Where AR represents the autoregressive model, z i It contains all the information up to the i-th record, and n represents the number of transactions.

[0030] The purpose of the self-supervised contrastive learning training is to utilize the hidden state z of account u. i Constructing the representation vector, specifically:

[0031] According to current historical data u1 ,…,e ui Records of the future u,i+1 To distinguish between them, the method includes the following steps:

[0032] Given the encoded representations of the first i transaction records x1,…,x i The hidden states z1,…,z are obtained through an autoregressive model. i The representation of the (i+1)th transaction record is x. i+1 Hidden state zi The degree of match with this record is represented by a bilinear function:

[0033]

[0034] Among them, W bilinear For parameter matrices;

[0035] A single transaction record e′ is randomly selected from all transaction log data as a negative sample, and its feature representation x is obtained through an event encoder. (event) The time representation is obtained by using the time of the (i+1)th real record. Thus, the feature representation of the negative samples is obtained:

[0036]

[0037] Similarly, calculate the hidden state z i Degree of match with the negative sample:

[0038]

[0039] The probability of successfully predicting future events is:

[0040]

[0041] Therefore, the optimization objective of self-supervised contrastive learning is defined as maximizing the log-likelihood of successful predictions:

[0042]

[0043] Finally, the feature representation of account u is obtained by concatenation:

[0044]

[0045] Among them, z n This represents the hidden state of all information up to the nth record.

[0046] The graph-based enhancement module enhances the representation matrix in the following ways:

[0047] The transaction history of each account can be viewed as an edge, and the account and its counterparty can be viewed as nodes in the graph. The graph is represented as G = (V, A), where V is the set of nodes in the graph. It is an adjacency matrix, assuming S is the set of transaction flows.

[0048] Then, the adjacency matrix of the graph is represented as:

[0049]

[0050] Among them, u iThis represents the i-th account and its counterparty v. j This corresponds to the j-th account, where t represents the timestamp and attr represents the attribute;

[0051] Let D = diag(d1, d2, ..., d n ) is the degree matrix of the graph, where d i =∑ j A ij ;

[0052] Using graph neural networks for information propagation on interactive graphs: Let L represent the normalized adjacency matrix with added self-loops.

[0053]

[0054] in

[0055] The update of the representation matrix is ​​then represented as sparse matrix multiplication:

[0056]

[0057] in, The enhanced representation matrix, d is the representation matrix generated by the autoregressive model through self-supervised contrastive learning training, W is the weight matrix of the graph neural network, and h is the number of network layers.

[0058] The autoregressive model can be implemented using LSTM, GRU, TCN, or Transformer sequence models. The machine learning classification model can be implemented using Bayesian classification, decision trees, logistic regression, support vector machines, or neural network models. The machine learning classification model implemented using logistic regression yields the following probability for account u to be abnormal:

[0059]

[0060] in, Let q be the representation vector of account u, and q be the weight matrix of the logistic regression model.

[0061] The loss function for the logistic regression model is:

[0062]

[0063] Where y u ∈{0,1, where ∈{0,1,1} represents the actual label corresponding to account u.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) This invention is based on a deep autoregressive model to model the embedded representations of the temporal and business features of transaction flow data. A representation model of the transaction flow data is obtained through self-supervised comparative learning. It can automatically capture more complex and abstract information from massive amounts of data and learn multi-level abstract representations of the data. On the one hand, it eliminates the need for domain experts to manually design features, reducing manual workload and enabling rapid model application even in the absence of business expert experience, thus solving the limitations of traditional methods. On the other hand, by timely capturing new behavioral patterns in the data, existing models can be transferred and applied to new application scenarios, overcoming the lag challenge faced by traditional methods.

[0066] (2) The present invention constructs a graph enhancement module based on a graph neural network, which realizes the extraction of account interaction information in the transaction flow and trains and updates the graph neural network through the backpropagation mechanism. Thus, based on the temporal and business feature representation, the account interaction information is further utilized to enhance the representation model, which can significantly improve the representation capability.

[0067] (3) Based on time series graph representation learning, this invention adopts a machine learning classification model. It uses the representation of transaction flow and account labels to construct a machine learning classification model, which can fully explore the implicit time series, interaction and business relationship between account transaction flow and labels. This financial behavior analysis and prediction method based on time series graph representation learning can be widely applied in the fields of anti-fraud, risk identification and prevention, abnormal behavior pattern detection and other fields. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention;

[0069] Figure 2 A schematic diagram illustrating the process of converting transaction flow data into a time series diagram;

[0070] Figure 3 This is the model architecture of the present invention. Detailed Implementation

[0071] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0072] A single transaction often involves two accounts, which can be organized into a format of (account, transaction account, timestamp, associated attributes). The interaction behavior can be constructed as a time-series interaction graph. Figure 2 An example of converting transaction log data into a time series diagram.

[0073] The time series graph representation learning model can automatically extract highly abstract behavioral data features from massive amounts of data in transaction data without the aid of feature engineering, and learn multi-level abstract representations of the data, thereby achieving a high-dimensional representation of transaction flow.

[0074] Therefore, this invention proposes a financial behavior pattern analysis and prediction method based on time-series graph representation learning for account transaction flow data. First, the business characteristics and time-series characteristics of each transaction flow are input into the event encoder and time encoder respectively to obtain a transaction flow embedding vector. Then, an autoregressive model is used to model the transaction flow embedding vector, and a self-supervised contrastive learning approach is used to train the model to obtain a representation model of the transaction flow. Next, a graph augmentation module based on a graph neural network is used to enhance the representation model. Based on the enhanced representation and account labels, a machine learning classification model is constructed, which can uncover the implicit time-series, interaction, and business relationships between account transaction flow and labels, achieving accurate pattern analysis and prediction.

[0075] The core of this invention is to automatically generate a representation model and a machine learning classification model for financial transactions using a time series graph representation learning method, and then perform pattern analysis and prediction based on these models. This invention can be applied to tasks such as anti-fraud, risk identification and prevention, and abnormal behavior pattern detection.

[0076] Specifically, this embodiment provides a method for analyzing and predicting financial behavior patterns based on time-series graph representation learning, such as... Figure 1 As shown, it includes the following steps:

[0077] 1) Obtain transaction log data.

[0078] Transaction flow data is represented as follows:

[0079] S u =(e u1 ,e u2 ,…,e un )

[0080] Among them, S u This represents the transaction history data for account u, e ui For the i-th transaction record, e ui =(u,v) ui ,t ui ,attr ui ), which is a quadruple, v ui Indicates the counterparty in the transaction, t ui Indicates timestamp, attr ui Represents an attribute.

[0081] This embodiment uses Figure 2(a) Taking the simulated transaction data shown as an example, the data is preprocessed. This dataset has five fields. The numerical field is the transaction amount, which is normalized. For categorical fields, such as the counterparty identifier, one-hot encoding is used. Assuming there are three categorical values ​​a, b, and c, they will be encoded as three-dimensional features:

[0082] {a:[1,0,0],b:[0,1,0],c:[0,0,1]}.

[0083] To address the issue that an excessive number of categories may lead to an overly sparse feature matrix, this embodiment selects the top-k categories as the main features and unifies the other categories with a smaller proportion into one category.

[0084] 2) Input the business characteristics and time-series characteristics of each transaction data of account u into the event encoder and time encoder respectively to obtain the transaction embedding vector.

[0085] The event encoder encodes the attribute information in a transaction event:

[0086]

[0087] Where, num i For all numerical features in the i-th transaction record, cat ij Let W,W be the j-th categorical feature in the i-th transaction record. j is the feature transformation matrix, used to map data into a d-dimensional vector, and Emb(·) represents the embedding of categorical attributes.

[0088] The time encoder encodes the timestamp information of the transaction event: for time-series coding, the goal is to construct a continuous function mapping from the time domain to the vector space. It also possesses translation invariance, meaning there exists a ψ that satisfies...

[0089] The time encoder used in this embodiment is:

[0090]

[0091] Time encoding is obtained through a time encoder.

[0092] The encoding of each event in the pipeline embedding vector is the sum of the encoding results of the event encoder and the time encoder:

[0093]

[0094] In this embodiment, the fields corresponding to event characteristics (such as counterparty bank identifier, transaction amount, etc.) in the preprocessed transaction log data are input into the event encoder, and the timestamp field is input into the time encoder. Therefore, the i-th transaction event characteristic is represented as:

[0095]

[0096] Specifically, the characteristics of the first event are as follows:

[0097] Simultaneously utilize The function encodes the timestamp, resulting in:

[0098]

[0099] Similarly, the event and time characteristics of other transactions can be represented, and through... Obtain the encoding.

[0100] 3) The flow embedding vector of account u is learned based on the auto-regressive model (AR) and trained through self-supervised contrastive learning to obtain the representation model.

[0101] The learning of the flow embedding vector based on the autoregressive model is specifically as follows:

[0102] The encoded pipeline embedding vector is represented as (x1,…,x). n The input is fed into an autoregressive model, which is then used to extract the latent states of the flowing embedding vectors.

[0103] z1,…,z i =AR(x1,,…,x i ), i = 1, ..., n

[0104] Among them, z i It contains all the information up to the i-th record, and n represents the number of transactions.

[0105] For example, z1 = [1.5331, -2.1234, ..., 0.9824].

[0106] In this embodiment, the autoregressive model can be implemented using sequence models such as LSTM, GRU, TCN, or Transformer.

[0107] Taking a single-layer Transformer as an example, its calculation process is as follows:

[0108]

[0109]

[0110] Here, W is the parameter matrix. Of course, a multi-layered, multi-head Transformer model can also be used.

[0111] The purpose of training through self-supervised comparative learning is to train based on the current history e of account u. u1 ,…,e ui Records of the future u,i+1 To distinguish them, the method is as follows:

[0112] Given the encoded representations of the first i transaction records x1,…,x i The hidden states z1,…,z are obtained through an autoregressive AR model. i The representation of the (i+1)th transaction record is x. i+1 Hidden state z i The degree of match with this record is represented by a bilinear function:

[0113]

[0114] Among them, W bilinear It is a parameter matrix.

[0115] The matching degree calculated in this embodiment

[0116] Randomly select a transaction record e′ from all transaction logs (which may be from other accounts) as a negative sample, and obtain its feature representation x using an event encoder. (event) The time representation is obtained by using the time of the (i+1)th real record. Thus, the feature representation of the negative samples is obtained:

[0117]

[0118] Similarly, calculate the hidden state z i Degree of match with the negative sample:

[0119]

[0120] The matching degree calculated in this embodiment

[0121] The probability of successfully predicting future events is:

[0122]

[0123] Therefore, the optimization objective of self-supervised contrastive learning is defined as maximizing the log-likelihood of successful predictions:

[0124]

[0125] This embodiment calculates the gradient of the sample based on the negative log-likelihood and optimizes the model parameters using gradient descent, with W... bilinear For example:

[0126]

[0127] Where lr is the learning rate.

[0128] Finally, the feature representation of account u is obtained by concatenation:

[0129]

[0130] Among them, Z n This represents the hidden state of all information up to the nth record.

[0131] The above steps can be used to obtain the representation vector Z of account u. u For example, Z1 = [1.5123, 0.6412, ..., 3.4212].

[0132] 4) Construct an interactive information graph using accounts and counterparties in the transaction data as nodes and transaction flow as connecting edges.

[0133] 5) Use graph neural networks to model the node interactions on the graph structure features to form a graph augmentation module. The graph augmentation module can enhance the representation matrix's ability to represent account interaction information, thus obtaining the weight matrix of the graph neural network.

[0134] To incorporate information from the account transaction interaction graph into the representation model, this embodiment introduces a graph augmentation module. Each account's transaction history is considered as an edge, and the account and its counterparties are considered nodes in the graph, represented as G = 9V, A), where V is the set of nodes (accounts) in the graph, and A ∈ R. |V|×|V| It is an adjacency matrix, assuming S is the set of transaction flows.

[0135] The adjacency matrix of the graph is represented as:

[0136]

[0137] Among them, u i This represents the i-th account and its counterparty v. j This corresponds to the j-th account, where t represents the timestamp and attr represents the attribute.

[0138] Based on the above calculation process, the adjacency matrix A of the graph calculated in this embodiment is:

[0139]

[0140] Let D = diag(d1, d2, ..., d n Let d be the degree matrix of the graph, where d i =∑ j A ij ;

[0141] Using graph neural networks for information propagation on interactive graphs: Let L represent the normalized adjacency matrix with added self-loops.

[0142]

[0143] in

[0144] The node representation update is then represented as sparse matrix multiplication:

[0145]

[0146] in, The enhanced representation matrix, d is the representation matrix generated by the autoregressive model through self-supervised contrastive learning training, W is the weight matrix of the graph neural network, and h is the number of network layers.

[0147] This example yields the enhanced representation vector of account u.

[0148] Based on the above method, it is possible to effectively learn the representation of financial transaction data, which can be integrated into downstream tasks such as fraud and anomaly detection to achieve financial behavior analysis and prediction. Its model structure diagram is shown below. Figure 3 As shown.

[0149] 6) Construct a machine learning classification model based on the enhanced representation matrix and account labels, and obtain the weight matrix of the machine learning classification model.

[0150] 7) The weight matrix of the machine learning classification model and the weight matrix of the graph neural network are updated by backpropagation to train the model and realize the analysis and prediction of financial behavior patterns.

[0151] In this embodiment, the machine learning classification model is implemented using methods such as Bayesian classification, decision trees, logistic regression, support vector machines, and neural network models.

[0152] In this embodiment, the financial behavior pattern analysis and prediction is for fraudulent account identification. The machine learning classification model is implemented through a logistic regression model, and the probability that account u is abnormal is:

[0153]

[0154] in, Let q be the representation vector of account u, and q be the weight matrix of the logistic regression model.

[0155] The loss function for the logistic regression model is:

[0156]

[0157] Where y u ∈{0,1}, where represents the true label corresponding to account u. This embodiment obtains the probability that account u is abnormal:

[0158]

[0159] Using the method of this invention, the model can automatically learn the transaction flow representation of each account at the latest moment and apply it to subsequent financial behavior pattern analysis and prediction tasks. The transaction flow representation not only includes business characteristic information but also covers temporal and interaction characteristic information, thus making the distinction between normal and abnormal accounts in the vector space more obvious. In a real anti-fraud business scenario, this embodiment compared the effect of this invention with the rule model and found that this invention can greatly make up for the shortcomings of the rule model in both public and private scenarios. In the field of anti-telecom fraud, it achieves a high accuracy and high coverage in identifying fraudulent accounts, with an F1 score improvement of 10% (private) and 48% (public) compared to the previous rule model.

[0160] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for analyzing and predicting financial behavior patterns based on time-series graph representation learning, characterized in that, Includes the following steps: Obtain transaction log data; The business characteristics and time-series characteristics of each transaction data are input into the event encoder and time encoder respectively to obtain the transaction embedding vector; The flow embedding vector is learned based on an autoregressive model and trained through self-supervised contrastive learning to obtain the representation matrix; Using accounts and counterparties in the transaction data as nodes and transaction flow as edges, an interactive information graph is constructed. Graph neural networks are used to model the node interactions on graph structural features, forming a graph enhancement module. The representation matrix is ​​then enhanced based on the graph enhancement module to obtain the weight matrix of the graph neural network. A machine learning classification model is constructed based on the enhanced representation matrix and account labels, and the weight matrix of the machine learning classification model is obtained. By updating the weight matrix of the machine learning classification model and the weight matrix of the graph neural network through backpropagation, the model is trained to achieve financial behavior pattern analysis and prediction.

2. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 1, characterized in that, The transaction log data is represented as follows: in, Indicates account u Transaction flow data, n For the number of transactions, For the first Transaction records , which is a quadruple. Indicates the counterparty in the transaction. Represents a timestamp. Represents an attribute.

3. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 2, characterized in that, The event encoder is for the account. u Encode the attribute information in the transaction event: in, For the first All numerical characteristics in the transaction record, For the first The first in the transaction record Each categorical feature, The feature transformation matrix is ​​used to map data to... dimensional vector, This indicates that a categorical attribute is embedded and mapped.

4. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 3, characterized in that, The time encoder is for the account. u The timestamp information of the transaction event is encoded: For time-series coding, the purpose is to construct a continuous function mapping from the time domain to the vector space. It also possesses translation invariance, that is, it exists. satisfy The time encoder is: Time encoding is obtained through a time encoder. .

5. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 4, characterized in that, The encoding of each event in the pipeline embedding vector is the sum of the encoding results of the event encoder and the encoding results of the time encoder: 。 6. A method for analyzing and predicting financial behavior patterns based on time-series graph representation learning according to claim 1 or 5, characterized in that, The account based on the autoregressive model u Feature extraction is performed on the pipeline embedding vector to construct the hidden state, specifically: The encoded flow embedding vector The input is fed into an autoregressive model, which is then used to extract the latent states of the flow embedding vector: in, This represents an autoregressive model. Contains the first This record contains all the information from the previous records. n This indicates the number of transactions.

7. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 6, characterized in that, The purpose of the self-supervised comparative learning training is to utilize accounts. u Hidden state Constructing the representation vector is as follows: Based on current history Records of the future To distinguish between them, the method includes the following steps: Before Encoded representation of each transaction record The hidden state is obtained through an autoregressive model. , No. The representation of a transaction record is as follows Hidden state The degree of match with this record is represented by a bilinear function: in, For parameter matrices; Randomly select one transaction record from all transaction data. As a negative sample, its feature representation is obtained through an event encoder. , using the The time of a real record is represented by time. Thus, the feature representation of the negative samples is obtained: Similarly, calculate the hidden states. Degree of match with the negative sample: The probability of successfully predicting future events is: Therefore, the optimization objective of self-supervised contrastive learning is defined as maximizing the log-likelihood of successful predictions: Ultimately, the account was obtained by piecing together the information. u Representation vector: in, For the first The hidden state of all previous information in the record.

8. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 1, characterized in that, The graph-based enhancement module enhances the representation matrix in the following ways: Viewing each account's transaction history as an edge, and the account and its counterparty as nodes in the graph, the graph is represented as follows: ,in It is the set of nodes in a graph. It is an adjacency matrix, assuming A collection of transaction flows. Then, the adjacency matrix of the graph is represented as: in, Indicates the first Individual account, counterparty Corresponding to the first Individual accounts, t Represents a timestamp. attr Indicates an attribute; remember Let be the degree matrix of the graph, where ; Using graph neural networks for information propagation on interactive graphs: Let... Represents the normalized adjacency matrix with added self-loops. in ; The representation update is then expressed as sparse matrix multiplication: in, The enhanced representation matrix, The representation matrix is ​​generated by the autoregressive model through self-supervised contrastive learning training. Dimensions for representing account features W Here is the weight matrix of the graph neural network. This represents the number of network layers.

9. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 1, characterized in that, The autoregressive model can be implemented using LSTM, GRU, TCN, or Transformer sequence models, while the machine learning classification model can be implemented using Bayesian classification, decision trees, logistic regression, support vector machines, or neural network models.

10. The financial behavior pattern analysis and prediction method based on time series graph representation learning according to claim 9, characterized in that, The machine learning classification model implemented using logistic regression obtains the account. The probability of it being abnormal is: in, For account u The representation vector, q This is the weight matrix for the logistic regression model; The loss function for the logistic regression model is: in For account The corresponding real tags.

Images