CN-116541755-B - Financial behavior pattern analysis and prediction method based on time sequence diagram representation learning

CN116541755BCN 116541755 BCN116541755 BCN 116541755BCN-116541755-B

Abstract

The invention belongs to the technical field of financial transaction safety, and particularly relates to a financial behavior pattern analysis and prediction method based on time sequence diagram representation learning, which comprises the following steps of obtaining transaction flow data; the method comprises the steps of obtaining a running water embedded vector based on an event encoder and a time encoder, learning the running water embedded vector based on an autoregressive model, obtaining a characterization matrix through self-supervision contrast learning training, constructing an interaction graph, modeling node interaction on graph structural features by using a graph neural network, forming a graph enhancement module, enhancing the characterization matrix, and constructing a machine learning classification model to realize analysis and prediction of financial behavior modes. Compared with the prior art, the method and the system can automatically and fully capture more complex and abstract information from massive information, mine hidden time sequence, interaction and business association between account running water and labels, and realize rapid and wide application of the model under the condition of lacking business expert experience analysis.

Inventors

WANG GUANGZHONG
QIAN FEI
CHEN HAORAN

Assignees

交通银行股份有限公司

Dates

Publication Date: 20260512
Application Date: 20230327

Claims (10)

1. The financial behavior pattern analysis and prediction method based on time sequence diagram representation learning is characterized by comprising the following steps of: Acquiring transaction flow data; the business characteristics and the time sequence characteristics of each transaction stream data are respectively input into an event encoder and a time encoder to obtain a stream embedded vector; learning the running water embedded vector based on the autoregressive model, and obtaining a characterization matrix through self-supervision contrast learning training; taking accounts and transaction opponents in the stream data as nodes, and taking transaction stream as a continuous edge to construct an interactive information graph; Modeling node interaction on the graph structural features by using a graph neural network to form a graph enhancement module, and enhancing the representation matrix based on the graph enhancement module to obtain a weight matrix of the graph neural network; constructing a machine learning classification model based on the enhanced characterization matrix and the account label to obtain a weight matrix of the machine learning classification model; And updating a weight matrix of the machine learning classification model and a weight matrix of the graph neural network through back propagation, and training the model to realize analysis and prediction of the financial behavior mode.
2. The method for predicting a financial behavior pattern analysis based on timing diagram characterization learning of claim 1, wherein the transaction flow data is represented as: Wherein, the Transaction flow data representing account u, n being the transaction number, Is the first A record of the transaction, The method is a four-element group, Representing the opponent of the transaction, The time stamp is indicated as such, Representing the attributes.
3. The method for predicting a pattern analysis of a financial behavior based on timing diagram representation learning of claim 2, the event encoder encodes attribute information in transaction events of account u: Wherein, the Is the first All of the numerical features in a transaction record, Is the first Item of transaction record The characteristics of the individual categories of features, Is a feature transformation matrix for mapping data into The vector of dimensions is used to determine, Representing the embedding mapping of category type attributes.
4. The method of claim 3, wherein the time encoder encodes time stamp information of transaction events of account u for time series encoding with the objective of constructing a continuous function map from time domain to vector space And has translational invariance, i.e. exists Satisfy the following requirements The time encoder is: time encoding by a time encoder 。
5. The method of claim 4, wherein the code of each event in the pipelined embedded vector is a sum of the code result of the event encoder and the code result of the time encoder: 。
6. The method for analyzing and predicting the financial behavior pattern based on time sequence diagram representation learning according to claim 1 or 5, wherein the feature extraction is performed on the flowing water embedded vector of the account u based on the autoregressive model, so as to construct a hidden state, specifically: embedding the encoded stream into a vector Inputting the hidden states of the running water embedded vector into an autoregressive model, and extracting the hidden states of the running water embedded vector by using the autoregressive model: Wherein, the Representing an autoregressive model of the model, Is implication of the first The bar records all previous information, n representing the number of transactions.
7. The method for analysis and prediction of financial behavior pattern based on time chart characterization learning according to claim 6, wherein the purpose of learning by self-supervision contrast is to use hidden state of account u Constructing a characterization vector, specifically: according to the current history For future recordings The implementation method for distinguishing comprises the following steps: Given front Coded representation of a transaction record Obtaining hidden states by means of an autoregressive model First, the The characterization of the strip transaction record is as follows Hidden state The degree of matching to the record is represented by a bilinear function: Wherein, the Is a parameter matrix; randomly extracting a transaction record from all transaction stream data As a negative sample and by means of an event encoder, a characteristic representation thereof is obtained By the first Time-to-time characterization of a real record of a bar Thereby yielding a characteristic representation of the negative sample: similarly, calculate hidden state Degree of match with the negative sample: the probability of successfully predicting a future event is: The optimization objective of the self-supervision contrast learning is defined as the log likelihood of maximizing successful prediction: Finally, the characterization vector of the account u is obtained by stitching: Wherein, the Is the first The bar records the hidden status of all previous information.
8. The method for analyzing and predicting a financial behavior pattern based on time chart characterization learning according to claim 1, wherein the enhancing the characterization matrix based on the graph enhancing module specifically comprises: Considering the transaction sequence of each account as a continuous edge, an account and its transaction counterpart as nodes in the graph, the graph is shown as Wherein Node set of graph Is an adjacency matrix, assuming that For a collection of transaction flows, Then, the adjacency matrix of the graph is expressed as: Wherein, the Represent the first Personal account, transaction opponent Corresponds to the first Accounts, t represents a timestamp, attr represents an attribute; Recording device Is a degree matrix of a graph, in which ; Information propagation on interaction graph using graph neural network Normalized adjacency matrix representing addition of self-loops Wherein the method comprises the steps of ; The token update is expressed as sparse matrix multiplication: Wherein, the In order to enhance the characterization matrix, A characterization matrix generated for the autoregressive model through self-supervision contrast learning training, Representing dimensions for account features, W is a weight matrix of the graph neural network, For the number of network layers.
9. The method according to claim 1, wherein the implementation manner of the autoregressive model comprises LSTM, GRU, TCN or a Transformer sequence model, and the implementation manner of the machine learning classification model comprises a bayesian classification method, a decision tree, a logistic regression, a support vector machine, and a neural network model.
10. The method for analysis and prediction of financial behavior pattern based on time chart characterization learning according to claim 9, wherein the account is obtained by a machine learning classification model implemented by a logistic regression model The probability of being abnormal is: Wherein, the The representation vector of the account u, q is the weight matrix of the logistic regression model; The loss function of the logistic regression model is: Wherein the method comprises the steps of For account Corresponding real labels.

Description

Financial behavior pattern analysis and prediction method based on time sequence diagram representation learning Technical Field The invention relates to the technical field of financial transaction safety, in particular to a financial behavior pattern analysis and prediction method based on time sequence diagram representation learning. Background With the development of information technology in the financial field, financial transaction behaviors are recorded, so that massive business flows are formed. The business flow data contains rich information, and is the data which is indispensable for improving financial services and preventing financial risks. The historical behavior of the account is modeled through business flow, so that analysis and prediction of financial behavior patterns are realized, and the method is applied to the fields of anti-fraud, risk identification and prevention, abnormal behavior pattern detection and the like. Traditional feature extraction only carries out simple change on original data, and cannot fully capture more complex and abstract concepts in information. In an actual application scene, the expert analysis is used for summarizing service characteristics, so that new behavior models in transaction flow data cannot be automatically and timely captured, and certain hysteresis and limitation exist. CN 111797177A discloses a method for classifying financial time series for detecting abnormal financial account and application, the method can construct and expand financial time series data set of financial account from transaction flow data of abnormal financial account and normal financial account, using neural network model of stacking multiple blocks (each Block contains LocalBiLSTM, self-Attention, residual connection, layerNormalization, position-wise Feed-ForwardNetworks) to extract local and global mode characteristics of sequence from financial time series, finally using softmax classifying layer to classify financial time series, finally realizing detecting function of abnormal financial account. However, the feature representation capability of the method on transaction flow data is still insufficient, and the method does not extract and utilize account interaction information in the transaction flow, so that the model prediction effect is limited to a certain extent. Disclosure of Invention The invention aims to provide a financial behavior pattern analysis and prediction method based on time sequence diagram representation learning, which is independent of human feature engineering, can well capture time sequence, interaction and business information in transaction flow and can improve the prediction effect. The aim of the invention can be achieved by the following technical scheme: A financial behavior pattern analysis and prediction method based on time sequence diagram representation learning comprises the following steps: Acquiring transaction flow data; the business characteristics and the time sequence characteristics of each transaction stream data are respectively input into an event encoder and a time encoder to obtain a stream embedded vector; learning the running water embedded vector based on the autoregressive model, and obtaining a characterization matrix through self-supervision contrast learning training; taking accounts and transaction opponents in the stream data as nodes, and taking transaction stream as a continuous edge to construct an interactive information graph; Modeling node interaction on the graph structural features by using a graph neural network to form a graph enhancement module, wherein the graph enhancement module can enhance the representation capability of the representation matrix on account interaction information to obtain a weight matrix of the graph neural network; constructing a machine learning classification model based on the enhanced characterization matrix and the account label to obtain a weight matrix of the machine learning classification model; And updating a weight matrix of the machine learning classification model and a weight matrix of the graph neural network through back propagation, and training the model to realize analysis and prediction of the financial behavior mode. The transaction pipeline data is expressed as: Su＝(eu1,eu2,…,eun) Wherein S u represents transaction flow data of account u, n is transaction number, e ui is ith transaction record, e ui＝(u,vui,tui,attrui) is a quadruple, v ui represents transaction opponents, t ui represents time stamp, attr ui represents attribute. The event encoder encodes attribute information in a transaction event of account u: The num i is all the numerical type features in the ith transaction record, the cat ij is the j type features in the ith transaction record, and the W and W j are feature transformation matrices for mapping data into d-dimensional vectors, and the Emb (·) represents embedding mapping of the type attribute. The time encoder encodes the time sta