Training method, system, computer equipment and medium for income assessment model

Abstract

The present application relates to a training method, system, computer equipment and medium for an income assessment model, the method comprising obtaining a user's structured financial data, time series behavior data and unstructured text data, and constructing semantic and logical associations between data based on a graph neural network to generate a multi-dimensional feature graph; based on a Transformer and a long short-term memory network, trend analysis is performed on the time series behavior data and the multi-dimensional feature graph, and feature extraction is performed on the unstructured text data in combination with a preset language model to obtain basic feature data; based on the key financial indicators and basic feature data of the structured financial data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain the corresponding high-dimensional feature representation; using a reinforcement learning algorithm, and iteratively training and parameter adjustment of the preset model based on the high-dimensional feature representation to obtain an income assessment model. The present application has the effect of improving the accuracy of income assessment.

Description

Technical Field

[0001] The present application relates to the technical field of income assessment, and in particular to a training method, system, computer device and medium for an income assessment model. Background Art

[0002] At present, income assessment technology is widely used in the financial field, especially in credit approval, risk control and customer management. Income assessment models are usually based on users' financial data, behavioral data and other relevant information to predict their repayment ability or credit risk.

[0003] Existing,income assessment methods mostly rely on simple statistical models and use,structured financial data for prediction.,They lack the effective use of unstructured data such as social,media data and consumer behavior data, and do not consider the dynamic changes of,time series data, resulting in low prediction accuracy and lack of strong generalization,ability.

[0004] The above-mentioned existing technical solutions have the following defects: the existing income assessment methods have a low consideration dimension for data, resulting in inaccurate prediction accuracy, so there is room for improvement. Summary of the invention

[0005] In order to improve the accuracy of income assessment, the present application provides a training method, system, computer device and medium for an income assessment model.

[0006] The above-mentioned invention objective of the present application is achieved through the following technical solutions: A method for training an income assessment model, the method comprising: Acquire the user's structured financial data, time series behavior data, and unstructured text data, and construct semantic and logical associations among the structured financial data, the time series behavior data, and the unstructured text data based on a graph neural network to generate a multi-dimensional feature graph; Based on Transformer and long short-term memory network, trend analysis is performed on the time series behavior data and the multi-dimensional feature graph, and feature extraction is performed on the unstructured text data in combination with a preset language model to obtain basic feature data; According to the key financial indicators of the structured financial data and the basic feature data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain corresponding high-dimensional feature representation; By using a reinforcement learning algorithm and performing iterative training and parameter adjustment on a preset model according to the high-dimensional feature representation, an income assessment model is obtained.

[0007] By adopting the above technical solution, by acquiring the user's structured financial data, time series behavior data and unstructured text data, and constructing semantic and logical associations between data based on graph neural networks, a multi-dimensional feature graph is generated, which can effectively integrate different types of data and help the system accurately capture the user's multi-dimensional features, thereby improving the accuracy of the model's assessment of user income; by performing trend analysis on time series behavior data and multi-dimensional feature graphs based on Transformer and long short-term memory networks, and extracting features from unstructured text data in combination with a preset language model, the user's dynamic behavior and consumption habits can be more comprehensively understood, thereby making income assessment more accurate and flexible; by performing cross-modal feature fusion based on key financial indicators and basic feature data of structured financial data, the advantages of various types of data can be effectively combined to generate high-dimensional feature representations, further improving the accuracy and generalization ability of the income assessment model; by using reinforcement learning algorithms and iteratively training and adjusting parameters of the preset model based on high-dimensional feature representations, the performance of the model can be optimized based on feedback, thereby improving the accuracy and adaptability of income assessment.

[0008] In one example, the present application may be further configured as follows: constructing the semantic and logical associations among the structured financial data, the time series behavior data, and the unstructured text data based on the graph neural network to generate a multi-dimensional feature graph, specifically including: The key financial indicators in the structured financial data are represented as node features, and node weights are generated based on the time series behavior data, and corresponding feature vectors are generated in combination with the unstructured text data, and then graph nodes are constructed according to the node features, the node weights and the feature vectors; Generate edge weights between the graph nodes and establish semantic and logical associations between the nodes through statistical analysis and causal inference methods; Graph neural networks are used to embed nodes and edges, and neighborhood node features are aggregated through a message passing mechanism to capture high-order semantic relationships between nodes. The embedded features are dynamically updated based on a time-aware mechanism to generate the time series-aware multi-dimensional feature graph.

[0009] By adopting the above technical solution, by representing the key financial indicators in structured financial data as node features, generating node weights based on time series behavior data, and combining unstructured text data to generate feature vectors, different types of data can be effectively integrated into graph nodes, helping the model to more accurately understand and analyze the user's comprehensive financial status and behavioral characteristics, thereby improving the model's ability to assess income.

[0010] In one example, the present application may be further configured as follows: the key financial indicators in the structured financial data are represented as node features, and node weights are generated based on the time series behavior data, and corresponding feature vectors are generated in combination with the unstructured text data, and graph nodes are constructed according to the node features, the node weights and the feature vectors, specifically including: Extracting key financial indicators such as salary, bonus, debt ratio, etc. of the user from the structured financial data, and establishing an initial feature vector of the node according to the data dimension; Performing time characteristic analysis on the time series behavior data, and dynamically generating a weight factor of a node based on the time characteristic to adjust initial node characteristics; Extracting features from the unstructured text data to generate a text embedding vector, and concatenating the vector with the initial feature vector and the weight factor of the structured financial data to form a fused node feature representation; The graph nodes are defined and constructed according to the fused node feature representations.

[0011] By adopting the above technical solutions, by extracting key financial indicators such as users' wages, bonuses, debt ratios, etc. from structured financial data, it is possible to provide accurate initial feature representation for each node, thereby enhancing the graph neural network's ability to understand data; by performing time characteristic analysis on time series behavior data and dynamically generating node weight factors, it is possible to accurately capture the dynamic changes in user behavior and improve the adaptability of the model; by extracting features from unstructured text data and concatenating it with financial data and weight factors, it is possible to enhance the multi-dimensional information of node features, thereby improving the model's revenue prediction accuracy.

[0012] In one example, the present application may be further configured as follows: based on the Transformer and the long short-term memory network, the time series behavior data and the multi-dimensional feature graph are subjected to trend analysis, and the unstructured text data is subjected to feature extraction in combination with a preset language model to obtain basic feature data, specifically including: The Transformer is used to encode the time series behavior data, and the long-term dependencies between different time points in the time series are captured through a multi-head attention mechanism; In combination with the long short-term memory network, the periodic fluctuation characteristics in the time series behavior data are modeled to extract the dynamic characteristics of user income and consumption behavior; Using the preset language model to perform feature extraction on the unstructured text data, converting the text into a vector representation through an embedding layer, and extracting corresponding key semantic features using a contextual attention mechanism; The trend features of the time series behavior data and the key semantic features of the unstructured text data are preliminarily integrated to generate the basic feature data.

[0013] By adopting the above technical solutions, by using Transformer to encode time series behavior data and capturing long-term dependencies through a multi-head attention mechanism, it is possible to effectively analyze the long-term trends of user income and consumption, thereby improving the accuracy of the model in dynamic income assessment; by combining the long short-term memory network to model the periodic fluctuation characteristics, it is possible to better identify the cyclical changes in user income and consumption, thereby accurately assessing their future income performance; by using a preset language model to extract features from unstructured text data, it is possible to extract key semantic information and enhance the model's understanding of user preferences and behaviors; by preliminarily fusing the trend features of time series behavior data with the semantic features of unstructured text data, it is possible to integrate multiple information sources to obtain more accurate basic feature data and improve the overall performance of income assessment.

[0014] In one example, the present application may be further configured as follows: performing feature fusion mapping based on the key financial indicators of the structured financial data and the basic feature data through cross-modal feature fusion technology to obtain corresponding high-dimensional feature representation, specifically including: Representing the key financial indicators in the structured financial data as numerical vectors of fixed dimensions, and performing standardization processing on the time series trend features and text semantic features in the basic feature data; The numerical vector and the standardized basic feature data are embedded into a unified feature space through a cross-modal feature fusion network, and the weights of different modal features are dynamically adjusted using an attention mechanism according to their contribution to income assessment, thereby obtaining corresponding fusion features; The fused features are further dimensionally upgraded through a nonlinear mapping method to generate the high-dimensional feature representation.

[0015] By adopting the above technical solution, by representing the key financial indicators in the structured financial data as numerical vectors of fixed dimensions and standardizing the time series trend features and text semantic features in the basic feature data, it is possible to ensure that the features of different data sources are processed under a unified standard, thereby improving the effectiveness of feature fusion; by embedding the numerical vectors and the standardized feature data into a unified feature space through a cross-modal feature fusion network, and dynamically adjusting the weights of different modalities using the attention mechanism, the feature fusion process can be refined according to the contribution of each modal feature to the income assessment, thereby enhancing the prediction accuracy of the model; by increasing the dimension of the fused features through the nonlinear mapping method, more nonlinear features can be captured, and the expressive power of the high-dimensional feature representation can be improved, thereby improving the evaluation effect of the model.

[0016] In one example, the present application may be further configured as follows: the reinforcement learning algorithm is used, and the preset model is iteratively trained and parameter adjusted according to the high-dimensional feature representation to obtain an income assessment model, specifically including: By means of the proxy mechanism of the reinforcement learning algorithm, the contribution of the high-dimensional feature representation to the model output result is analyzed to obtain weight values ​​of different features, and the model parameters are dynamically adjusted according to the weight values; By utilizing a multi-round training optimization mechanism, the model weights and hyperparameters are iteratively updated according to the reward function feedback in each round of training, thereby completing the global parameter adjustment and generating the income evaluation model.

[0017] By adopting the above technical solution, through the proxy mechanism of the reinforcement learning algorithm, the contribution of high-dimensional feature representation to the model output results can be analyzed, and the model parameters can be automatically adjusted to optimize the model performance and improve the model's adaptability to new data; through the multi-round training optimization mechanism, the model's weights and hyperparameters can be gradually updated according to the reward function feedback in each round of training, thereby completing the optimization of global parameters, improving the accuracy and stability of the model, and ultimately generating a high-performance income assessment model to enhance the long-term reliability of the model.

[0018] The second object of the invention is achieved by the following technical solutions: A training system for an income assessment model, characterized in that the training system for the income assessment model comprises: A data collection module is used to obtain the user's structured financial data, time series behavior data and unstructured text data, and to construct semantic and logical associations among the structured financial data, the time series behavior data and the unstructured text data based on a graph neural network to generate a multi-dimensional feature graph; A trend analysis module, which is used to perform trend analysis on the time series behavior data and the multi-dimensional feature graph based on Transformer and long short-term memory network, and extract features from the unstructured text data in combination with a preset language model to obtain basic feature data; A feature fusion module, used to perform feature fusion mapping based on the key financial indicators of the structured financial data and the basic feature data through cross-modal feature fusion technology, thereby obtaining a corresponding high-dimensional feature representation; The dynamic optimization module is used to utilize a reinforcement learning algorithm and iteratively train and adjust parameters of a preset model according to the high-dimensional feature representation to obtain an income evaluation model.

[0019] By adopting the above technical solution, by acquiring the user's structured financial data, time series behavior data and unstructured text data, and constructing semantic and logical associations between data based on graph neural networks, a multi-dimensional feature graph is generated, which can effectively integrate different types of data and help the system accurately capture the user's multi-dimensional features, thereby improving the accuracy of the model's assessment of user income; by performing trend analysis on time series behavior data and multi-dimensional feature graphs based on Transformer and long short-term memory networks, and extracting features from unstructured text data in combination with a preset language model, the user's dynamic behavior and consumption habits can be more comprehensively understood, thereby making income assessment more accurate and flexible; by performing cross-modal feature fusion based on key financial indicators and basic feature data of structured financial data, the advantages of various types of data can be effectively combined to generate high-dimensional feature representations, further improving the accuracy and generalization ability of the income assessment model; by using reinforcement learning algorithms and iteratively training and adjusting parameters of the preset model based on high-dimensional feature representations, the performance of the model can be optimized based on feedback, thereby improving the accuracy and adaptability of income assessment.

[0020] The third objective of the present application is achieved through the following technical solutions: A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned training method for the income assessment model when executing the computer program.

[0021] The fourth objective of the present application is achieved through the following technical solutions: A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned income assessment model training method.

[0022] In summary, this application includes the following beneficial technical effects: 1. By acquiring the user's structured financial data, time series behavior data, and unstructured text data, and building semantic and logical associations between data based on graph neural networks, and generating multi-dimensional feature graphs, different types of data can be effectively integrated to help the system accurately capture the user's multi-dimensional features, thereby improving the accuracy of the model's assessment of user income; by performing trend analysis on time series behavior data and multi-dimensional feature graphs based on Transformer and long short-term memory networks, and extracting features from unstructured text data in combination with a preset language model, users' dynamic behaviors and consumption habits can be more comprehensively understood, making income assessment more accurate and flexible; by performing cross-modal feature fusion based on key financial indicators and basic feature data of structured financial data, the advantages of various types of data can be effectively combined to generate high-dimensional feature representations, further improving the accuracy and generalization ability of the income assessment model; by using reinforcement learning algorithms and iteratively training and adjusting parameters of the preset model based on high-dimensional feature representations, the performance of the model can be optimized based on feedback, thereby improving the accuracy and adaptability of income assessment; 2. By representing key financial indicators in structured financial data as node features, generating node weights based on time series behavioral data, and combining unstructured text data to generate feature vectors, different types of data can be effectively integrated into graph nodes, helping the model to more accurately understand and analyze the user's comprehensive financial status and behavioral characteristics, thereby improving the model's ability to assess income; 3. By extracting key financial indicators such as users' wages, bonuses, debt ratios, etc. from structured financial data, it can provide accurate initial feature representation for each node, thereby enhancing the graph neural network's ability to understand data; by analyzing the time characteristics of time series behavior data and dynamically generating node weight factors, it can accurately capture the dynamic changes of user behavior and improve the adaptability of the model; by extracting features from unstructured text data and splicing it with financial data and weight factors, it can enhance the multi-dimensional information of node features, thereby improving the model's revenue prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of a method for training an income assessment model in one embodiment of the present application; Figure 2 It is a flowchart for implementing step S10 in the training method of the income assessment model in one embodiment of the present application; Figure 3 It is a flowchart for implementing step S11 in the training method of the income assessment model in one embodiment of the present application; Figure 4 is a flowchart for implementing step S20 in the training method of the income assessment model in one embodiment of the present application; Figure 5is a flowchart for implementing step S30 in the training method of the income assessment model in one embodiment of the present application; Figure 6 is a flowchart for implementing step S40 in the training method of the income assessment model in one embodiment of the present application; Figure 7 It is a principle block diagram of a training system for an income assessment model in one embodiment of the present application; Figure 8 It is a schematic diagram of a device in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The present application is further described in detail below in conjunction with the accompanying drawings.

[0025] In one embodiment, if Figure 1 As shown, the present application discloses a training method for an income assessment model, which specifically includes the following steps: S10: Obtain the user's structured financial data, time series behavior data, and unstructured text data, and build semantic and logical associations between the structured financial data, time series behavior data, and unstructured text data based on a graph neural network to generate a multi-dimensional feature graph.

[0026] Specifically, by connecting to the API interfaces of banks, tax platforms and other platforms, users' financial data, including structured data such as income, expenditure, and liabilities, are automatically collected, and crawler technology is used to obtain unstructured text data from sources such as social media and transaction records. Graph neural networks are then used to represent these data as graph nodes, and edge weights are generated based on the semantic relationships and logical associations between the data. A multi-dimensional feature graph is constructed to represent the correlation between different types of data.

[0027] S20: Based on Transformer and long short-term memory networks, trend analysis is performed on time series behavior data and multi-dimensional feature graphs, and features are extracted from unstructured text data in combination with a preset language model to obtain basic feature data.

[0028] Specifically, the Transformer model is used to encode time series behavior data, and the long-term dependencies between different time points are captured through the multi-head attention mechanism. The long short-term memory network is used to model the periodic fluctuation characteristics in the data and extract the user's income and consumption behavior trends. At the same time, the preset language model is used to extract features from unstructured text data and convert it into vector representation. The key information is extracted through the contextual attention mechanism to generate the final basic feature data.

[0029] S30: Based on the key financial indicators and basic feature data of the structured financial data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain the corresponding high-dimensional feature representation.

[0030] Specifically, key financial indicators in structured financial data, such as salary, bonus, debt ratio, etc., are represented as numerical vectors, and the basic feature data extracted from time series behavioral data and unstructured text data are standardized. Subsequently, these data of different modalities are fused into a unified feature space through a cross-modal feature fusion network, and the feature weights of different modalities are dynamically adjusted according to their contribution to income assessment using the attention mechanism, ultimately generating a high-dimensional feature representation.

[0031] S40: Using the reinforcement learning algorithm, iteratively train and adjust the parameters of the preset model according to the high-dimensional feature representation to obtain the income assessment model.

[0032] Specifically, the agent mechanism of the reinforcement learning algorithm is utilized to represent the contribution of the model output results according to the high-dimensional feature representation analysis, the weight value of each feature is calculated, and the model parameters are dynamically adjusted according to the weight value. Subsequently, a multi-round training optimization mechanism is adopted. In each round of training, the model weights and hyperparameters are iteratively updated according to the reward function feedback to improve the prediction accuracy and generalization ability of the model. After multiple rounds of optimization, an income assessment model with high stability and predictive ability is generated.

[0033] In one embodiment, if Figure 2 As shown, in step S10, the semantic and logical associations between structured financial data, time series behavior data and unstructured text data are constructed based on the graph neural network to generate a multi-dimensional feature graph, which specifically includes: S11: Represent the key financial indicators in the structured financial data as node features, generate node weights based on time series behavior data, and generate corresponding feature vectors in combination with unstructured text data. Then, construct graph nodes based on node features, node weights and feature vectors.

[0034] Specifically, key financial indicators such as users' wages, bonuses, and debt ratios are extracted from structured financial data, the initial feature vectors of nodes are established according to the data dimensions, and time characteristic analysis is performed based on time series behavior data to generate node weight factors to adjust the initial node features. Next, feature extraction is performed on unstructured text data to generate text embedding vectors, which are then concatenated with the initial feature vectors and weight factors of structured financial data to form a fused node feature representation, and finally graph nodes are constructed.

[0035] S12: Generate edge weights between graph nodes through statistical analysis methods and causal inference methods, and establish semantic and logical association relationships between nodes.

[0036] Specifically, statistical analysis methods, such as the Pearson correlation coefficient and cosine similarity, are used to evaluate the strength of association between different nodes and generate edge weights to represent the semantic relationship between nodes. In addition, causal inference methods are applied to infer the causal relationship between nodes based on the data characteristics and external variables between nodes, further deepen the understanding of the logical association between nodes, and generate an edge weight matrix for graph neural network input to ensure that the nodes and edges in the graph can correctly reflect the true semantics and logical connections.

[0037] S13: Use graph neural networks to embed nodes and edges, aggregate neighborhood node features through a message passing mechanism, and capture high-order semantic relationships between nodes. The embedded features are dynamically updated based on a time-aware mechanism to generate a time-series-aware multi-dimensional feature graph.

[0038] Specifically, graph neural networks are used to embed features of nodes and edges, and the feature information of neighboring nodes is aggregated through a message passing mechanism, so that each node can obtain the information of its neighboring nodes, thereby improving the semantic understanding between nodes in the graph; then, combined with the time-aware mechanism, the embedded features of the nodes are updated in real time according to the dynamics of time series data, so that the features can capture changes in the time dimension, and generate a multi-dimensional feature graph that can reflect the characteristics of the time series and contain high-order semantic relationships between nodes, which is used for subsequent model training.

[0039] In one embodiment, if Figure 3 As shown, in step S11, the key financial indicators in the structured financial data are represented as node features, and the node weights are generated based on the time series behavior data, and the corresponding feature vectors are generated in combination with the unstructured text data. The graph nodes are constructed according to the node features, node weights and feature vectors, which specifically include: S111: Extract key financial indicators such as user's salary, bonus, debt ratio, etc. from structured financial data, and establish initial feature vectors of nodes according to data dimensions.

[0040] Specifically, key financial indicators such as the user's salary, bonus, debt ratio, etc. are selected, and these indicators are standardized and converted into numerical vectors of fixed dimensions as the initial feature vectors of the node; according to different types of financial data, each data item is encoded as a feature vector and used as the input of subsequent graph neural network nodes.

[0041] S112: Perform time characteristic analysis on the time series behavior data, and dynamically generate weight factors of nodes based on the time characteristics to adjust initial node features.

[0042] Specifically, the time characteristics such as periodic changes and trend fluctuations in time series behavior data are analyzed, and node weight factors related to these characteristics are generated. These weight factors are used to dynamically adjust the characteristics of the nodes to ensure that the characteristics of the data in different time periods can accurately reflect the user's behavior changes.

[0043] S113: Extract features from unstructured text data to generate text embedding vectors, and concatenate them with the initial feature vectors and weight factors of structured financial data to form a fused node feature representation.

[0044] Specifically, natural language processing technology is used to extract features from unstructured text data and convert the text data into an embedding vector. The embedding vector is concatenated with the initial feature vector extracted from structured financial data and the dynamically adjusted weight factor to form a fused node feature representation, providing richer input data for subsequent graph neural network models.

[0045] S114: Define and construct graph nodes according to the fused node feature representation.

[0046] Specifically, the fused node feature representation is used as the graph node input to create a graph structure, and these nodes are connected with other nodes in the graph through edges to form a complete graph structure. The features of the graph nodes and edges are used for subsequent graph neural network processing.

[0047] In one embodiment, if Figure 4 As shown, in step S20, based on Transformer and long short-term memory network, trend analysis is performed on time series behavior data and multi-dimensional feature graphs, and features are extracted from unstructured text data in combination with a preset language model to obtain basic feature data, specifically including: S21: Use Transformer to encode time series behavior data and capture long-term dependencies between different time points in the time series through a multi-head attention mechanism.

[0048] Specifically, the Transformer model is used to encode time series behavior data, and the long-term dependencies between different time points are analyzed through the multi-head attention mechanism to capture, for example, the volatility and cyclical changes of user income, ensuring that the key patterns in the time series data are fully mined, thereby providing accurate trend data for subsequent analysis.

[0049] S22: Combined with long short-term memory networks, the periodic fluctuation characteristics in time series behavior data are modeled to extract the dynamic characteristics of user income and consumption behavior.

[0050] Specifically, the long short-term memory network is used to model time series data, especially the periodic fluctuation characteristics, such as the seasonal fluctuations of user income and the periodic characteristics of consumption behavior. The LSTM model is used to capture the long-term dependencies in the time series and extract the dynamic behavior characteristics of users, providing in-depth dynamic analysis for evaluating income capacity.

[0051] S23: Use the preset language model to extract features from unstructured text data, convert the text into vector representation through the embedding layer, and use the contextual attention mechanism to extract the corresponding key semantic features.

[0052] Specifically, a preset language model such as BERT is used to process unstructured text data. The text is first converted into a vector representation through an embedding layer, and the contextual attention mechanism is further used to extract key semantics in the text, such as the user's consumption preferences and living habits, to generate corresponding semantic features to support income assessment.

[0053] S24: Preliminarily integrate the trend features of the time series behavior data and the key semantic features of the unstructured text data to generate basic feature data.

[0054] Specifically, the trend features of time series data processed by Transformer and LSTM are fused with the key semantic features of text data extracted by the language model, and a unified feature vector is generated by splicing or weighting to form basic feature data, providing a basis for subsequent feature fusion and model training.

[0055] In one embodiment, if Figure 5 As shown, in step S30, feature fusion mapping is performed based on the key financial indicators and basic feature data of the structured financial data through cross-modal feature fusion technology to obtain the corresponding high-dimensional feature representation, which specifically includes: S31: Represent the key financial indicators in the structured financial data as numerical vectors of fixed dimensions, and standardize the time series trend features and text semantic features in the basic feature data.

[0056] Specifically, the key financial indicators in the structured financial data, such as salary, bonus, debt ratio, etc., are converted into numerical vectors of fixed dimensions through preset numerical mapping, and the time series behavioral data and text semantic features are standardized to ensure that various features have the same scale, which facilitates subsequent cross-modal fusion and processing.

[0057] S32: Through the cross-modal feature fusion network, the numerical vector and the standardized basic feature data are embedded into a unified feature space, and the weights of different modal features are dynamically adjusted using the attention mechanism according to their contribution to income assessment, thereby obtaining the corresponding fusion features.

[0058] Specifically, a cross-modal feature fusion network is used to fuse the numerical vector and the standardized feature data, and the contribution of data from different modalities is weighted through the attention mechanism, so that the features that have a greater impact on income assessment receive more attention, thereby generating a fused feature representation.

[0059] S33: Through the nonlinear mapping method, the fused features are further upgraded to generate high-dimensional feature representation.

[0060] Specifically, nonlinear mapping methods, such as multi-layer perceptron (MLP), are used to further upgrade the fusion features to generate high-dimensional feature representations suitable for subsequent model training, thereby improving the model's expressive power and its ability to fit complex patterns.

[0061] In one embodiment, if Figure 6 As shown, in step S40, the preset model is iteratively trained and parameter adjusted using a reinforcement learning algorithm according to the high-dimensional feature representation to obtain an income evaluation model, which specifically includes: S41: Through the proxy mechanism of the reinforcement learning algorithm, the contribution of high-dimensional feature representation to the model output results is analyzed, the weight values ​​of different features are obtained, and the model parameters are dynamically adjusted according to the weight values.

[0062] Specifically, the reinforcement learning algorithm uses an agent mechanism to analyze the contribution of high-dimensional feature representation to the model output results, calculate the weight values ​​of different features, and dynamically adjust the model parameters based on these weight values, so that the model can effectively learn the optimal parameter configuration and improve prediction accuracy.

[0063] S42: Utilize a multi-round training optimization mechanism to iteratively update the model weights and hyperparameters based on the reward function feedback in each round of training, and then complete the global parameter adjustment to generate an income evaluation model.

[0064] Specifically, a multi-round training mechanism in reinforcement learning is adopted. In each round of training, the weights and hyperparameters of the model are adjusted according to the feedback information of the reward function. Through the iterative optimization process, the performance of the model is gradually improved, and finally a stable income assessment model with high predictive ability is generated.

[0065] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0066] In one embodiment, a training system for an income assessment model is provided, and the training system for the income assessment model corresponds one-to-one to the training method for the income assessment model in the above embodiment. Figure 7 As shown, the training system of the income assessment model includes a data acquisition module, a trend analysis module, a feature fusion module, and a dynamic optimization module. The detailed description of each functional module is as follows: The data collection module is used to obtain the user's structured financial data, time series behavior data, and unstructured text data, and build semantic and logical associations between structured financial data, time series behavior data, and unstructured text data based on the graph neural network to generate a multi-dimensional feature graph; The trend analysis module is used to perform trend analysis on time series behavior data and multi-dimensional feature graphs based on Transformer and long short-term memory networks, and extract features from unstructured text data in combination with a preset language model to obtain basic feature data; The feature fusion module is used to perform feature fusion mapping based on the key financial indicators and basic feature data of the structured financial data through cross-modal feature fusion technology, and then obtain the corresponding high-dimensional feature representation; The dynamic optimization module is used to utilize the reinforcement learning algorithm and iteratively train and adjust the parameters of the preset model according to the high-dimensional feature representation to obtain the income evaluation model.

[0067] Optionally, the data acquisition module specifically includes: The node construction submodule is used to represent the key financial indicators in the structured financial data as node features, generate node weights based on the time series behavior data, generate corresponding feature vectors in combination with the unstructured text data, and then construct graph nodes based on the node features, node weights and feature vectors; The association relationship generation submodule is used to generate edge weights between graph nodes and establish semantic association relationships and logical association relationships between nodes through statistical analysis methods and causal inference methods; The graph embedding submodule is used to embed nodes and edges using graph neural networks, aggregate neighborhood node features through a message passing mechanism, capture high-order semantic relationships between nodes, and dynamically update embedded features based on a time-aware mechanism to generate a time-series-aware multi-dimensional feature graph.

[0068] Optionally, the node construction submodule specifically includes: The node initialization unit is used to extract key financial indicators such as user's salary, bonus, debt ratio, etc. from structured financial data, and establish the initial feature vector of the node according to the data dimension; A time characteristic analysis unit is used to perform time characteristic analysis on time series behavior data and dynamically generate a weight factor of a node based on the time characteristic to adjust the initial node characteristics; The text feature fusion unit is used to extract features from unstructured text data, generate text embedding vectors, and concatenate them with the initial feature vectors and weight factors of structured financial data to form a fused node feature representation; The graph node construction unit is used to define and construct graph nodes according to the fused node feature representation.

[0069] Optionally, the trend analysis module specifically includes: The time series encoding submodule is used to encode time series behavior data using Transformer and capture the long-term dependencies between different time points in the time series through a multi-head attention mechanism; The periodic modeling submodule is used to combine the long short-term memory network to model the periodic fluctuation characteristics in the time series behavior data and extract the dynamic characteristics of user income and consumption behavior; The text semantic extraction submodule is used to extract features from unstructured text data using a preset language model, convert the text into a vector representation through an embedding layer, and extract the corresponding key semantic features using a contextual attention mechanism; The feature fusion submodule is used to preliminarily fuse the trend features of time series behavior data and the key semantic features of unstructured text data to generate basic feature data.

[0070] Optionally, the feature fusion module specifically includes: The data standardization submodule is used to represent the key financial indicators in the structured financial data as numerical vectors of fixed dimensions, and to standardize the time series trend features and text semantic features in the basic feature data; The cross-modal fusion submodule is used to embed the numerical vector and the standardized basic feature data into a unified feature space through a cross-modal feature fusion network, and dynamically adjust the weights of different modal features using the attention mechanism according to their contribution to income assessment, thereby obtaining the corresponding fusion features; The feature dimension upgrading submodule is used to further upgrade the fusion features through nonlinear mapping methods to generate high-dimensional feature representations.

[0071] Optionally, the dynamic optimization module specifically includes: The reinforcement learning agent submodule is used to analyze the contribution of high-dimensional feature representation to the model output results through the agent mechanism of the reinforcement learning algorithm, obtain the weight values ​​of different features, and dynamically adjust the model parameters according to the weight values; The training optimization submodule is used to utilize a multi-round training optimization mechanism to iteratively update the model weights and hyperparameters according to the reward function feedback in each round of training, thereby completing the global parameter adjustment and generating a revenue evaluation model.

[0072] The specific definition of the training system of the income assessment model can be found in the definition of the training method of the income assessment model above, which will not be repeated here. Each module in the above-mentioned training system of the income assessment model can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0073] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a training method for an income assessment model is implemented.

[0074] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program: Obtain the user's structured financial data, time series behavior data, and unstructured text data, and build semantic and logical associations between structured financial data, time series behavior data, and unstructured text data based on graph neural networks to generate multi-dimensional feature graphs; Based on Transformer and long short-term memory networks, trend analysis is performed on time series behavior data and multi-dimensional feature graphs, and features are extracted from unstructured text data in combination with a preset language model to obtain basic feature data; According to the key financial indicators and basic feature data of structured financial data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain the corresponding high-dimensional feature representation; By using the reinforcement learning algorithm and iteratively training and adjusting the parameters of the preset model according to the high-dimensional feature representation, an income assessment model is obtained.

[0075] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented: Obtain the user's structured financial data, time series behavior data, and unstructured text data, and build semantic and logical associations between structured financial data, time series behavior data, and unstructured text data based on graph neural networks to generate multi-dimensional feature graphs; Based on Transformer and long short-term memory networks, trend analysis is performed on time series behavior data and multi-dimensional feature graphs, and features are extracted from unstructured text data in combination with a preset language model to obtain basic feature data; According to the key financial indicators and basic feature data of structured financial data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain the corresponding high-dimensional feature representation; By using the reinforcement learning algorithm and iteratively training and adjusting the parameters of the preset model according to the high-dimensional feature representation, an income assessment model is obtained.

[0076] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0077] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0078] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A training method for an income assessment model, characterized in that: The training method of the income assessment model includes: Acquire the user's structured financial data, time series behavior data, and unstructured text data, and construct semantic and logical associations among the structured financial data, the time series behavior data, and the unstructured text data based on a graph neural network to generate a multi-dimensional feature graph; Based on Transformer and long short-term memory network, trend analysis is performed on the time series behavior data and the multi-dimensional feature graph, and feature extraction is performed on the unstructured text data in combination with a preset language model to obtain basic feature data; According to the key financial indicators of the structured financial data and the basic feature data, feature fusion mapping is performed through cross-modal feature fusion technology to obtain corresponding high-dimensional feature representation; By using a reinforcement learning algorithm and performing iterative training and parameter adjustment on a preset model according to the high-dimensional feature representation, an income assessment model is obtained.

2. The training method of the income assessment model according to claim 1, characterized in that: The constructing of semantic and logical associations among the structured financial data, the time series behavior data and the unstructured text data based on the graph neural network to generate a multi-dimensional feature graph specifically includes: The key financial indicators in the structured financial data are represented as node features, and node weights are generated based on the time series behavior data, and corresponding feature vectors are generated in combination with the unstructured text data, and then graph nodes are constructed according to the node features, the node weights and the feature vectors; Generate edge weights between the graph nodes and establish semantic and logical associations between the nodes through statistical analysis and causal inference methods; Graph neural networks are used to embed nodes and edges, and neighborhood node features are aggregated through a message passing mechanism to capture high-order semantic relationships between nodes. The embedded features are dynamically updated based on a time-aware mechanism to generate the time series-aware multi-dimensional feature graph.

3. The training method of the income assessment model according to claim 2, characterized in that: The step of representing the key financial indicators in the structured financial data as node features, generating node weights based on the time series behavior data, generating corresponding feature vectors in combination with the unstructured text data, and constructing graph nodes according to the node features, the node weights and the feature vectors specifically includes: Extracting key financial indicators such as salary, bonus, debt ratio, etc. of the user from the structured financial data, and establishing an initial feature vector of the node according to the data dimension; Performing time characteristic analysis on the time series behavior data, and dynamically generating a weight factor of a node based on the time characteristic to adjust initial node characteristics; Extracting features from the unstructured text data to generate a text embedding vector, and concatenating the vector with the initial feature vector and the weight factor of the structured financial data to form a fused node feature representation; The graph nodes are defined and constructed according to the fused node feature representations.

4. The method for training an income assessment model according to claim 1, characterized in that: The Transformer and LSTM network are used to perform trend analysis on the time series behavior data and the multi-dimensional feature graph, and feature extraction is performed on the unstructured text data in combination with a preset language model to obtain basic feature data, specifically including: The Transformer is used to encode the time series behavior data, and the long-term dependencies between different time points in the time series are captured through a multi-head attention mechanism; In combination with the long short-term memory network, the periodic fluctuation characteristics in the time series behavior data are modeled to extract the dynamic characteristics of user income and consumption behavior; Using the preset language model to perform feature extraction on the unstructured text data, converting the text into a vector representation through an embedding layer, and extracting corresponding key semantic features using a contextual attention mechanism; The trend features of the time series behavior data and the key semantic features of the unstructured text data are preliminarily integrated to generate the basic feature data.

5. The method for training an income assessment model according to claim 1, characterized in that: The key financial indicators of the structured financial data and the basic feature data are used to perform feature fusion mapping through cross-modal feature fusion technology to obtain corresponding high-dimensional feature representation, specifically including: The key financial indicators in the structured financial data are represented as numerical vectors of fixed dimensions, and the time series trend features and text semantic features in the basic feature data are standardized; The numerical vector and the standardized basic feature data are embedded into a unified feature space through a cross-modal feature fusion network, and the weights of different modal features are dynamically adjusted using an attention mechanism according to their contribution to income assessment, thereby obtaining corresponding fusion features; The fused features are further dimensionally upgraded through a nonlinear mapping method to generate the high-dimensional feature representation.

6. The method for training an income assessment model according to claim 1, characterized in that: The reinforcement learning algorithm is used to iteratively train and adjust parameters of the preset model according to the high-dimensional feature representation to obtain the income evaluation model, which specifically includes: By means of the proxy mechanism of the reinforcement learning algorithm, the contribution of the high-dimensional feature representation to the model output result is analyzed to obtain weight values ​​of different features, and the model parameters are dynamically adjusted according to the weight values; By utilizing a multi-round training optimization mechanism, the model weights and hyperparameters are iteratively updated according to the reward function feedback in each round of training, thereby completing the global parameter adjustment and generating the income evaluation model.

7. A training system for an income assessment model, characterized in that: The training system of the income assessment model includes: A data collection module is used to obtain the user's structured financial data, time series behavior data and unstructured text data, and to construct semantic and logical associations among the structured financial data, the time series behavior data and the unstructured text data based on a graph neural network to generate a multi-dimensional feature graph; A trend analysis module, which is used to perform trend analysis on the time series behavior data and the multi-dimensional feature graph based on Transformer and long short-term memory network, and extract features from the unstructured text data in combination with a preset language model to obtain basic feature data; A feature fusion module, used to perform feature fusion mapping based on the key financial indicators of the structured financial data and the basic feature data through cross-modal feature fusion technology, thereby obtaining a corresponding high-dimensional feature representation; The dynamic optimization module is used to utilize a reinforcement learning algorithm and iteratively train and adjust parameters of a preset model according to the high-dimensional feature representation to obtain an income evaluation model.

8. The training system for the income assessment model according to claim 7, characterized in that: The data acquisition module specifically includes: A node construction submodule, used to represent the key financial indicators in the structured financial data as node features, generate node weights based on the time series behavior data, generate corresponding feature vectors in combination with the unstructured text data, and then construct graph nodes according to the node features, the node weights and the feature vectors; The association relationship generation submodule is used to generate edge weights between the graph nodes and establish semantic association relationships and logical association relationships between the nodes through statistical analysis methods and causal inference methods; The graph embedding submodule is used to embed nodes and edges using a graph neural network, aggregate neighborhood node features through a message passing mechanism, capture high-order semantic relationships between nodes, and dynamically update the embedded features based on a time-aware mechanism to generate the multi-dimensional feature graph that is aware of time series.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the processor implements the steps of the method for training the income assessment model according to any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for training an income assessment model as claimed in any one of claims 1 to 6 are implemented.

Images