A user credit analysis system based on big data

Abstract

The present invention relates to the technical field of big data analysis, and provides a user credit analysis system based on big data; a data collection module of the system collects structured data and unstructured data; sentiment features are extracted from unstructured data such as text, image, audio and video data through a multimodal fusion network, and fusion feature data is generated by combining numerical features in structured data; a strategy combining dictionary annotation and reinforcement learning is introduced to train SVM to obtain a LeELSVM classification model to evaluate user credit risk; through the above steps, the system can comprehensively and accurately analyze user credit status, effectively improve the accuracy and efficiency of credit risk management, and provide strong technical support for intelligent risk control and user management in the financial industry.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and in particular to a user credit analysis system based on big data. Background Art

[0002] With the rapid development of big data technology and financial technology, existing user credit assessment methods have become difficult to cope with complex and changeable user behavior patterns and multi-dimensional data types; existing credit analysis systems usually rely on single structured data, ignoring the potential emotional and behavioral characteristics in unstructured data, resulting in low sensitivity and insufficient accuracy in identifying user credit risks; in addition, existing methods have limited processing capabilities in feature extraction and data fusion, making it difficult to fully capture the deep correlation between multimodal data; this results in poor system performance when faced with complex credit risk scenarios; therefore, there is an urgent need for a user credit analysis system that can integrate multimodal data feature extraction, deep feature fusion and intelligent classification models to improve the processing capabilities of multi-dimensional data and enhance the accuracy and reliability of credit risk assessment. Summary of the invention

[0003] The existing credit analysis system has shown high accuracy and stability when processing traditional data sources and fixed features, but it still has certain limitations when dealing with complex and changeable user behavior patterns and multi-dimensional data; the sentiment and behavior characteristics in unstructured data are not sufficiently mined, resulting in room for improvement in the recognition accuracy and response speed of risks in some complex scenarios; in view of these shortcomings, the present invention proposes a user credit analysis system based on big data, which deeply integrates unstructured data such as text, images, audio and video with structured data by constructing a multimodal fusion feature extraction method to more comprehensively represent the credit characteristics of users; at the same time, the present invention introduces a strategy combining dictionary annotation and reinforcement learning to train SVM and obtain a LeELSVM classification model, which effectively improves the accuracy and generalization ability of credit risk assessment; the system of the present invention further improves the accuracy and response speed of credit assessment on the basis of the existing technology, better adapts to diverse user behavior patterns and complex risk environments, and achieves significant improvement in application effect and reliability.

[0004] The present invention provides a user credit analysis system based on big data, which includes a data collection module, a data preprocessing module, a feature engineering module, a model building module and a risk assessment and decision support module;

[0005] The data collection module collects the user's unstructured data and structured data;

[0006] The data preprocessing module preprocesses unstructured data and structured data. The unstructured data is processed by denoising, word segmentation, normalization and standardization to generate unstructured preprocessed data. The structured data is processed by data cleaning, integrity check and standardization to generate structured preprocessed data.

[0007] The feature engineering module uses a multimodal fusion network to extract weighted fusion emotional feature data based on unstructured preprocessed data, and directly extracts numerical feature data based on structured preprocessed data; the weighted fusion emotional feature data is fused with the numerical feature data to obtain emotional-numerical feature data;

[0008] The model building module adopts the strategy of optimizing SVM through reinforcement learning and dictionary annotation to obtain the EL-Le-SVMMT algorithm. According to the sentiment-numerical feature data, the EL-Le-SVMMT algorithm is used to generate the LeELSVM classification model. The EL-Le-SVMMT algorithm specifically includes the following steps: data partitioning and annotation, identifying query data, updating the data pool, checking the data pool, training the SVM model and improving the SVM model.

[0009] The risk assessment and decision support module obtains a user credit risk assessment report based on the LeELSVM classification model, and performs risk assessment and decision support based on the user credit risk assessment report.

[0010] Furthermore, the unstructured preprocessed data includes text preprocessed data, image preprocessed data, audio preprocessed data and video preprocessed data; the multimodal fusion network includes text emotion feature extraction, image emotion feature extraction, audio emotion feature extraction, video emotion feature extraction, feature fusion and dynamic attention fusion;

[0011] Furthermore, in the feature engineering module, the process of extracting text sentiment features specifically includes the following steps:

[0012] Step B1: Establish a BERT-BiGRU model, which includes a BERT layer and a BiGRU layer;

[0013] Step B2: Input the text preprocessing data into the BERT-BiGRU model. The BERT layer performs bidirectional encoding on the text preprocessing data and extracts context-related sentiment features.

[0014] Step B3: Input the context-related sentiment features into the BiGRU layer. The BiGRU layer extracts the temporal features by processing the sequence data of the text preprocessing data and generates a text sentiment feature vector.

[0015] Furthermore, in the feature engineering module, the process of extracting image emotion features specifically includes establishing a ResNet model, inputting the image preprocessing data into the ResNet model for convolution operation, extracting visual features, and analyzing the emotion information through facial expressions to obtain the image emotion feature vector.

[0016] Furthermore, in the feature engineering module, the process of extracting audio emotion features specifically includes integrating an open source audio processing library to obtain the Librosa tool, using the Librosa tool to extract audio features of audio preprocessing data, and performing speech emotion analysis to obtain an audio emotion feature vector.

[0017] Furthermore, in the feature engineering module, the process of extracting video emotion features specifically includes establishing an LSTM model, using the LSTM model to process the temporal features of the video preprocessing data, and combining the audio emotion feature vector to obtain the video emotion feature vector.

[0018] Furthermore, in the feature engineering module, the feature fusion process specifically includes splicing the text emotion feature vector, the image emotion feature vector, the audio emotion feature vector and the video emotion feature vector to generate a multimodal fusion feature vector.

[0019] Furthermore, in the feature engineering module, the process of dynamic attention fusion specifically includes combining multimodal fusion feature vectors, introducing learnable parameters and graph neural networks to capture the deep correlation of text emotion feature vectors, image emotion feature vectors, audio emotion feature vectors and video emotion feature vectors, and obtaining weighted fusion emotion feature data. The formula used is as follows:

[0020] ;

[0021] in, Represents the index, Representing modality The modality contains text emotion feature vector, image emotion feature vector, audio emotion feature vector and video emotion feature vector. represents the weighted fusion sentiment feature data, Indicates at time Lower mode The weight of represents a learnable parameter, Represents a graph neural network module that captures the deep correlation between data modalities. Represents the tuning parameters of the modal eigenvectors.

[0022] Furthermore, the process of generating the LeELSVM classification model using the EL-Le-SVMMT algorithm based on the sentiment-numerical feature data specifically includes the following steps:

[0023] Step S1: Data division and labeling: 10% of the sentiment-numerical feature data is used as test data, 85% of the sentiment-numerical feature data is used as training data, and the data is labeled using the Vader dictionary to obtain initial labeled training data; 5% of the sentiment-numerical feature data is used as a data pool, and extreme sentiment data is labeled using the Vader dictionary to obtain data pool feature data;

[0024] Step S2: Identify query data: Establish an ELSVM model, learn the initial labeled training data through the ELSVM model, identify the initial labeled training data with uncertain classification as query data, obtain query feature data, and annotate the query feature data again through the Vader dictionary to obtain annotated query feature data;

[0025] Step S3: updating the data pool: adding the labeled query feature data to the data pool feature data, and deleting the labeled query feature data from the initial labeled training data, to obtain updated data pool feature data;

[0026] Step S4: Check the data pool: set a preset threshold, check whether the size of the updated data pool feature data exceeds the preset threshold, if not, continue to execute steps S2 and S3, when it exceeds the preset threshold, stop executing, and finally obtain the complete data pool feature data and the remaining labeled training data;

[0027] Step S5: training the SVM model: establishing the SVM model, using the complete data pool feature data and the remaining labeled training data to train the SVM model, and obtaining a preliminary SVM model;

[0028] Step S6: Improve the SVM model: Input the test data into the preliminary SVM model, evaluate the preliminary SVM model by calculating the accuracy, precision, recall, F1 score and confusion matrix of the preliminary SVM model, and obtain the evaluation results. According to the evaluation results, improve the preliminary SVM model to obtain the LeELSVM classification model. The formula used is as follows:

[0029] Accuracy formula:

[0030] ;

[0031] in, Indicates accuracy, Represents the true class, represents the true negative class, represents the false positive class, represents the false negative class;

[0032] Precision formula:

[0033] ;

[0034] in, Indicates accuracy;

[0035] Recall formula:

[0036] ;

[0037] in, represents the recall rate;

[0038] F1 score formula:

[0039] ;

[0040] in, Represents the harmonic mean of precision and recall.

[0041] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0042] The present invention extracts sentiment features from unstructured data through a multimodal fusion network and fuses them with structured data, thereby realizing multi-dimensional and multi-level user credit analysis. The multimodal fusion network can not only effectively extract sentiment features from various data forms such as text, images, audio and video, but also deeply associate these features through the feature fusion layer and the dynamic attention fusion layer to generate weighted fused sentiment feature data. This multimodal feature fusion method improves the user credit analysis system's ability to mine hidden sentiment information in unstructured data, significantly enhances the model's accuracy and adaptability in identifying user credit risks in complex scenarios, overcomes the limitation of existing methods that they cannot effectively process multimodal data, and lays a solid foundation for comprehensive user credit risk assessment.

[0043] In addition, the present invention constructs a LeELSVM classification model by combining dictionary annotation and reinforcement learning, which solves the shortcomings of existing models in dealing with the complexity and dynamics of sentiment data; first, the Vader dictionary is used to accurately annotate the sentiment-numerical feature data to ensure the high quality of the model input data; then, the reinforcement learning support vector machine (LeELSVM) is used to dynamically optimize the data pool, thereby improving the model's ability to handle classification uncertainty data; the model is continuously iterated and updated during the training process, combined with the data pool management strategy, through the intelligent expansion and contraction of the data pool, effectively improving the generalization ability and robustness of the model, making the credit risk assessment results more accurate and stable, and providing a scientific basis for user management and decision support.

[0044] In summary, the present invention has achieved remarkable results in credit risk assessment and decision support; through the combination of multimodal fusion network and LeELSVM classification model, the deep fusion and precise classification of sentiment features and numerical features of multi-source data are realized; especially in the refined analysis of user credit risk and abnormal behavior identification, the system has demonstrated excellent performance, which not only improves the accuracy and reliability of credit risk prediction, but also significantly enhances the adaptability and intelligence level to complex data scenarios; overall, the present invention introduces a new technical path of multimodal sentiment analysis and reinforcement learning classification for the user credit analysis system, which has made outstanding contributions to the fields of financial risk control and market analysis while improving the intelligence and accuracy of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A module schematic diagram of a user credit analysis system based on big data proposed by the present invention;

[0046] Figure 2 Flowchart of extracting unstructured data using the multimodal fusion network of the feature engineering module proposed in the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Embodiment 1, according to Figure 1 The present invention provides a user credit analysis system based on big data, which includes a data collection module, a data preprocessing module, a feature engineering module, a model building module and a risk assessment and decision support module;

[0049] The data collection module collects the user's unstructured data and structured data;

[0050] The data preprocessing module preprocesses unstructured data and structured data. The unstructured data is processed by denoising, word segmentation, normalization and standardization to generate unstructured preprocessed data. The structured data is processed by data cleaning, integrity check and standardization to generate structured preprocessed data.

[0051] The feature engineering module uses a multimodal fusion network to extract weighted fusion emotional feature data based on unstructured preprocessed data, and directly extracts numerical feature data based on structured preprocessed data; the weighted fusion emotional feature data is fused with the numerical feature data to obtain emotional-numerical feature data;

[0052] The model building module adopts the strategy of optimizing SVM through reinforcement learning and dictionary annotation to obtain the EL-Le-SVMMT algorithm. According to the sentiment-numerical feature data, the EL-Le-SVMMT algorithm is used to generate the LeELSVM classification model. The EL-Le-SVMMT algorithm specifically includes the following steps: data partitioning and annotation, identifying query data, updating the data pool, checking the data pool, training the SVM model and improving the SVM model.

[0053] The risk assessment and decision support module obtains a user credit risk assessment report based on the LeELSVM classification model, and performs risk assessment and decision support based on the user credit risk assessment report.

[0054] Embodiment 2: This embodiment is based on the above embodiment, and the data collection module collects the user's unstructured data and structured data.

[0055] Unstructured data includes:

[0056] Text data: including social media posts, comments, blog posts;

[0057] Image data: expressions and scenes in pictures uploaded by users;

[0058] Audio data: user’s tone, speed, and volume;

[0059] Video data: facial expressions, body language, and voice information of users in the video;

[0060] Structured data includes: financial data, transaction data, asset data, occupational data and default data.

[0061] Embodiment 3, based on the above embodiment, the unstructured preprocessed data includes text preprocessed data, image preprocessed data, audio preprocessed data and video preprocessed data; the multimodal fusion network includes text emotion feature extraction, image emotion feature extraction, audio emotion feature extraction, video emotion feature extraction, feature fusion and dynamic attention fusion;

[0062] Embodiment 4, based on the above embodiment, in the feature engineering module, the process of extracting text sentiment features specifically includes the following steps:

[0063] Step B1: Establish a BERT-BiGRU model, which includes a BERT layer and a BiGRU layer;

[0064] Step B2: Input the text preprocessing data into the BERT-BiGRU model. The BERT layer performs bidirectional encoding on the text preprocessing data and extracts context-related sentiment features.

[0065] Step B3: Input the context-related sentiment features into the BiGRU layer. The BiGRU layer extracts the temporal features by processing the sequence data of the text preprocessing data and generates a text sentiment feature vector.

[0066] Embodiment 5, based on the above embodiment, in the feature engineering module, the process of image emotion feature extraction specifically includes establishing a ResNet model, inputting the image preprocessing data into the ResNet model for convolution operation, extracting visual features, and analyzing the emotion information through facial expressions to obtain the image emotion feature vector.

[0067] Embodiment 6: This embodiment is based on the above embodiment. In the feature engineering module, the process of extracting audio emotion features specifically includes integrating an open source audio processing library to obtain the Librosa tool, using the Librosa tool to extract audio features of audio preprocessing data, and performing speech emotion analysis to obtain an audio emotion feature vector.

[0068] Embodiment 7: This embodiment is based on the above embodiment. In the feature engineering module, the process of extracting video emotion features specifically includes establishing an LSTM model, using the LSTM model to process the temporal features of video preprocessing data, and combining the audio emotion feature vector to obtain the video emotion feature vector.

[0069] Embodiment 8: This embodiment is based on the above embodiment. In the feature engineering module, the feature fusion process specifically includes splicing the text emotion feature vector, the image emotion feature vector, the audio emotion feature vector and the video emotion feature vector to generate a multimodal fusion feature vector.

[0070] Embodiment 9, based on the above embodiment, in the feature engineering module, the process of dynamic attention fusion specifically includes combining multimodal fusion feature vectors, introducing learnable parameters and graph neural networks to capture the deep association of text emotion feature vectors, image emotion feature vectors, audio emotion feature vectors and video emotion feature vectors, and obtaining weighted fusion emotion feature data. The formula used is as follows:

[0071] ;

[0072] in, Represents the index, Representing modality The modality contains text emotion feature vector, image emotion feature vector, audio emotion feature vector and video emotion feature vector. represents the weighted fusion sentiment feature data, Indicates at time Lower mode The weight of represents a learnable parameter, Represents a graph neural network module that captures the deep correlation between data modalities. Represents the tuning parameters of the modal eigenvectors.

[0073] Embodiment 10, based on the above embodiment, the process of generating the LeELSVM classification model using the EL-Le-SVMMT algorithm according to the emotion-numerical feature data specifically includes the following steps:

[0074] Step S1: Data division and labeling: 10% of the sentiment-numerical feature data is used as test data, 85% of the sentiment-numerical feature data is used as training data, and the data is labeled using the Vader dictionary to obtain initial labeled training data; 5% of the sentiment-numerical feature data is used as a data pool, and extreme sentiment data is labeled using the Vader dictionary to obtain data pool feature data;

[0075] Step S2: Identify query data: Establish an ELSVM model, learn the initial labeled training data through the ELSVM model, identify the initial labeled training data with uncertain classification as query data, obtain query feature data, and annotate the query feature data again through the Vader dictionary to obtain annotated query feature data;

[0076] Step S3: updating the data pool: adding the labeled query feature data to the data pool feature data, and deleting the labeled query feature data from the initial labeled training data, to obtain updated data pool feature data;

[0077] Step S4: Check the data pool: set a preset threshold, check whether the size of the updated data pool feature data exceeds the preset threshold, if not, continue to execute steps S2 and S3, when it exceeds the preset threshold, stop executing, and finally obtain the complete data pool feature data and the remaining labeled training data;

[0078] Step S5: training the SVM model: establishing the SVM model, using the complete data pool feature data and the remaining labeled training data to train the SVM model, and obtaining a preliminary SVM model;

[0079] Step S6: Improve the SVM model: Input the test data into the preliminary SVM model, evaluate the preliminary SVM model by calculating the accuracy, precision, recall, F1 score and confusion matrix of the preliminary SVM model, and obtain the evaluation results. According to the evaluation results, improve the preliminary SVM model to obtain the LeELSVM classification model. The formula used is as follows:

[0080] Accuracy formula:

[0081] ;

[0082] in, Indicates accuracy, Represents the true class, represents the true negative class, represents the false positive class, represents the false negative class;

[0083] Precision formula:

[0084] ;

[0085] in, Indicates accuracy;

[0086] Recall formula:

[0087] ;

[0088] in, represents the recall rate;

[0089] F1 score formula:

[0090] ;

[0091] in, Represents the harmonic mean of precision and recall.

[0092] Embodiment 10: This embodiment is based on the above embodiment. The risk assessment and decision support module obtains a user credit risk assessment report according to the LeELSVM classification model, and performs risk assessment and decision support according to the user credit risk assessment report.

[0093] In this embodiment:

[0094] User A obtains the following user credit risk assessment report based on the LeELSVM classification model:

[0095] (1) User credit score:

[0096] 650 (score range is 300-900, with higher scores indicating lower credit risk);

[0097] (2) Risk level:

[0098] Medium risk (based on the credit score classification, medium risk means there is a certain credit risk, but it has not yet reached high risk);

[0099] (3) Main risk factors:

[0100] High debt level: The ratio of credit card debt to monthly income exceeds 200%, and repayment pressure is high;

[0101] Large emotional fluctuations: Social media and video sentiment analysis show that users' emotions have fluctuated greatly recently and their emotions are unstable;

[0102] Overdue record: There is one overdue payment record in the past year.

[0103] User A's decision support and management suggestions:

[0104] The credit card limit of user A is appropriately reduced by 10%.

Claims

1. A user credit analysis system based on big data, characterized by: It includes data collection module, data preprocessing module, feature engineering module and model building module; The data collection module collects the user's unstructured data and structured data; The data preprocessing module preprocesses unstructured data and structured data. The unstructured data is processed by denoising, word segmentation, normalization and standardization to generate unstructured preprocessed data. The structured data is processed by data cleaning, integrity check and standardization to generate structured preprocessed data. The feature engineering module uses a multimodal fusion network to extract weighted fusion emotional feature data based on unstructured preprocessed data, and directly extracts numerical feature data based on structured preprocessed data; the weighted fusion emotional feature data is fused with the numerical feature data to obtain emotional-numerical feature data; The model building module adopts the strategy of optimizing SVM by reinforcement learning and dictionary annotation to obtain the EL-Le-SVMMT algorithm. According to the sentiment-numerical feature data, the EL-Le-SVMMT algorithm is used to generate the LeELSVM classification model. The EL-Le-SVMMT algorithm specifically includes the following steps: data partitioning and annotation, identifying query data, updating the data pool, checking the data pool, training the SVM model and improving the SVM model; Unstructured data includes: Text data: including social media posts, comments, blog posts; Image data: expressions and scenes in pictures uploaded by users; Audio data: user’s tone, speed, and volume; Video data: facial expressions, body language, and voice information of users in the video; Structured data includes: financial data, transaction data, asset data, occupational data, and default data; The unstructured preprocessed data includes text preprocessed data, image preprocessed data, audio preprocessed data and video preprocessed data; the multimodal fusion network includes text emotion feature extraction, image emotion feature extraction, audio emotion feature extraction, video emotion feature extraction, feature fusion and dynamic attention fusion; The process of generating the LeELSVM classification model using the EL-Le-SVMMT algorithm based on the sentiment-numerical feature data specifically includes the following steps: Step S1: Data division and labeling: 10% of the sentiment-numerical feature data is used as test data, 85% of the sentiment-numerical feature data is used as training data, and the data is labeled using the Vader dictionary to obtain initial labeled training data; 5% of the sentiment-numerical feature data is used as a data pool, and extreme sentiment data is labeled using the Vader dictionary to obtain data pool feature data; Step S2: Identify query data: Establish an ELSVM model, learn the initial labeled training data through the ELSVM model, identify the initial labeled training data with uncertain classification as query data, obtain query feature data, and annotate the query feature data again through the Vader dictionary to obtain annotated query feature data; Step S3: updating the data pool: adding the labeled query feature data to the data pool feature data, and deleting the labeled query feature data from the initial labeled training data, to obtain updated data pool feature data; Step S4: Check the data pool: set a preset threshold, check whether the size of the updated data pool feature data exceeds the preset threshold, if not, continue to execute steps S2 and S3, when it exceeds the preset threshold, stop executing, and finally obtain the complete data pool feature data and the remaining labeled training data; Step S5: training the SVM model: establishing the SVM model, using the complete data pool feature data and the remaining labeled training data to train the SVM model, and obtaining a preliminary SVM model; Step S6: Improve the SVM model: Input the test data into the preliminary SVM model, evaluate the preliminary SVM model by calculating the accuracy, precision, recall, F1 score and confusion matrix of the preliminary SVM model, obtain the evaluation results, improve the preliminary SVM model according to the evaluation results, and obtain the LeELSVM classification model.

2. The user credit analysis system based on big data according to claim 1 is characterized by: In the feature engineering module, the process of text sentiment feature extraction includes the following steps: Step B1: Establish a BERT-BiGRU model, which includes a BERT layer and a BiGRU layer; Step B2: Input the text preprocessing data into the BERT-BiGRU model. The BERT layer performs bidirectional encoding on the text preprocessing data and extracts context-related sentiment features. Step B3: Input the context-related sentiment features into the BiGRU layer. The BiGRU layer extracts the temporal features by processing the sequence data of the text preprocessing data and generates a text sentiment feature vector.

3. The user credit analysis system based on big data according to claim 1 is characterized by: In the feature engineering module, the process of image emotion feature extraction specifically includes establishing a ResNet model, inputting the image preprocessing data into the ResNet model for convolution operation, extracting visual features, and analyzing emotional information through facial expressions to obtain the image emotion feature vector.

4. The user credit analysis system based on big data according to claim 1 is characterized by: In the feature engineering module, the process of audio emotion feature extraction specifically includes integrating the open source audio processing library to obtain the Librosa tool, using the Librosa tool to extract audio features of audio preprocessing data, and performing speech emotion analysis to obtain an audio emotion feature vector.

5. The user credit analysis system based on big data according to claim 4 is characterized by: In the feature engineering module, the process of video emotion feature extraction specifically includes establishing an LSTM model, using the LSTM model to process the temporal features of video preprocessing data, and combining the audio emotion feature vector to obtain the video emotion feature vector.

6. The user credit analysis system based on big data according to claim 5 is characterized by: In the feature engineering module, the feature fusion process specifically includes splicing the text emotion feature vector, the image emotion feature vector, the audio emotion feature vector and the video emotion feature vector to generate a multimodal fusion feature vector.

7. The user credit analysis system based on big data according to claim 6 is characterized by: In the feature engineering module, the process of dynamic attention fusion specifically includes combining multimodal fusion feature vectors, introducing learnable parameters and graph neural networks to capture the deep correlation between text sentiment feature vectors, image sentiment feature vectors, audio sentiment feature vectors and video sentiment feature vectors to obtain weighted fusion sentiment feature data.

Images