Automobile financial data processing method, platform and system based on big data

Abstract

The invention discloses an automobile industry financial data processing method based on big data, and relates to the field of data processing, and the method comprises the steps: obtaining automobile industry financial data; analyzing data circulation characteristics of the automobile financial data; constructing a business and financial data graph network according to the data circulation characteristics of the automobile business and financial data; sampling the business and financial data graph network based on the feature similarity to obtain a circulation feature sample set, and dividing the circulation feature sample set into an abnormal training set and an abnormal test set; performing negative sample sampling on the abnormal training set to obtain a mixed training set; after carrying out abnormal point marking on the mixed training set, inputting the mixed training set into a preset business and financial anomaly detection model for iterative training to obtain a target business and financial anomaly detection model; and inputting the abnormal test set into the target industry and financial anomaly detection model to obtain the data anomaly degree of the automobile industry and financial data. According to the invention, high-precision anomaly detection can be effectively carried out on data with large data volume and high dimension.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of business and financial data processing, and in particular, to a method, platform and system for processing automobile business and financial data based on big data. Background Art

[0002] With the rapid development of social economy, the automotive industry financial data stored in enterprises has shown a trend of large-scale growth. At the same time, due to the continuous rise of information technology, automotive industry financial data has achieved information exchange through big data. Automotive industry financial data covers data in finance, sales, market, customers and other aspects. Based on these data, reliable development policies and strategic decisions can be made for the subsequent development of enterprises. However, due to the wide coverage of data and the huge amount of data, some abnormal data are inevitable. For example, the sales data of the sales department has dropped sharply compared with the previous few months, and these abnormal data may be accompanied by some abnormal behaviors. These abnormal behaviors and abnormal data will affect the subsequent development of the enterprise to a certain extent. Therefore, it is of great significance to conduct abnormal analysis and evaluation of automotive industry financial data.

[0003] In the existing technology, the detection of business and financial data anomalies mainly relies on a certain distribution (such as normal distribution, Poisson distribution and binomial distribution, etc.) or probability model. The anomaly detection of business and financial data is achieved by judging whether a certain data point conforms to the distribution / model, that is, by distinguishing low-probability events. However, this method has low sensitivity and large computational complexity. For high-dimensional business and financial data, its anomaly detection effect has low accuracy and large errors. Summary of the invention

[0004] The embodiment of the present application provides an automobile industry financial data processing method based on big data, which is used to solve the problem that the existing technology cannot achieve high-precision anomaly detection for data with large volume and high dimension.

[0005] To achieve the above objectives, the embodiments of the present application adopt the following technical solutions:

[0006] In a first aspect, a method for processing automobile industry financial data based on big data is provided, the method comprising:

[0007] Obtain the automotive industry financial data of the target enterprise;

[0008] Performing data circulation feature analysis based on the automobile industry financial data to obtain data circulation features;

[0009] All departments of the target enterprise and multiple sub-departments corresponding to the departments are used as department nodes, and the data flow characteristics are used as feature edges between the department nodes to construct a business and financial data graph network of the target enterprise, wherein the business and financial data graph network is a high-dimensional data graph network;

[0010] Extracting the feature similarity between each of the department nodes and a plurality of neighboring nodes in the business and financial data graph network, screening the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain a plurality of data circulation feature samples, and integrating all of the data circulation feature samples to obtain a circulation feature sample set, wherein the neighboring nodes are a plurality of adjacent nodes around the department nodes;

[0011] Randomly dividing the circulation feature sample set into an abnormal training set and an abnormal test set;

[0012] Performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set, and obtaining a mixed training set after the negative samples are expanded;

[0013] A data grid space is constructed based on the mixed training set, a grid density of grid cells in the data grid space is calculated, all abnormal grids in the data grid space are marked according to the grid density, and all target training nodes in the abnormal grid are marked as abnormal points based on angle features and distance features between all target training nodes in the abnormal grid;

[0014] Input the mixed training set after the abnormal point marking is completed into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, and obtain the target business and financial anomaly detection model after the training is completed;

[0015] The abnormal test set is input into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

[0016] Optionally, the extracting of feature similarities between each of the department nodes and a plurality of neighboring nodes in the business and financial data graph network, screening the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarities to obtain a plurality of data circulation feature samples, and integrating all of the data circulation feature samples to obtain a circulation feature sample set comprises the following steps:

[0017] For each of the department nodes in the business and financial data graph network, extract feature vectors of the department node and all neighbor nodes from the business and financial data graph network to obtain a department feature vector and multiple neighbor feature vectors;

[0018] Calculating the feature similarity between the department feature vector and each of the neighbor feature vectors using a similarity formula;

[0019] If there is a target feature similarity that exceeds a preset similarity threshold, the department node, the target neighbor node corresponding to the target feature similarity, and the target feature edge between the department node and the target neighbor node are extracted as data flow feature samples;

[0020] After traversing all the department nodes, all the data circulation feature samples are integrated to obtain a circulation feature sample set.

[0021] Optionally, the performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set to obtain the mixed training set after the negative sample expansion comprises the following steps;

[0022] Extracting a plurality of the data circulation feature samples from the abnormal training set, and equally dividing and integrating the data circulation feature samples to obtain a first feature sample set and a second feature sample set;

[0023] Extracting the sample feature vector of the first feature sample set and the sample feature value corresponding to the sample feature vector by principal component analysis;

[0024] Filter out the sample feature vectors whose sample feature values ​​are greater than or equal to a preset feature value threshold to obtain a plurality of principal components, where the principal components are key features of the first feature sample set;

[0025] The key features are used as feature subspaces, and feature perturbations are performed on all the feature subspaces to generate multiple feature negative samples;

[0026] Extracting multiple subgraphs of the second feature sample set, and using all of the subgraphs as structural subspaces;

[0027] Performing structural perturbations on all the structural subspaces to generate multiple structural negative samples;

[0028] All of the feature negative samples and the structure negative samples are merged into the abnormal training set to obtain a mixed training set.

[0029] Optionally, constructing a data grid space based on the mixed training set, calculating the grid density of grid cells in the data grid space, marking all abnormal grids in the data grid space according to the grid density, and marking all target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all target training nodes in the abnormal grid comprises the following steps:

[0030] Performing dimensionality reduction processing on the mixed training set according to rough set theory;

[0031] Performing data density estimation according to the data distribution status of the mixed training set in each dimension after dimensionality reduction, and constructing multiple data dense spaces of the mixed training set in different dimensions according to the data density estimation results;

[0032] Acquire multiple hyper-rectangular sequences of the mixed training set through the data dense space, and calculate the sequence volume of each of the hyper-rectangular sequences respectively;

[0033] Gridding the data dense space in each dimension based on the super-rectangular sequence with the smallest sequence volume to obtain a data grid space;

[0034] After mapping all training nodes in the hybrid training set to the data grid space, calculating the grid density of each grid unit in the data grid space;

[0035] Taking the grid unit with the largest grid density as the central grid;

[0036] Based on the grid density, respectively calculate the relative density difference between the adjacent grids of each of the central grids and the central grid;

[0037] If the relative density difference is less than or equal to a preset density difference threshold, marking the adjacent grid as a normal grid;

[0038] If the relative density difference is greater than the density difference threshold, marking the adjacent grid as an abnormal grid;

[0039] Repeating the grid classification and marking step based on the grid density until each of the grid units has participated in the calculation of the relative density difference at least once;

[0040] Marking all nodes in the normal grid and the central grid as normal nodes;

[0041] Marking the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid;

[0042] The grid classification and marking step comprises:

[0043] Taking any normal grid that is not used as a reference grid as a reference grid, and calculating the reference relative density difference between the adjacent grids of each reference grid and the reference grid based on the grid density;

[0044] If the reference relative density difference is less than or equal to the density difference threshold, marking the adjacent grid corresponding to the reference grid as the normal grid;

[0045] If the reference relative density difference is greater than the density difference threshold, the adjacent grid corresponding to the reference grid is marked as the abnormal grid.

[0046] Optionally, the dimensionality reduction processing of the mixed training set according to the rough set theory comprises the following steps:

[0047] Extracting edge features and node features of each training node in the mixed training set respectively, and integrating the node features and edge features into comprehensive features corresponding to the training nodes;

[0048] Based on all the comprehensive features and according to the rough set theory, the feature weights of the comprehensive features are calculated respectively. The specific formula is as follows:

[0049]

[0050] Wherein, U is the mixed training set, P(R1) is the positive domain of the edge feature in the comprehensive feature, P(R2) is the positive domain of the node feature in the comprehensive feature, I(k i ,k j )-I(k i ) is the edge feature k in the comprehensive feature i Subject to the edge feature k j The impact of I(p i ,p j )-I(p i ) The node feature p in the comprehensive feature i Affected by the node feature p j The impact of I(k i ) and I(p i ) is the primary feature weight of the mixed training set, x is the total number of edge features in the comprehensive feature, and y is the total number of node features in the comprehensive feature;

[0051] If the feature weight of the comprehensive feature is less than the preset feature weight threshold, the corresponding training node is removed from the mixed training set.

[0052] Optionally, marking the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid comprises the following steps:

[0053] Randomly select a target training node from the abnormal grid as a reference node, and perform an abnormal point identification and marking step on the reference node until the abnormal point identification and marking step is performed on each of the target training nodes at least once;

[0054] The abnormal point identification and marking step comprises:

[0055] Determine a plurality of reference neighboring nodes of the reference node in the abnormal grid;

[0056] Adding a connection edge between the reference node and the reference neighboring node;

[0057] Calculate the angle characteristic of the angle with the reference node as the vertex using the cosine formula;

[0058] Calculating the distance features between the reference node and all the reference neighboring nodes respectively, and assigning weights to the angle features according to the distance features;

[0059] The abnormality of the reference node is obtained according to the angle feature after completing the weight assignment and using the abnormality calculation formula, and the abnormality calculation formula is as follows:

[0060]

[0061] Wherein, A is the reference node, B1 and B2 are the reference neighbor nodes, Wn is a set of n target training nodes, is the vector between the reference node A and the reference neighboring node B1, is the vector between the reference node A and the reference neighboring node B2, is the distance between the reference node A and the reference neighboring node B1, is the distance between the reference node A and the reference neighboring node B2, <·, ·> represents the dot product between two vectors, and VAR represents the variance calculation function;

[0062] If the abnormality of the reference node is greater than a preset abnormality threshold, marking the reference node as an abnormal node;

[0063] If the abnormality of the reference node is less than or equal to the abnormality threshold, the reference node is marked as the normal node.

[0064] Optionally, the mixed training set after the abnormal point marking is input into a preset business and financial anomaly detection model for iterative training until the number of iterative training reaches a preset maximum number of training times, and the target business and financial anomaly detection model after training is obtained is as follows:

[0065] The mixed training set after the abnormal point marking is put into the preset business and financial anomaly detection model;

[0066] Aggregating the comprehensive features of all the training nodes in the mixed training set using an aggregation function to obtain node aggregation features;

[0067] The node aggregation features are forward propagated, and the parameters of the business and financial anomaly detection model are adjusted until the number of parameter adjustments reaches a preset maximum number of training times, thereby obtaining a target business and financial anomaly detection model.

[0068] Optionally, the aggregation function is as follows:

[0069]

[0070] Among them, W is the node aggregation feature, Tx is the number of nodes in the mixed training set, F is the fully connected layer in the business and financial anomaly detection model, is the weight of the edge feature, is the weight of the node feature, and is the node information of the training node i in the mixed training set and all the corresponding neighboring nodes.

[0071] In a second aspect, the present application provides an automobile industry financial data processing platform based on big data, the platform comprising:

[0072] Information acquisition module, used to obtain the target enterprise's automotive industry financial data;

[0073] A feature analysis module, used to perform data circulation feature analysis based on the automobile industry financial data to obtain data circulation features;

[0074] A network construction module, used to construct a business and financial data graph network of the target enterprise by taking all departments of the target enterprise and multiple sub-departments corresponding to the departments as department nodes, and taking the data flow characteristics as feature edges between the department nodes, wherein the business and financial data graph network is a high-dimensional data graph network;

[0075] The first sampling module is used to extract the feature similarity between each of the department nodes and multiple neighboring nodes in the business and financial data graph network, screen the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain multiple data circulation feature samples, and integrate all the data circulation feature samples to obtain a circulation feature sample set, wherein the neighboring nodes are multiple adjacent nodes around the department nodes;

[0076] A sample division module, used for randomly dividing the circulation feature sample set into an abnormal training set and an abnormal test set;

[0077] A second sampling module is used to perform negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set to obtain a mixed training set after the negative samples are expanded;

[0078] An information marking module, used to construct a data grid space based on the mixed training set, calculate the grid density of the grid cells in the data grid space, mark all abnormal grids in the data grid space according to the grid density, and mark all the target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all the target training nodes in the abnormal grid;

[0079] A model training module is used to input the mixed training set after the abnormal point marking is completed into a preset business and financial anomaly detection model for iterative training until the number of iterative training reaches a preset maximum number of training times, thereby obtaining a target business and financial anomaly detection model after the training is completed;

[0080] The data detection module is used to input the abnormal test set into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

[0081] In a third aspect, an automobile industry financial data processing system based on big data includes:

[0082] a memory configured to store instructions; and

[0083] The processor is configured to call the instructions from the memory and implement a method for processing automobile business financial data based on big data according to any one of the first aspects when executing the instructions.

[0084] Through the above technical scheme, the target enterprise's automobile business and financial data is used to construct the target enterprise's business and financial data graph network, which is convenient for the subsequent extraction of the business and financial training set to train the pre-built business and financial anomaly detection model. Then, the feature similarity between the department node and the corresponding neighbor node in the business and financial data graph network is extracted, and the business and financial data graph network is sampled using the feature similarity to obtain the business and financial training set. In order to ensure that the positive and negative samples in the training set used for model training are approximately balanced and there is a certain similarity between the positive and negative samples, it is necessary to use the business and financial training set to sample negative samples, so as to obtain a mixed training set after the negative samples are expanded. Using the mixed training set to train the business and financial anomaly detection model can effectively improve the robustness of the business and financial anomaly detection model, so that the subsequent monitoring results of abnormal data monitoring of the target enterprise are more reliable. After obtaining the mixed training set, first mark the outliers in the mixed training set to improve the generalization ability of the business and financial anomaly detection model, which can effectively optimize the model. Finally, input the mixed training set with marked outliers into the business and financial anomaly detection model to obtain the target business and financial anomaly detection model. The target business and financial anomaly detection model is then used to identify anomalies in the target company's automotive business and financial data. In summary, a method is provided that can achieve high-precision anomaly detection for data with large amounts of data and high dimensions, which can accurately identify the abnormal data of the target company and ensure the long-term development of the target company.

[0085] Other features and advantages of the embodiments of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A flowchart of a method for processing automobile industry financial data based on big data provided in an embodiment of the present application;

[0087] Figure 2 A schematic diagram of the structure of the business and financial anomaly detection model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific implementation methods described herein are only used to illustrate and explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0089] It should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back...), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0090] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present application, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0091] Figure 1 The following schematically shows a flow chart of a method for processing automobile industry financial data based on big data according to an embodiment of the present application. Figure 1 As shown, the embodiment of the present application provides a method for processing automobile business financial data based on big data, which may include the following steps:

[0092] S101. Obtain automobile industry financial data of the target enterprise.

[0093] In this embodiment, the automobile business financial data can be obtained through the database inside the target enterprise, which contains all the data information inside the target enterprise, such as the personnel appointments of the target enterprise, the revenue status of the target enterprise, the research and development progress of the target enterprise, etc. The automobile business financial data includes the operating status of each department of the target enterprise and the business interaction data between the departments, wherein the operating status of the departments includes the sales of the marketing department, the research and development progress of the research and development department, the production efficiency of the production department, the customer maintenance information of the after-sales service department, etc. The business interaction data between the departments includes the manufacturability assessment report transmitted by the production department to the research and development department, and the vehicle quality monitoring report sent by the quality management department to the production department.

[0094] S102. Analyze data circulation characteristics based on automobile industry financial data to obtain data circulation characteristics.

[0095] In this embodiment, the automobile industry financial data is analyzed to determine whether there is data flow between various departments and the multiple sub-departments corresponding to each department. If data flow exists, the frequency and method of data flow are analyzed. These characteristics are data flow characteristics. Data flow refers to the information exchange between departments. For example, the legal department will exchange information with R&D, production and other departments to conduct legal risk assessments for these departments to ensure that the new automobile products produced and developed by the target company comply with relevant legal provisions. The R&D, production and sales departments will also jointly provide the basis for the project budget to the finance department for review.

[0096] S103. All departments of the target enterprise and the multiple sub-departments corresponding to the departments are taken as department nodes, and the data flow characteristics are taken as the characteristic edges between the department nodes to construct the business and financial data graph network of the target enterprise. The business and financial data graph network is a high-dimensional data graph network.

[0097] In this embodiment, each department and its corresponding sub-department are used as nodes of the data graph network, departments with information exchange are connected, and the connection relationship between them, the frequency and mode of data exchange and other data flow characteristics are used as the characteristic edges of the data graph network, and finally a business and financial data graph network is formed. Among them, departments include R&D departments, production departments, sales departments, etc., and sub-departments are departments under these departments. For example, the R&D department has product development department, engineering department, testing and verification department, etc., and the production department includes manufacturing department, process department, equipment maintenance department, etc. There is also information exchange between these sub-departments.

[0098] S104. Extract the feature similarity between each department node and multiple neighbor nodes in the business and financial data graph network, screen the department nodes and feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain multiple data circulation feature samples, and integrate all data circulation feature samples to obtain a circulation feature sample set, where the neighbor nodes are multiple adjacent nodes around the department nodes.

[0099] In this embodiment, for different department nodes in the business and financial data graph network, there are certain differences in their node degree values. The higher the node degree value, the more important the node is, because it has more direct neighbor nodes, which may have a greater impact on the network structure of the entire business and financial data graph network. Therefore, the local information represented by each department node is also different. If these department nodes are simply randomly sampled, nodes containing more information may be ignored, increasing the computational burden. Therefore, it is necessary to adopt a non-uniform or weighted sampling method for sampling. The node features of each department node and the neighbor nodes corresponding to the department node are extracted. The node features refer to the operating status of the department, the work tasks of the department, etc. The extracted node features are reduced to a low-dimensional vector space to obtain the feature vectors of the department node and the neighbor nodes, that is, the department feature vector and multiple neighbor feature vectors. Then, based on the department feature vector and the multiple neighbor feature vectors and using the similarity formula, the feature similarity between each department node and the corresponding multiple neighbor nodes is calculated.

[0100] The larger the cosine similarity, the greater the similarity between the department node and the neighbor node. The department nodes with high similarity, the corresponding target neighbor nodes, and the target feature edges between the department nodes and the target neighbor nodes are extracted as data flow feature samples. After calculating the similarity between all department nodes and neighbor nodes and completing the screening based on the similarity, all the obtained data flow feature samples are integrated to obtain the flow feature sample set. This method of calculating the similarity between department nodes and neighbor nodes for priority sampling can reduce the complexity of the business and financial anomaly detection model as much as possible and improve the versatility of the business and financial anomaly detection model.

[0101] S105. Randomly divide the circulation feature sample set into an abnormal training set and an abnormal test set.

[0102] In this embodiment, the circulation feature sample set is randomly divided into an abnormal training set and an abnormal test set to facilitate subsequent model training and model testing.

[0103] S106 , performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set, and obtaining a mixed training set after negative sample expansion.

[0104] In this embodiment, since the negative samples in the abnormal training set are far less than the positive samples, and the imbalance of positive and negative samples often makes it difficult to train the business and financial anomaly detection model, affecting the accuracy of the subsequent business and financial anomaly detection model for business and financial data anomaly detection, it is necessary to expand the abnormal training set through negative sample sampling so that the positive and negative samples in the abnormal training set are approximately balanced. Negative samples can be generated using the subspace perturbation method. The subspace perturbation method is designed based on the assumption that "positive samples are surrounded by negative samples". The negative sample data generated by the subspace perturbation method is close to the positive sample, which is conducive to the model to effectively distinguish positive and negative samples through training and improve the generalization of the model.

[0105] Specifically, some data circulation feature samples are extracted from the abnormal training set, and then the data circulation feature samples are perturbed by adding noise, deleting feature edges, etc., to obtain negative samples with a certain similarity to the data circulation feature samples, and all the obtained negative samples are added to the abnormal training set. The abnormal training set is expanded to obtain a mixed training set with a balanced positive and negative samples.

[0106] S107. Construct a data grid space based on the mixed training set, calculate the grid density of the grid cells in the data grid space, mark all abnormal grids in the data grid space according to the grid density, and mark all target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all target training nodes in the abnormal grid.

[0107] In this embodiment, since the automobile business and financial data is high-dimensional data, the business and financial graph data network is a high-dimensional data graph network, and the distance distinction between department nodes is not obvious, the evaluation result obtained by evaluating the outliers only based on the distance between department nodes is not accurate, so the angle variance between nodes is introduced to identify and mark the outliers of the training nodes in the mixed training set. In addition, since the amount of data in the training set is large, if the angle variance is calculated between all the training nodes in the training set, the calculation efficiency will be greatly reduced, resulting in a waste of resources. Therefore, before the angle variance calculation, the mixed training set is first gridded, and the abnormal grids and normal grids are screened out. The training nodes in the normal grid are first marked as normal nodes, and then the angle features and distance features of the target training nodes in the abnormal grid are calculated. The angle variance between the target training nodes is calculated based on the calculated angle features and distance features, and the abnormality of the target training nodes is obtained based on the angle variance. The target training nodes are marked as abnormal points based on the abnormality, that is, the target training nodes with an abnormality greater than the preset abnormality threshold are marked as abnormal nodes, and the target training nodes with an abnormality less than or equal to the abnormality threshold are marked as normal nodes. By marking normal nodes and abnormal nodes to facilitate subsequent training of the business and financial anomaly detection model, the business and financial anomaly detection model pays more attention to normal samples during the training process, which helps the business and financial anomaly detection model capture the true distribution of data, improve the robustness of the business and financial anomaly detection model, and improve the generalization ability of the business and financial anomaly detection model.

[0108] S108, inputting the mixed training set after the abnormal point marking into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, thereby obtaining the target business and financial anomaly detection model after the training is completed.

[0109] After the hybrid training set with the anomaly point marking is put into the preset business and financial anomaly detection model, the hybrid training set is firstly information-transferred to obtain the updated new training node, and then multiple neighboring nodes of the new training node in the hybrid training set that has completed information transmission and information update are obtained, and all new training nodes are aggregated with the corresponding neighboring nodes to obtain node aggregation features. The node aggregation features are forward propagated, and the parameters of the business and financial anomaly detection model are adjusted until the number of parameter adjustments reaches the preset maximum number of training times, the model training is completed, and the target business and financial anomaly detection model is obtained.

[0110] S109: Input the abnormal test set into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

[0111] The abnormal test set is input into the target business and financial anomaly detection model to obtain the data anomaly degree of each department node in the automotive business and financial data. Based on the obtained data anomaly degree, it is possible to analyze whether there is any abnormal behavior in each department of the target enterprise. When abnormal behavior is detected, an alarm message will be automatically output, which is convenient for the target enterprise to timely investigate and eliminate abnormal behavior.

[0112] In one embodiment, extracting the feature similarity between each department node and multiple neighboring nodes in the business and financial data graph network, screening the department nodes and feature edges of the department nodes in the business and financial data graph network according to the feature similarity, obtaining multiple data circulation feature samples, and integrating all data circulation feature samples to obtain a circulation feature sample set includes the following steps:

[0113] For each department node in the business and financial data graph network, extract feature vectors of the department node and all neighbor nodes from the business and financial data graph network to obtain a department feature vector and multiple neighbor feature vectors;

[0114] The similarity formula is used to calculate the feature similarity between the department feature vector and each neighbor feature vector;

[0115] If there is a target feature similarity that exceeds a preset similarity threshold, the department node, the target neighbor node corresponding to the target feature similarity, and the target feature edge between the department node and the target neighbor node are extracted as data flow feature samples;

[0116] After traversing all department nodes, all data circulation feature samples are integrated to obtain a circulation feature sample set.

[0117] In this embodiment, the node features of each department node and the neighbor nodes corresponding to the department node are extracted. The node features refer to the operating status of the department, the work tasks of the department, etc. The extracted node features are reduced to a low-dimensional vector space, and then a covariance matrix is ​​constructed based on the node features after the dimension reduction. After the covariance matrix is ​​subjected to eigenvalue decomposition, the feature vectors of the department nodes and the neighbor nodes are obtained, namely, the department feature vector and the neighbor feature vector. The department feature vector and the neighbor feature vector are input into the similarity formula, and the feature similarity between the department feature vector and the neighbor feature vector is calculated. This feature similarity is used to measure the similarity between the department node and the corresponding neighbor node. Commonly used similarity formulas include the cosine similarity formula, the Jaccard similarity calculation formula, etc. Taking the cosine similarity formula as an example, the calculation formula is as follows:

[0118]

[0119] Among them, a x is the department feature vector of department node x, b y is the neighbor feature vector of neighbor node y.

[0120] After calculating the feature similarity of all department nodes, the department nodes with feature similarity greater than the preset feature similarity threshold, the corresponding target neighbor nodes, and the target feature edges between the department nodes and the target neighbor nodes are extracted as data flow feature samples. The similarity between all department nodes and neighbor nodes is calculated in the same way, and they are screened according to the similarity. All data flow feature samples are integrated to obtain a flow feature sample set. This method of calculating the similarity between department nodes and neighbor nodes for priority sampling can reduce the complexity of the business and financial anomaly detection model as much as possible and improve the versatility of the business and financial anomaly detection model.

[0121] In one embodiment, it is characterized in that performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set, and obtaining the mixed training set after the negative sample expansion includes the following steps:

[0122] Extract multiple data flow feature samples from the abnormal training set, divide and integrate the data flow feature samples to obtain a first feature sample set and a second feature sample set;

[0123] Extracting the sample feature vector of the first feature sample set and the sample feature value corresponding to the sample feature vector by principal component analysis;

[0124] The sample feature vectors whose sample feature values ​​are greater than or equal to a preset feature value threshold are screened out to obtain a plurality of principal components, where the principal components are key features of the first feature sample set;

[0125] The key features are used as feature subspaces, and all feature subspaces are perturbed to generate multiple feature negative samples;

[0126] Extracting multiple subgraphs of the second feature sample set, and using all the subgraphs as structural subspaces;

[0127] Perform structural perturbations on all structural subspaces to generate multiple structural negative samples;

[0128] All feature negative samples and structural negative samples are merged into the abnormal training set to obtain a mixed training set.

[0129] In this embodiment, a plurality of data circulation feature samples are first extracted from the abnormal training set, and the data circulation feature samples are evenly divided and integrated to obtain two feature sample sets, namely the first feature sample set and the second feature sample set. This is to ensure that the generated negative samples have a certain degree of similarity with the data circulation feature samples in the abnormal training set, and to ensure that the trained business and financial anomaly detection model has a stronger generalization ability. The principal component analysis method is used to extract a plurality of the most important key features of the first feature sample set. The specific steps are: firstly, the first feature sample set is cleaned, and then the feature sample set after data cleaning is standardized. After the standardization is completed, the feature sample set is calculated to calculate its covariance matrix, and then the covariance matrix is ​​decomposed by eigenvalue to obtain the eigenvalue and the corresponding eigenvector, and the eigenvalue is arranged in descending order, and the sample feature vector whose sample eigenvalue is greater than or equal to the preset eigenvalue threshold is selected as the principal component. The obtained principal component is the key feature of the first feature sample set, and the obtained key feature is used as the feature subspace. In addition, the feature subspace is the key feature of the data circulation feature sample. The key features include the data characteristics, degree and clustering coefficient of the data circulation feature sample in the first feature sample set. Among them, the data characteristics represent the data characteristics of each department node in the data circulation feature sample. For example, the data of the sales department node has the characteristics of timeliness, sharing, and bidirectionality, and the data of the R&D department node has the characteristics of security, integrity, clarity, etc.

[0130] Perform feature perturbation on the feature subspace to generate multiple feature negative samples. Feature perturbation refers to randomly adding Gaussian noise or uniform noise to the feature subspace to obtain new key features, and the data flow feature samples with new key features are used as feature negative samples. The structural subspace refers to the data graph network extracted from the second data flow feature sample, that is, the subgraph. Structural perturbation refers to randomly adding or deleting feature edges to these subgraphs, thereby changing the connection relationship between department nodes, and obtaining data flow feature samples with different structures. These data flow feature samples with different structures are used as structural negative samples. The obtained feature negative samples and structural negative samples are merged into the abnormal training set as negative samples to obtain a mixed training set.

[0131] In one embodiment, constructing a data grid space based on a mixed training set, calculating the grid density of grid cells in the data grid space, marking all abnormal grids in the data grid space according to the grid density, and marking all target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all target training nodes in the abnormal grid include the following steps:

[0132] According to the rough set theory, the mixed training set is reduced in dimension;

[0133] According to the data distribution of the mixed training set in each dimension after dimensionality reduction, data density estimation is performed, and multiple data dense spaces of the mixed training set in different dimensions are constructed according to the data density estimation results;

[0134] Obtain multiple hyper-rectangular sequences of the mixed training set through the data dense space, and calculate the sequence volume of each hyper-rectangular sequence respectively;

[0135] Based on the hyper-rectangular sequence with the smallest sequence volume, the data dense space is gridded in each dimension to obtain the data grid space;

[0136] After mapping all training nodes in the hybrid training set to the data grid space, the grid density of each grid cell in the data grid space is calculated;

[0137] The grid cell with the largest grid density is taken as the central grid;

[0138] Based on the grid density, the relative density difference between the adjacent grids of each central grid and the central grid is calculated respectively;

[0139] If the relative density difference is less than or equal to the preset density difference threshold, the adjacent grid is marked as a normal grid;

[0140] If the relative density difference is greater than the density difference threshold, the adjacent grid is marked as an abnormal grid;

[0141] Repeat the grid classification and marking steps based on the grid density until each grid cell has participated in at least one relative density difference calculation;

[0142] Mark all nodes in the normal grid and the central grid as normal nodes;

[0143] Mark the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid;

[0144] The grid classification and labeling steps include:

[0145] Taking any normal grid that is not used as a reference grid as a reference grid, the reference relative density difference between the adjacent grids of each reference grid and the reference grid is calculated based on the grid density;

[0146] If the reference relative density difference is less than or equal to the density difference threshold, the adjacent grid corresponding to the reference grid is marked as a normal grid;

[0147] If the reference relative density difference is greater than the density difference threshold, the adjacent grid corresponding to the reference grid is marked as an abnormal grid.

[0148] In this embodiment, the mixed training set is first reduced in dimension using the rough set theory. The rough set theory is a mathematical tool for dealing with uncertainty. It can characterize incompleteness and uncertainty, effectively analyze various incomplete information such as imprecise, inconsistent, and incomplete, and can also analyze and reason the data to discover implicit knowledge and reveal potential laws. Through the rough set theory, unnecessary features in the data set can be identified, redundant features can be removed, and core features can be retained, thereby reducing the dimension of the data. This method achieves the effect of dimensionality reduction without losing the amount of information.

[0149] The mixed training set after dimensionality reduction is mapped to multiple one-dimensional spaces, and the data distribution status on each one-dimensional space is counted, that is, the number of training nodes on each one-dimensional space. The training node refers to the department node in the mixed training set. Then each one-dimensional space is evenly divided into multiple one-dimensional grids, and the data density in each one-dimensional grid is calculated. The specific calculation method is to divide the number of samples by the corresponding one-dimensional grid area, screen out the one-dimensional grids with data density greater than the preset data density threshold to obtain a dense one-dimensional grid, aggregate adjacent dense one-dimensional grids to obtain multiple data dense intervals, and obtain multiple data dense spaces according to the multiple data dense intervals of the mixed training set in different dimensions. Specifically, the total number of training nodes in the mixed training set is T, the data dense interval is [Si,Si] (i=1,2,……,d), and the data dense space is G=[S1,S1]×[S2,S2]×……×[Sd,Sd], i refers to the dimension of the mixed training set after dimensionality reduction, and d refers to the dth one-dimensional space. Then calculate the centroid of the data dense space, find the data dense space that contains the centroid and only contains √T training nodes in the entire data dense space as the hyperrectangular sequence, and calculate the volume of the hyperrectangular sequence. The centroid calculation formula is as follows:

[0150]

[0151] Among them, O j is the center of gravity of the jth data dense space, H ij is the i-th training node in the j-th data dense space, and N is the number of training nodes in the data dense space.

[0152] The formula for calculating the volume of a hyperrectangular sequence is as follows:

[0153]

[0154] in, and are the maximum and minimum values ​​of the hyperrectangle sequence in the nth dimension, respectively.

[0155] The length of the smallest hyperrectangle sequence in each dimension is used as the standard for dividing the data grid space. The space is divided according to this standard to obtain the data grid space, which is composed of multiple grid cells. All training nodes in the mixed training set are mapped to the data grid space. After completing the training node mapping, the number of training nodes falling into the grid cell is counted, and the total number of training nodes falling into each grid cell is used as the grid density of the grid cell. The grid cell with the largest grid density is taken as the central grid. The grid density difference between the central grid and the adjacent grids with overlapping edges of the central grid is calculated according to the grid density to obtain the relative density difference between the central grid and the adjacent grids. The adjacent grids with a relative density difference less than or equal to the relative density difference threshold are marked as normal grids. The adjacent grids with a relative density difference greater than the relative density difference threshold are marked as abnormal grids. Repeat the grid classification and marking steps until each grid cell has participated in at least one relative density difference calculation. After completing the grid classification and marking steps, the training nodes in the normal grid and the central grid are marked as normal nodes. This step is to reduce the amount of calculation when marking abnormal points in the subsequent process. Only the training nodes in the abnormal grid need to be marked as abnormal points in the subsequent process. When marking outliers, the angle is introduced as an evaluation factor for outlier marking, and the distance is used as the angle weight to calculate the angle variance between the target training node and other target training nodes. The calculated angle variance is used as the outlier degree of the target training node, and the target training nodes with an outlier degree greater than the preset outlier degree threshold are marked as outliers, and the target training nodes with an outlier degree less than or equal to the outlier degree threshold are marked as normal nodes, completing the marking of all target training nodes in the abnormal grid.

[0156] Specifically, the grid classification and marking step includes: taking a normal grid that is not used as a reference grid as a reference grid, calculating a reference relative density difference between the reference grid and the adjacent grids of the reference grid based on the grid density, marking the adjacent grids corresponding to the reference grid whose reference relative density difference is less than or equal to the density difference threshold as normal grids, and marking the adjacent grids corresponding to the reference grid whose reference relative density difference is greater than the density difference threshold as abnormal grids. The reference grid refers to the difference between the grid density of the grid and the grid density of the corresponding adjacent grid calculated based on the grid density of the grid.

[0157] In one embodiment, the dimensionality reduction process of the mixed training set according to the rough set theory includes the following steps:

[0158] Extract edge features and node features of each training node in the mixed training set respectively, and integrate the node features and edge features into the comprehensive features of the corresponding training nodes;

[0159] Based on all comprehensive features and according to the rough set theory, the feature weights of each comprehensive feature are calculated respectively. The specific formula is as follows:

[0160]

[0161] Among them, U is the mixed training set, P(R1) is the positive domain of edge features in the comprehensive features, P(R2) is the positive domain of node features in the comprehensive features, and I(k i ,k j )-I(k i ) is the edge feature k in the comprehensive feature i Edge characteristic k j The impact of I(p i ,p j )-I(p i )Node feature p in comprehensive feature i Affected by node feature p j The impact of I(k i ) and I(p i ) is the primary feature weight of the mixed training set, x is the total number of edge features in the comprehensive feature, and y is the total number of node features in the comprehensive feature;

[0162] If the feature weight of the comprehensive feature is less than the preset feature weight threshold, the corresponding training node will be removed from the mixed training set.

[0163] In this embodiment, the edge features and node features of the feature edges of the mixed training set and the department nodes are combined. The edge features refer to the frequency and mode of information flow between department nodes, and the node features refer to the node features of the department nodes, which refer to the operating status of the department, the work tasks of the department, etc. For example, the node features of the R&D department nodes refer to the current research progress of the R&D department, the type of products being studied, etc. The extracted node features and edge features are integrated to obtain comprehensive features. The rough set theory is used to calculate all feature weights of the mixed training set. The rough set theory can identify unnecessary features in the data set, remove redundant features, and retain core features, thereby reducing the dimension of the data. This method achieves the effect of dimensionality reduction without losing the amount of information. The feature weight of each feature in the comprehensive feature is calculated using the rough set theory calculation formula, and the features whose feature weights are less than the preset feature weight threshold are eliminated to obtain the mixed training set after dimensionality reduction. In addition, the positive domain of the comprehensive feature is a commonly used concept in the rough set theory, which refers to the intersection of the upper similarity and the lower similarity of the comprehensive feature. The primary feature weight is the feature weight whose calculation method is inaccurate, so it is necessary to further calculate to obtain the accurate feature weight.

[0164] In one embodiment, it is characterized in that marking the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid comprises the following steps:

[0165] Randomly select a target training node from the abnormal grid as a reference node, and perform an abnormal point identification and marking step on the reference node until each target training node has been subjected to the abnormal point identification and marking step at least once;

[0166] The steps of outlier identification and marking include:

[0167] Determine a plurality of reference neighboring nodes of the reference node in the abnormal grid;

[0168] Add connecting edges between the benchmark node and the benchmark neighboring nodes;

[0169] The angle characteristic of the angle with the reference node as the vertex is calculated using the cosine formula;

[0170] Calculate the distance features between the reference node and all reference neighboring nodes respectively, and assign weights to the angle features according to the distance features;

[0171] According to the angle characteristics of the completed weight distribution and using the abnormality calculation formula to obtain the abnormality of the reference node, the abnormality calculation formula is as follows:

[0172]

[0173] Among them, A is the reference node, B1 and B2 are the reference neighbor nodes, Wn is the set of n target training nodes, is the vector between the reference node A and the reference neighboring node B1, is the vector between the reference node A and the reference neighboring node B2, is the distance between the reference node A and the reference neighboring node B1, is the distance between the reference node A and the reference neighbor node B2, <·,·> represents the dot product between two vectors, and VAR represents the variance calculation function;

[0174] If the abnormality of the reference node is greater than the preset abnormality threshold, the reference node is marked as an abnormal node;

[0175] If the abnormality of the reference node is less than or equal to the abnormality threshold, the reference node is marked as a normal node.

[0176] In this embodiment, a target training node is randomly selected from the abnormal grid as a reference node. The reference node is selected here to facilitate the subsequent calculation of the abnormality of the reference node, and multiple (can be two) target training nodes are randomly selected as reference neighbor nodes of the reference node. The reference neighbor node is selected to construct an angle with the reference node as the vertex with the reference node, and a connecting edge is added between the reference node and the reference neighbor node, that is, the reference node and the reference neighbor node are connected as an angle, and the reference node is used as the vertex. The cosine formula is used to calculate the cosine value of the angle, and the cosine value is the angle feature of the angle. Then, the distance between the reference node and the reference neighbor node is calculated as the distance feature, and the angle feature is weighted according to the distance feature. The formula representing the distance weighting in the abnormality calculation formula is as follows:

[0177]

[0178] The formula for the angle feature is as follows:

[0179]

[0180] The calculated angular variance is used as the abnormality of the target training node. The target training nodes with an abnormality greater than the preset abnormality threshold are marked as abnormal nodes, and the target training nodes with an abnormality less than or equal to the abnormality threshold are marked as normal nodes, completing the marking of all target training nodes in the abnormal grid.

[0181] In one embodiment, the mixed training set after the abnormal point marking is input into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, and the target business and financial anomaly detection model after training is obtained is as follows:

[0182] The mixed training set after the abnormal point marking is put into the preset business and financial anomaly detection model;

[0183] The aggregation function is used to aggregate the comprehensive features of all training nodes in the mixed training set to obtain node aggregation features;

[0184] The node aggregation features are forward propagated, and the parameters of the business and financial anomaly detection model are adjusted until the number of parameter adjustments reaches the preset maximum number of training times to obtain the target business and financial anomaly detection model.

[0185] In this embodiment, refer to Figure 2 , the mixed training set after the abnormal point marking is put into the preset business and financial anomaly detection model for training. The preset business and financial anomaly detection model is built based on the graph neural network model. The steps to build the business and financial anomaly detection model are as follows:

[0186] 1.1. Collect appropriate graph data, i.e., business and financial data graph grid, including department nodes, feature edges, node features, and edge features, and standardize the collected graph data;

[0187] 1.2. Construct a node feature matrix containing the node features of each department node;

[0188] 1.3. Choose the appropriate architecture of the graph neural grid model, such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), etc.

[0189] 1.4. Define the hierarchical structure of the graph neural grid model, including graph convolutional layers, activation functions, etc., and select appropriate aggregation methods (such as average, maximum, weighted, etc.) to aggregate the information of neighbor nodes;

[0190] 1.5. According to the specific task (such as node classification, graph classification, link prediction, etc.), select the appropriate loss function, such as cross entropy loss, mean square error, etc., and finally obtain the business anomaly detection model to be trained.

[0191] First, use the K-nearest neighbor algorithm to obtain multiple neighboring nodes of the new training node. The K-nearest neighbor algorithm is a basic machine learning algorithm. Its core idea is: if most of the K nearest neighboring samples of a sample in the feature space belong to a certain category, then the sample also belongs to this category. The KNN algorithm calculates the distance between the sample to be classified and the known category sample, selects the K samples with the closest distance, and determines the category of the sample to be classified based on the category of these neighbors. First, calculate the distance between any new training node and other new training nodes, and take other new training nodes with a distance less than the preset distance threshold as the neighboring nodes of the new training node. This method is used to calculate the neighboring nodes of all new training nodes.

[0192] According to the propagation mechanism of the graph neural grid model, information is first transferred to the mixed training set, and then information is transferred between different training nodes along the feature edges connecting different training nodes, so that each training node can obtain the information of its neighboring nodes, complete the information update, and obtain a new training node.

[0193] Then, an aggregation function is used to aggregate information of the new training node and the corresponding multiple neighboring nodes to obtain node aggregation features. Information aggregation refers to feature aggregation of the new training node and the corresponding multiple neighboring nodes. The obtained node aggregation features refer to the features of the new training node after feature aggregation. The information of the new training node refers to the information of various departments of the target enterprise. For example, the information of the sales department includes sales performance, customer information, etc.

[0194] The node aggregation features are forward propagated to obtain the model output value, and the error of the model output value is calculated according to the pre-set loss function. Commonly used loss functions include cross entropy loss function and mean square error loss function, etc. Then the gradient of the loss function is calculated through the back propagation algorithm. The model parameters are updated according to the gradient of the loss function and the optimization algorithm. When the number of parameter updates reaches the preset maximum number of training times, the model training is completed and the target business and financial anomaly detection model is obtained.

[0195] In this embodiment, the aggregation function is as follows:

[0196]

[0197] Among them, W is the node aggregation feature, Tx is the number of nodes in the mixed training set, and F is the fully connected layer in the business and financial anomaly detection model. is the weight of the edge feature, α is the weight of the node feature, e (i) is the node information of training node i and all corresponding neighboring nodes in the mixed training set.

[0198] In one embodiment, a big data-based automobile industry financial data processing platform is characterized in that the platform includes:

[0199] Information acquisition module, used to obtain the target enterprise's automotive industry financial data;

[0200] The feature analysis module is used to analyze the data circulation features based on the automobile industry financial data to obtain the data circulation features;

[0201] A network construction module is used to construct a business and financial data graph network of the target enterprise by taking all departments of the target enterprise and multiple sub-departments corresponding to the departments as department nodes, and taking data flow characteristics as characteristic edges between department nodes. The business and financial data graph network is a high-dimensional data graph network;

[0202] The first sampling module is used to extract the feature similarity between each department node and multiple neighbor nodes in the business and financial data graph network, screen the department nodes and feature edges of the department nodes in the business and financial data graph network according to the feature similarity, obtain multiple data circulation feature samples, and integrate all data circulation feature samples to obtain a circulation feature sample set, where the neighbor nodes are multiple adjacent nodes around the department nodes;

[0203] A sample division module is used to randomly divide the circulation feature sample set into an abnormal training set and an abnormal test set;

[0204] The second sampling module is used to perform negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set to obtain a mixed training set after the negative samples are expanded;

[0205] An information marking module is used to construct a data grid space based on the mixed training set, calculate the grid density of the grid cells in the data grid space, mark all abnormal grids in the data grid space according to the grid density, and mark all target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all target training nodes in the abnormal grid;

[0206] The model training module is used to input the mixed training set after the abnormal point marking into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, and obtain the target business and financial anomaly detection model after the training is completed;

[0207] The data detection module is used to input the abnormal test set into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

[0208] In this embodiment, the information acquisition module acquires the automobile business financial data of the target enterprise through the database inside the target enterprise, which contains all the data information inside the target enterprise, such as personnel appointments of the target enterprise, revenue status of the target enterprise, research and development progress of the target enterprise, etc. The automobile business financial data includes the operating status of each department of the target enterprise and the business interaction data between the departments, wherein the operating status of the departments includes the sales of the marketing department, the research and development progress of the research and development department, the production efficiency of the production department, the customer maintenance information of the after-sales service department, etc. The business interaction data between the departments includes the manufacturability assessment report transmitted by the production department to the research and development department, and the vehicle quality monitoring report sent by the quality management department to the production department.

[0209] The feature analysis module is responsible for analyzing the automotive industry financial data to see if there is data flow between the various departments and the multiple sub-departments corresponding to each department. If there is data flow, the frequency and method of data flow are analyzed. These characteristics are data flow characteristics. Data flow refers to the information exchange between departments. For example, the legal department will exchange information with the R&D, production and other departments to conduct legal risk assessments for these departments to ensure that the new automotive products produced and developed by the target company comply with relevant legal provisions. The R&D, production and sales departments will also jointly provide the basis for the project budget to the finance department for review.

[0210] The network construction module is used to use each department and its corresponding sub-department as nodes of the data graph network, connect departments with information interaction, and use the connection relationship between them, the frequency and method of data flow and other data flow characteristics as the characteristic edges of the data graph network, ultimately forming a business and financial data graph network.

[0211] The first sampling module is responsible for sampling the business and financial data graph network according to the feature similarity between the department node and multiple neighbor nodes. Specifically, the node features of each department node and the neighbor nodes corresponding to the department node are extracted. The node features refer to the operating status of the department, the work tasks of the department, etc. The extracted node features are reduced to a low-dimensional vector space to obtain the feature vectors of the department node and the neighbor nodes, that is, the department feature vector and multiple neighbor feature vectors. Then, based on the department feature vector and multiple neighbor feature vectors and using the similarity formula, the feature similarity between each department node and the corresponding multiple neighbor nodes is calculated. The department nodes with high similarity, the corresponding target neighbor nodes, and the target feature edges between the department nodes and the target neighbor nodes are extracted as data circulation feature samples. The similarity between all department nodes and neighbor nodes is calculated and after screening based on the similarity, all the obtained data circulation feature samples are integrated to obtain a circulation feature sample set.

[0212] The sample partitioning module is used to randomly divide the circulation feature sample set into an abnormal training set and an abnormal test set to facilitate subsequent model training and model testing.

[0213] The second sampling module is responsible for expanding the abnormal training set through negative sample sampling, so that the positive and negative samples in the abnormal training set are approximately balanced. Negative samples can be generated using the subspace perturbation method. Specifically, some data flow feature samples are extracted from the abnormal training set, and then the data flow feature samples are perturbed by adding noise, deleting feature edges, etc., to obtain negative samples with a certain degree of similarity to the data flow feature samples, and all the obtained negative samples are added to the abnormal training set, and the abnormal training set is expanded to obtain a mixed training set with balanced positive and negative samples. The balance of positive and negative samples is conducive to the model to effectively distinguish positive and negative samples through training, thereby improving the generalization of the model.

[0214] The information marking module is used to identify and mark the abnormal points of the training nodes of the mixed training set according to the angle variance between the nodes. Due to the large amount of data in the training set, if the angle variance is calculated between all the training nodes in the training set, the calculation efficiency will be greatly reduced and resources will be wasted. Therefore, before the angle variance calculation, the mixed training set is first gridded, and the abnormal grids and normal grids are screened out. The training nodes in the normal grid are first marked as normal nodes, and then the angle features and distance features of the target training nodes in the abnormal grid are calculated. The angle variance between the target training nodes is calculated based on the calculated angle features and distance features, and the abnormality of the target training nodes is obtained based on the angle variance. The target training nodes are marked as abnormal points based on the abnormality, that is, the target training nodes with an abnormality greater than the preset abnormality threshold are marked as abnormal nodes, and the target training nodes with an abnormality less than or equal to the abnormality threshold are marked as normal nodes. By marking normal nodes and abnormal nodes to facilitate subsequent training of the business and financial anomaly detection model, the business and financial anomaly detection model pays more attention to normal samples during the training process, which helps the business and financial anomaly detection model capture the true distribution of data, improve the robustness of the business and financial anomaly detection model, and improve the generalization ability of the business and financial anomaly detection model.

[0215] The model training module is used to train the preset business and financial anomaly detection model based on the mixed training set with completed anomaly point marking. Specifically, after the mixed training set with completed anomaly point marking is inserted into the preset business and financial anomaly detection model, the mixed training set is information-transferred to obtain an updated new training node, and then multiple neighboring nodes of the new training node in the mixed training set that has completed information transmission and information update are obtained, and all new training nodes are aggregated with the corresponding neighboring nodes to obtain node aggregation features. The node aggregation features are forward propagated, and the parameters of the business and financial anomaly detection model are adjusted until the number of parameter adjustments reaches the preset maximum number of training times, the model training is completed, and the target business and financial anomaly detection model is obtained.

[0216] The data detection module is responsible for testing the target business and financial anomaly detection model using the anomaly test set, so as to analyze whether the target enterprise has abnormal behavior. Specifically, the anomaly test set is input into the target business and financial anomaly detection model to obtain the data anomaly degree of each department node in the automotive business and financial data. Based on the obtained data anomaly degree, it can be analyzed whether each department of the target enterprise has some abnormal behavior. When abnormal behavior is detected, an alarm message will be automatically output. This is convenient for the target enterprise to timely investigate and eliminate abnormal behavior.

[0217] In one embodiment, a big data-based automobile business financial data processing system is characterized by comprising:

[0218] a memory configured to store instructions; and

[0219] A processor is configured to call instructions from a memory and implement a method for processing automotive business financial data based on big data according to any of the above items when executing the instructions.

[0220] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0221] Among them, the memory can be an internal storage unit of a computer device, such as a hard disk or memory of a computer device, or an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device, etc., and the memory can also be a combination of an internal storage unit and an external storage device of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or is to be output, and this application does not impose any restrictions on this.

[0222] An embodiment of the present application also provides a machine-readable storage medium, on which instructions are stored, and the instructions are used to enable a machine to execute the above-mentioned method for processing automobile business financial data based on big data.

[0223] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0224] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0225] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0226] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0227] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0228] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0229] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0230] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0231] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for processing automobile industry financial data based on big data, characterized in that: The method comprises the following steps: Obtain the automotive industry financial data of the target enterprise; Performing data circulation feature analysis based on the automobile industry financial data to obtain data circulation features; All departments of the target enterprise and multiple sub-departments corresponding to the departments are used as department nodes, and the data flow characteristics are used as feature edges between the department nodes to construct a business and financial data graph network of the target enterprise, wherein the business and financial data graph network is a high-dimensional data graph network; Extracting the feature similarity between each of the department nodes and a plurality of neighboring nodes in the business and financial data graph network, screening the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain a plurality of data circulation feature samples, and integrating all of the data circulation feature samples to obtain a circulation feature sample set, wherein the neighboring nodes are a plurality of adjacent nodes around the department nodes; Randomly dividing the circulation feature sample set into an abnormal training set and an abnormal test set; Performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set, and obtaining a mixed training set after the negative samples are expanded; A data grid space is constructed based on the mixed training set, a grid density of grid cells in the data grid space is calculated, all abnormal grids in the data grid space are marked according to the grid density, and all target training nodes in the abnormal grid are marked as abnormal points based on angle features and distance features between all target training nodes in the abnormal grid; Input the mixed training set after the abnormal point marking is completed into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, and obtain the target business and financial anomaly detection model after the training is completed; The abnormal test set is input into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

2. The method according to claim 1, characterized in that The step of extracting the feature similarity between each of the department nodes and a plurality of neighboring nodes in the business and financial data graph network, screening the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain a plurality of data circulation feature samples, and integrating all of the data circulation feature samples to obtain a circulation feature sample set comprises the following steps: For each of the department nodes in the business and financial data graph network, extract feature vectors of the department node and all neighbor nodes from the business and financial data graph network to obtain a department feature vector and multiple neighbor feature vectors; Calculating the feature similarity between the department feature vector and each of the neighbor feature vectors using a similarity formula; If there is a target feature similarity that exceeds a preset similarity threshold, the department node, the target neighbor node corresponding to the target feature similarity, and the target feature edge between the department node and the target neighbor node are extracted as data flow feature samples; After traversing all the department nodes, all the data circulation feature samples are integrated to obtain a circulation feature sample set.

3. The method according to claim 1, characterized in that The step of performing negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set to obtain the mixed training set after the negative samples are expanded comprises the following steps: Extracting a plurality of the data circulation feature samples from the abnormal training set, and equally dividing and integrating the data circulation feature samples to obtain a first feature sample set and a second feature sample set; Extracting the sample feature vector of the first feature sample set and the sample feature value corresponding to the sample feature vector by principal component analysis; Filter out the sample feature vectors whose sample feature values ​​are greater than or equal to a preset feature value threshold to obtain a plurality of principal components, where the principal components are key features of the first feature sample set; The key features are used as feature subspaces, and feature perturbations are performed on all the feature subspaces to generate multiple feature negative samples; Extracting multiple subgraphs of the second feature sample set, and using all of the subgraphs as structural subspaces; Performing structural perturbations on all the structural subspaces to generate multiple structural negative samples; All of the feature negative samples and the structure negative samples are merged into the abnormal training set to obtain a mixed training set.

4. The method according to claim 1, characterized in that: The step of constructing a data grid space based on the mixed training set, calculating the grid density of grid cells in the data grid space, marking all abnormal grids in the data grid space according to the grid density, and marking all target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all target training nodes in the abnormal grid comprises the following steps: Performing dimensionality reduction processing on the mixed training set according to rough set theory; Performing data density estimation according to the data distribution status of the mixed training set in each dimension after dimensionality reduction, and constructing multiple data dense spaces of the mixed training set in different dimensions according to the data density estimation results; Acquire multiple hyper-rectangular sequences of the mixed training set through the data dense space, and calculate the sequence volume of each of the hyper-rectangular sequences respectively; Gridding the data dense space in each dimension based on the super-rectangular sequence with the smallest sequence volume to obtain a data grid space; After mapping all training nodes in the hybrid training set to the data grid space, calculating the grid density of each grid unit in the data grid space; Taking the grid unit with the largest grid density as the central grid; Based on the grid density, respectively calculate the relative density difference between the adjacent grids of each of the central grids and the central grid; If the relative density difference is less than or equal to a preset density difference threshold, marking the adjacent grid as a normal grid; If the relative density difference is greater than the density difference threshold, marking the adjacent grid as an abnormal grid; Repeating the grid classification and marking step based on the grid density until each of the grid units has participated in the calculation of the relative density difference at least once; Marking all nodes in the normal grid and the central grid as normal nodes; Marking the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid; The grid classification and marking step comprises: Taking any normal grid that is not used as a reference grid as a reference grid, and calculating the reference relative density difference between the adjacent grids of each reference grid and the reference grid based on the grid density; If the reference relative density difference is less than or equal to the density difference threshold, marking the adjacent grid corresponding to the reference grid as the normal grid; If the reference relative density difference is greater than the density difference threshold, the adjacent grid corresponding to the reference grid is marked as the abnormal grid.

5. The method according to claim 4, characterized in that The dimensionality reduction process of the mixed training set according to the rough set theory comprises the following steps: Extracting edge features and node features of each training node in the mixed training set respectively, and integrating the node features and edge features into comprehensive features corresponding to the training nodes; Based on all the comprehensive features and according to the rough set theory, the feature weights of the comprehensive features are calculated respectively. The specific formula is as follows: Wherein, U is the mixed training set, P(R1) is the positive domain of the edge feature in the comprehensive feature, P(R2) is the positive domain of the node feature in the comprehensive feature, I(k i ,k j )-I(k i ) is the edge feature k in the comprehensive feature i Subject to the edge feature k j The impact of I(p i ,p j )-I(p i ) The node feature p in the comprehensive feature i Affected by the node feature p j The impact of I(k i ) and I(p i ) is the primary feature weight of the mixed training set, x is the total number of edge features in the comprehensive feature, and y is the total number of node features in the comprehensive feature; If the feature weight of the comprehensive feature is less than the preset feature weight threshold, the corresponding training node is removed from the mixed training set.

6. The method according to claim 4, characterized in that The step of marking the target training nodes as abnormal points according to the angle features and distance features between all the target training nodes in the abnormal grid comprises the following steps: Randomly select a target training node from the abnormal grid as a reference node, and perform an abnormal point identification and marking step on the reference node until the abnormal point identification and marking step is performed on each of the target training nodes at least once; The abnormal point identification and marking step comprises: Determine a plurality of reference neighboring nodes of the reference node in the abnormal grid; Adding a connection edge between the reference node and the reference neighboring node; Calculate the angle characteristic of the angle with the reference node as the vertex using the cosine formula; Calculating the distance features between the reference node and all the reference neighboring nodes respectively, and assigning weights to the angle features according to the distance features; The abnormality of the reference node is obtained according to the angle feature after completing the weight assignment and using the abnormality calculation formula, and the abnormality calculation formula is as follows: Wherein, A is the reference node, B1 and B2 are the reference neighbor nodes, Wn is a set of n target training nodes, is the vector between the reference node A and the reference neighboring node B1, is the vector between the reference node A and the reference neighboring node B2, is the distance between the reference node A and the reference neighboring node B1, is the distance between the reference node A and the reference neighboring node B2, <·, ·> represents the dot product between two vectors, and VAR represents the variance calculation function; If the abnormality of the reference node is greater than a preset abnormality threshold, marking the reference node as an abnormal node; If the abnormality of the reference node is less than or equal to the abnormality threshold, the reference node is marked as the normal node.

7. The method according to claim 5, characterized in that The mixed training set after the abnormal point marking is input into the preset business and financial anomaly detection model for iterative training until the number of iterative training reaches the preset maximum number of training times, and the target business and financial anomaly detection model after the training is obtained is as follows: The mixed training set after the abnormal point marking is put into the preset business and financial anomaly detection model; Aggregating the comprehensive features of all the training nodes in the mixed training set using an aggregation function to obtain node aggregation features; The node aggregation features are forward propagated, and the parameters of the business and financial anomaly detection model are adjusted until the number of parameter adjustments reaches a preset maximum number of training times, thereby obtaining a target business and financial anomaly detection model.

8. The method according to claim 7, characterized in that The aggregation function is as follows: Wherein, W is the node aggregation feature, Tx is the number of nodes in the mixed training set, F is the fully connected layer in the business and financial anomaly detection model, is the weight of the edge feature, α is the weight of the node feature, e (i) is the node information of the training node i and all the corresponding neighboring nodes in the mixed training set.

9. An automotive industry financial data processing platform based on big data, characterized in that: The platform includes: Information acquisition module, used to obtain the target enterprise's automotive industry financial data; A feature analysis module, used to perform data circulation feature analysis based on the automobile industry financial data to obtain data circulation features; A network construction module, used to construct a business and financial data graph network of the target enterprise by taking all departments of the target enterprise and multiple sub-departments corresponding to the departments as department nodes, and taking the data flow characteristics as feature edges between the department nodes, wherein the business and financial data graph network is a high-dimensional data graph network; The first sampling module is used to extract the feature similarity between each of the department nodes and multiple neighboring nodes in the business and financial data graph network, screen the department nodes and the feature edges of the department nodes in the business and financial data graph network according to the feature similarity to obtain multiple data circulation feature samples, and integrate all the data circulation feature samples to obtain a circulation feature sample set, wherein the neighboring nodes are multiple adjacent nodes around the department nodes; A sample division module, used for randomly dividing the circulation feature sample set into an abnormal training set and an abnormal test set; A second sampling module is used to perform negative sample sampling on the abnormal training set to expand the negative samples in the abnormal training set to obtain a mixed training set after the negative samples are expanded; An information marking module, used to construct a data grid space based on the mixed training set, calculate the grid density of the grid cells in the data grid space, mark all abnormal grids in the data grid space according to the grid density, and mark all the target training nodes in the abnormal grid as abnormal points based on the angle features and distance features between all the target training nodes in the abnormal grid; A model training module is used to input the mixed training set after the abnormal point marking is completed into a preset business and financial anomaly detection model for iterative training until the number of iterative training reaches a preset maximum number of training times, thereby obtaining a target business and financial anomaly detection model after the training is completed; The data detection module is used to input the abnormal test set into the target business and financial anomaly detection model to obtain the data anomaly degree of the automobile business and financial data.

10. An automobile industry financial data processing system based on big data, characterized in that: include: a memory configured to store instructions; as well as A processor is configured to call the instructions from the memory and to implement a method for processing automobile business financial data based on big data according to any one of claims 1 to 8 when executing the instructions.

Images