Model dimension prediction consistency evaluation method based on multi-view learning

Abstract

The invention discloses a model dimension prediction consistency evaluation method based on multi-view learning. The method comprises the steps that S1, a basic prediction model and multi-dimensional feature data are acquired; s2, combining the multi-dimensional feature data in pairs in sequence, constructing a consistency constraint function based on the Euclidean distance of the two feature data, adding all the obtained constraint functions into a loss function, and performing parameter tuning to obtain a multi-view learning model; s3, selecting two or more dimensions of feature data to apply constraints, and respectively constructing a multi-view learning model according to the method in S2; s4, evaluating the prediction performance indexes of the multi-view learning model obtained in S2 and S3 on the test set; and S5, extracting a constraint value and a prediction performance index, and outputting an evaluation result. According to the method, the consistency among different dimensions in the model with multi-dimensional index system expression can be effectively and quantitatively evaluated, and a basis is provided for construction of the multi-dimensional index system and extraction of a potential relationship, so that the model effect is assisted to be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-dimensional index system. More particularly, the present application relates to a model dimension prediction consistency evaluation method based on multi-view learning. BACKGROUND

[0002] With the rapid development of information technology, the scale and dimension of data acquisition show exponential growth. Massive multi-source heterogeneous data (such as text, image, sensor data, time series data, etc.) provide a rich information base for the training of machine learning and artificial intelligence models. Fully mining and utilizing these multi-dimensional data can effectively improve the prediction performance of the model and enhance the generalization ability of the model.

[0003] Because data of different dimensions may come from different collection methods, representation angles or data distributions, their prediction results may conflict with each other, even contradict each other, thereby reducing the accuracy and robustness of the model. Traditional methods usually only simply concatenate or weightedly fuse data from different dimensions. Although these methods are simple to implement, they are difficult to effectively solve the problems of heterogeneity between data and inconsistency of prediction results. Multi-view learning is an effective data fusion method, and its core idea is to integrate data from different sources and different dimensions through certain constraints or optimization strategies, fully utilize the complementary information of multi-dimensional data, and make the information of each dimension in the model mutually collaborative, thereby improving the prediction accuracy. Existing multi-view learning methods include strategies based on collaborative training, subspace learning, deep neural network fusion, etc. These methods realize the collaborative representation between data through different optimization objectives.

[0004] However, there is currently a lack of effective evaluation methods to quantify the consistency degree of different dimensional data in the prediction results. The lack of consistency evaluation methods not only cannot quantitatively analyze the collaborative and conflict relationship between different data dimensions, but also makes it difficult to provide a reliable theoretical basis for subsequent dimension selection and revealing the potential mutual relationship between the data of each dimension. SUMMARY

[0005] An object of the present application is to provide a model dimension prediction consistency evaluation method based on multi-view learning, which can effectively quantitatively evaluate the consistency between different dimensions in a model expressed by a multi-dimensional index system, provide a basis for the construction of a multi-dimensional index system and the extraction of potential relationships, and thereby assist in improving the model effect.

[0006] In order to achieve the objects and other advantages of the present application, according to one aspect of the present application, the present application provides a model dimension prediction consistency evaluation method based on multi-view learning, characterized in that, comprising: S1: obtaining a basic prediction model and multi-dimensional feature data; S2: combining the multi-dimensional feature data two by two in turn to construct a consistency constraint function based on the Euclidean distance of the two feature data, adding all the constraint functions obtained as penalty terms to the loss function of the basic prediction model, and performing parameter optimization to obtain a multi-view learning model; S3: selecting two or more dimensional feature data to apply constraints, and constructing a multi-view learning model according to the method of S2; S4: evaluating the prediction performance indicators of the multi-view learning models obtained in S2 and S3 on the test set, the prediction performance indicators including accuracy, precision, recall and F1 score; S5: extracting the constraint values of the multi-view learning model in S2 and the prediction performance indicators of the multi-view learning model in S4, and outputting the evaluation results.

[0007] Further, in S2, the calculation formula of the consistency constraint function g(w) is:

[0008]

[0009] k represents the number of dimensions, is the combination number, X v p and X v q are feature data of different dimensions, and are coefficients corresponding to the feature data, λ m as a hyperparameter is used to make the consistency constraint function and the loss function of the basic prediction model have comparable orders of magnitude.

[0010] Further, in S2, the objective function J(w) is constructed for parameter optimization, and the specific calculation formula is:

[0011]

[0012] wherein h(w) is the loss function of the basic prediction model.

[0013] Further, the gradient descent method is used to solve the objective function, and during the solving process, when the weight change of adjacent two iterations is less than a predetermined threshold, or the maximum iteration number is reached, the iteration is stopped.

[0014] Further, in S4, it further comprises: using at least three independent test sets with significant distribution difference to verify the same multi-view learning model across scenes, and the distribution difference is quantitatively confirmed by a data collection time span threshold or a feature covariance offset angle; calculating the weighted average and standard deviation of each performance indicator on multiple test sets, wherein the weight of the test set is dynamically adjusted according to its data freshness, and the more recent data obtains a higher evaluation weight.

[0015] Further, in S4, the distribution difference of the independent test set is quantitatively confirmed by a preset data collection time span threshold or a feature covariance offset angle, wherein the time span threshold is greater than or equal to 6 months, and the covariance offset angle is obtained by calculating the similarity of the feature covariance matrix; the weight distribution of the weighted average value adopts a time decay function, and the test set weight is negatively correlated with the data collection time, and the specific weight value is an exponential decreasing function of the difference between the current time and the data collection time.

[0016] Further, in S5, FGSM fast gradient sign attack or PGD projected gradient descent adversarial algorithm is also used to generate a perturbation test set, wherein the perturbation amplitude is controlled in the interval of 15%-20% of the feature value standard deviation, and the absolute deviation mean of the accuracy, precision, recall and F1 score of the original test set and the perturbation test set is calculated to set a double trigger mechanism of a single indicator deviation rate threshold of 10% and a comprehensive deviation rate threshold of 8%, when any indicator deviation rate breaks through 10% or the comprehensive deviation rate exceeds 8%, the gradient distribution data in the historical training log of the last three months is automatically called to perform gradient backtracking analysis based on the reverse tracking of the activation function, to locate the sensitive neuron layer causing the robustness defect, and to generate an evaluation review report containing the adversarial sample vulnerability distribution map in parallel.

[0017] Further, S5 also includes: constructing a 95% confidence interval of the constraint value based on Bootstrap resampling 1000 times, introducing a stratified sampling strategy in the sampling process to ensure uniformity of coverage of each feature dimension, when the F1 score optimal value in cross-scene verification deviates from the confidence interval, automatically extracting the constraint value sequence, performance index matrix and feature covariance tensor in the last three historical evaluations, constructing a Bayesian posterior distribution model with Gaussian process regression as the core, and performing distribution alignment correction by calculating the KL divergence between the current evaluation result and the posterior distribution, wherein the KL divergence threshold is set to 0.05 and the maximum iteration number is 50, and the L2 norm stability of the feature weight vector is optimized synchronously in the correction process, and finally outputting the corrected constraint value distribution curve and the weight vector convergence diagnosis report.

[0018] Further, the evaluation result output by S5 is input into the adversarial sample generation process, and when the comprehensive deviation rate of the original test set and the perturbation test set does not exceed 8%, the evaluation result is directly output, and if the deviation rate exceeds the threshold, gradient backtracking analysis is started; the result of the gradient backtracking analysis is input into the Bootstrap confidence interval construction process simultaneously, and the correction is performed in the form of KL divergence distribution alignment, and the corrected constraint value covers the corresponding value in the original evaluation result of S5; finally, a composite evaluation report containing the gradient backtracking analysis report and the Bootstrap confidence interval data is generated, and the corrected constraint value and the cross-scene performance index are marked.

[0019] The present application at least includes the following beneficial effects:

[0020] The present application can quantitatively analyze and compare the consistency of prediction effects between different perspectives. By comparing the numerical values of the consistency constraint penalty terms between different perspectives, the consistency of data characteristics under different perspectives in the prediction effect can be intuitively evaluated, which provides a basis for the construction of a multi-dimensional index system and reveals the potential mutual relationship between the dimensions of data. The characteristic data of different perspectives may contain unique information that other perspectives cannot provide. In a multi-perspective environment, the consistency constraint based on collaborative regularization proposed in the present application can integrate the knowledge of each perspective, thereby accurately extracting key information, making the multi-perspective model achieve prediction consistency between perspectives, improving the collaborative ability of different dimensional data in the model on the prediction task, and thus helping to improve the prediction ability of the model.

[0021] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following description, and will be understood by those skilled in the art upon reading and understanding the specification. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 Flowchart for an embodiment of the present application;

[0023] Figure 2 Partial sample data diagram for an embodiment of the present application.

[0024] Figure 3 Consistency evaluation result diagram for an embodiment of the present application. DETAILED DESCRIPTION

[0025] The present application will be further described in detail below, so that those skilled in the art can implement it according to the description.

[0026] It should be understood that the terms such as "have", "contain" and "include" used in the embodiments of the present application do not exclude the presence or addition of one or more other elements. All directional indications (such as upper, lower, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain specific posture, and when the specific posture changes, the directional indications also change accordingly. When an element is referred to as "fixed to" or "disposed on" another element, it can be directly on another element or can have a middle element. When an element is referred to as "connected" to another element, it can be directly connected to another element or indirectly connected to another element through a middle element. The description of "first", "second" and the like in the embodiments of the present application is only for the purpose of description and cannot be understood as indicating or implying the relative importance of the technical features indicated or implying the number of technical features indicated. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features.

[0027] It should be noted that the technical solutions of various embodiments of the present application can be combined with each other, but it must be based on the realization of a person skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor within the scope of protection claimed in the present application.

[0028] Embodiment 1:

[0029] A model dimension prediction consistency evaluation method based on multi-view learning includes the following steps:

[0030] Step one, the feature data obtained from different ways or different levels is preprocessed, such as cleaning, denoising, standardization, missing value filling, etc. The data is divided into different dimension data subsets according to certain rules, and a data set containing not less than two dimension data subsets is constructed. Different dimensions of data can be used to construct each view in the multi-view learning method.

[0031] It should be noted that the data dimension classification principle and method can be reasonably set by relevant personnel, and the obtained data can be divided into two or more dimension data subsets according to the data acquisition source, data type or according to a specific classification logic method. The division method of the view in the present application is not limited.

[0032] A financial fraud identification model based on multi-view learning is provided in embodiment 3 of the present application, which is used to predict the probability of occurrence of financial fraud events of listed companies. In the embodiment, the view classification method based on behavior logic and fraud conditions divides the obtained feature data into the dimension of hidden poor management motivation, the dimension of enterprise internal collusion conditions, the dimension of fraud means based on accounting manipulability and the dimension of external collusion to make the audit institution compliant.

[0033] Step two, using the different dimension feature data selected in step one, a basic prediction model is constructed.

[0034] An example of a basic prediction model based on logistic regression is given in Example 3 of the present application, which is used to predict whether the annual financial report of a listed company is fraudulent, and the specific expression is as follows:

[0035]

[0036] In the above formula, f(X) is the response variable, representing the probability of occurrence of financial fraud events of a listed company, is the observation variable under each view; D is the number of all observation variables, D k represents the number of observation variables of each view; k represents the number of divided feature data views, and N represents the number of observation values in the sample period; is the coefficient of the feature data under each view in the model.

[0037] Step three, through the basic prediction model established in step two, although the feature data from different dimensions can be used, the complementary effect of different dimension data is not effectively exerted, and different dimension feature data may not bring consistent prediction results. Therefore, in this step, a multi-view learning model is constructed by using the collaborative regularization method, that is, a consistency constraint represented by a collaborative regularization constraint is imposed on the prediction results of each view in the basic model constructed in step two, and the multi-view model with the consistency constraint is parameterized to optimize the model as the optimal model.

[0038] Step 3.1: a consistency constraint g(w) represented by a collaborative regularization is imposed on each two views in different views, and the specific calculation formula is:

[0039]

[0040] λ m As a hyperparameter, the consistency constraint function and the loss function of the basic model have comparable orders of magnitude.

[0041] In Example 3 of the present application, the feature data is divided into four views, so there are 6 combination forms in two combinations.

[0042] Step 3.2: the consistency constraint function in step 3.1 is added to the loss function h(w) of the basic model constructed in step two as a multi-view learning model with a consistency constraint, and parameter optimization is performed to obtain an optimal model, and the specific calculation formula of the optimization objective J(w) of the multi-view learning model with a consistency constraint is:

[0043]

[0044] Step four, compare the numerical size of the consistency constraint function between different views in the optimal model obtained in step three, so as to evaluate the consistency of the feature data of different views in the prediction effect. In order to ensure the robustness of the evaluation results, select any two or more views to apply consistency constraint, construct different multi-view learning models, evaluate the prediction effect of different multi-view learning models, and further analyze whether the feature data under different views has prediction consistency.

[0045] Step 4.1: Compare the consistency constraint penalty term results of each view in the optimal model obtained in step three, and evaluate the prediction consistency of each view. Specifically, the more similar the prediction results generated by different dimensional information, that is, the stronger the consistency of the data features of two dimensions in the prediction effect, the smaller the penalty term; otherwise, the larger the penalty term.

[0046] Step 4.2: Only select any two or more views to apply consistency constraint, and construct multi-view learning model. Evaluate the out-of-sample prediction effect of different multi-view learning models on the test set, so as to evaluate whether the multi-view learning model with consistency constraint is helpful to improve the prediction ability of the model. The out-of-sample prediction effect of the model is also used to verify the robustness of the results in step 4.1. Specifically, applying consistency constraint to two or more dimensions with strong consistency can improve the model synergy and then improve the prediction effect of the model. The optional model evaluation indicators include accuracy, precision, recall and F1 score, and the specific calculation formula is as follows:

[0047]

[0048] Preferably, other models with multi-dimensional index system expression can be constructed in step two. The multi-view learning model based on logistic regression in the embodiment of the application is intended to explain the application and cannot be understood as a limitation of the application.

[0049] According to the above technical solution, compared with the prior art, the model dimension prediction consistency evaluation method based on multi-view learning proposed in the embodiment has the following three advantages:

[0050] (1) The embodiment can quantitatively analyze and compare the consistency of the prediction effect between different views. By comparing the numerical value of the consistency constraint penalty term between different views, the consistency of the data features under different views in the prediction effect can be intuitively evaluated, which provides a basis for the construction of multi-dimensional index system and reveals the potential relationship between the dimensions of data.

[0051] (2) Different view angle feature data may contain unique information that other view angles cannot provide. In a multi-view environment, the consistency constraint based on collaborative regularization proposed in the present application can integrate the knowledge of each view angle, thereby accurately extracting key information, enabling the multi-view model to achieve prediction consistency between view angles, improving the collaborative ability of different dimensional data in the model in the prediction task, and thereby improving the prediction ability of the model.

[0052] (3) The method proposed in the present application is not limited to the prediction model based on logistic regression proposed in the examples, but is also applicable to model dimension prediction consistency evaluation in other arbitrary models, which will be illustrated below.

[0053] Example 2: Application of model dimension prediction consistency evaluation method based on multi-view learning in automatic driving

[0054] S1: Obtain basic data and model

[0055] Obtain a trained basic prediction model (such as a vehicle trajectory prediction model based on a convolutional neural network), and multi-dimensional feature data used for training, including vehicle-mounted camera images, millimeter wave radar distance and speed data, laser radar three-dimensional point cloud data, vehicle motion parameters, etc. These data constitute feature inputs of different dimensions.

[0056] S2: Construct a multi-view learning model in pairs

[0057] The multi-dimensional feature data is combined in pairs in turn (such as camera image + millimeter wave radar, laser radar + vehicle speed, etc.), and multiple view angle combinations are divided. A consistency constraint function is constructed based on the Euclidean distance of each view angle feature data, for example, the Euclidean distance of the prediction results of the camera and the radar data is constrained, so that the model can fuse the complementary information of visual appearance and radar ranging in rainy and foggy weather. Add all the constraint functions as penalty terms to the basic model loss function, and use gradient descent method (iteratively calculate the loss gradient and update the parameters in reverse) to optimize, and obtain a multi-view learning model in pairs.

[0058] S3: Select a single view or multi-view to apply constraints to construct a model

[0059] Optionally, one or more view angle combinations are applied to constraints (such as individually constraining the laser radar view angle, or combining the camera / radar / vehicle data of three view angles), and a multi-view learning model is constructed according to the method of S2. For example, when the laser radar view angle is individually constrained, the prediction stability of the three-dimensional point cloud data is strengthened through the consistency constraint; when the three view angles are combined, the prediction consistency of the multi-source data is simultaneously constrained to improve the adaptability to complex road conditions.

[0060] S4: Evaluate the performance of the multi-view learning model

[0061] The model is evaluated using test sets containing different weather (sunny / rainy / snowy), road conditions (highway / urban / rural), and indicators including accuracy (correctness of trajectory prediction), precision (correctness of predicted trajectory), recall (capture rate of actual trajectory), and F1 score. At least three distribution difference test sets (such as summer urban / winter highway / rainy rural data with a time span of more than 6 months, confirmed by covariance matrix similarity) are used for cross-scene verification, and weighted average indicators and standard deviations are calculated based on time decay function (recent data has higher weight).

[0062] S5: Output evaluation results and deep analysis

[0063] The constraint values of the model in S2 (such as the constraint strength of camera and radar view angle) and the performance indicators of S4 are extracted to generate initial evaluation results. When the single indicator deviation exceeds 10% or the comprehensive deviation exceeds 8%, gradient backtracking analysis is started (based on historical gradient reverse tracking to locate sensitive neuron layer). Bootstrap resampling (1000 stratified samples) is combined to construct the 95% confidence interval of the constraint value, and if the F1 score deviates from the interval, it is corrected by Bayesian posterior model (Gaussian process regression) and KL divergence, finally generating a composite evaluation report containing gradient backtracking report and confidence interval.

[0064] Embodiment 3:

[0065] As shown in Figure 1 , the present embodiment exemplarily proposes a model dimension prediction consistency evaluation method based on multi-view learning, and the specific scene of the present embodiment is a financial fraud identification model used to predict whether the annual financial report of a listed company is fraudulent, which includes the following steps:

[0066] Step 1: Constructing a data set for the financial fraud identification model.

[0067] This embodiment first divides the feature data into four perspectives based on the logic of enterprise behavior and the conditions for achieving fraud, namely the perspective of hidden poor management motivation, the perspective of internal collusion conditions, the perspective of fraud means based on accounting manipulability, and the perspective of external collusion with audit institutions, obtaining an index system containing four first-level perspectives, 17 second-level perspectives, and 61 measurement indicators.

[0068] From the perspective of companies concealing poor performance or financial distress, this study constructs four secondary perspectives with a total of 19 variables. Financial ratios, which characterize corporate performance, are used as indicators, selecting four main categories: profitability, solvency, operational efficiency, and growth potential. Furthermore, in accordance with my country's "ST" (Special Treatment) system regulations, two dummy scalars representing whether a company incurred losses in the previous year or two are introduced into the profitability analysis to better represent operational distress. The 19 financial ratios or data for listed companies are sourced from the Wind Financial Database.

[0069] In terms of internal collusion, four secondary perspectives with a total of 17 variables were constructed: the risk appetite of the senior management team, and the internal power, supervision level and equity structure of senior management that reflect the characteristics of the corporate organizational structure. The relevant data came from the CSMAR database.

[0070] From the perspective of fraudulent methods based on accounting manipulation, variables were constructed around several accounting items with high room for manipulation, including accounts receivable, inventory, related party transactions, period expenses, goodwill, and operating revenue. The financial data of a total of 20 indicator variables came from the Wind database.

[0071] From the perspective of external collusion in the cooperation of auditing firms, five indicators were constructed to measure the relationship between enterprises and auditors and their firms from three secondary perspectives: the status and employment of accounting firms, audit fees, and audit opinions on financial statements. The original data of the variables came from the CSMAR database.

[0072] Step 2: Sample Selection

[0073] This embodiment uses 323 cases involving financial fraud from January 1, 2010 to June 30, 2022. From the fiscal years 2009 to 2021, 534 corporate annual reports were subject to administrative penalties. This application uses these fraudulent annual reports as the research sample. Due to the relatively unique financial reporting structure and business types of the financial industry, corporate annual reports from the financial industry were excluded during sample construction, leaving 527 annual reports. Some annual reports lacked sufficient disclosure of information related to corporate governance, resulting in missing data for variables related to internal collusion. These reports were then excluded, leaving 460 annual reports. Based on time, the research sample was divided into a training set and a test set: fraudulent annual reports from 2009 to 2016 served as the training set, and fraudulent annual reports from 2017 to 2020 served as the test set.

[0074] In this embodiment, the sampling matching method is to match fraudulent annual reports with normal samples. The research sample training set contains 270 fraudulent annual reports and 270 corresponding normal annual reports.

[0075] Step 3: Construct a basic prediction model based on logistic regression.

[0076] Step 3.1: Constructing the base prediction model based on the logistic regression, the specific expression is as follows:

[0077]

[0078] In the above formula, f(X) is the response variable, indicating the probability of occurrence of financial fraud events of listed companies, is the observation variable under each view; D k represents the number of observation variables of each view; k represents the number of view angles of the divided feature data, and N represents the number of observation values in the sample period; is the coefficient of the feature data under each view in the model.

[0079] Step 3.2: Apply the consistency constraint function g(w) expressed by the collaborative regularization to each two views in different views, and the specific calculation formula is as follows:

[0080]

[0081] λ m As a hyperparameter is to make the consistency constraint function and the loss function of the base model have comparable order of magnitude. In order to make the consistency between views comparable, the hyperparameters λ m are set to be the same size.

[0082] In step 1, the feature data is divided into four views according to the dimension of the data, so there are 6 combination forms in two combinations.

[0083] Step 4: This embodiment adopts the gradient descent method to optimize the parameters of the multi-view learning model with consistency constraint, and the specific method is as follows:

[0084] Step 4.1: The loss function J(w) of the model constructed in this embodiment is h(w)+g(w), wherein the first term is the negative log-likelihood term, which is used to calculate the error h(w) between the true value and the predicted value, and the regular term after that is the penalty term g(w) of the consistency constraint, which evaluates the consistency of the prediction results between different views. The smaller the prediction error is, and the more consistent the prediction results between different views are, the smaller the loss function is, and the specific calculation formula is as follows:

[0085]

[0086] Step 4.2: The gradient descent method is used to solve the objective function, and the gradient calculation formula of the first error term of the objective function is as follows:

[0087]

[0088] wherein,

[0089] The gradient calculation formula of the second part of the penalty term is as follows, wherein D p and D q respectively represent V p and V q characteristic view angles.

[0090]

[0091] Based on the gradient calculation formula of the above two parts, the global update strategy of the parameter w is as follows, wherein:

[0092]

[0093] Step 4.3: Two stopping iteration standards are adopted in this embodiment: one is the threshold of the weight change of the adjacent two iterations, and the other is the maximum number of iterations. λ m As a hyperparameter is to make the penalty term and the likelihood have comparable orders of magnitude, and to ensure that the loss functions under different view angle combinations are comparable. In order to be simple, in this embodiment, the parameters λ m of all terms in the regularization term are set to be equal, and λ m = 10 -3 is set according to the optimal out-of-sample performance F value.

[0094] The optimal multi-view learning model with consistency constraint obtained through the parameter optimization in step 4 is the optimal model obtained in this embodiment.

[0095] Step 5: Compare the loss function values and the decomposition error term and penalty term results when the multi-view learning algorithm with consistency constraint stops iterating on the training set samples, as shown in the second row in Table 1. Figure 3 From the penalty term brought by each constraint, the internal collusion condition dimension and the external collusion condition dimension have the smallest prediction difference, indicating that the two are the two view angles with the highest recognition result consistency. In addition, the prediction difference of the motivation dimension and the internal collusion and external collusion is also small, indicating that it also has good prediction consistency.

[0096] Step 6: In addition to constructing a multi-view learning model with consistency constraints on all four view angles, in order to further compare and analyze more clearly, three view angles are selected to impose consistency constraints (there are 4 combination forms in total), evaluate the consistency of the constructed multi-view learning model, and compare and analyze the prediction effects of different models. As shown in Table 2. Figure 3As shown in the results, the consistency constraint of reducing the accounting manipulability fraud and other dimensions reduces the model loss function compared to the loss function obtained by applying constraints to all four dimensions. The above results also prove the conclusion that the motivation dimension obtained in step 5 has good predictive consistency with internal collusion, external collusion and each other. Comparing the sample identification effect of the three models with all constraints, only constraints on the remaining three dimensions except for the fraud means and the three models without constraints on the test set, the results show that the multi-view learning model with constraints on the remaining three dimensions except for the fraud means has better recall rate and F1 value, reaching 0.801 and 0.757 respectively. This shows that the fusion of different dimensions, especially the fusion of three dimensions except for the financial fraud means, improves the prediction ability of the model. The above results also show that there is a strong correlation between the internal collusion condition dimension and the external collusion condition dimension.

[0097] In summary, the multi-view learning method based on collaborative regularization proposed in the embodiment can fully integrate the knowledge of each view, so that the multi-view learning model can more accurately extract key information, thereby improving the prediction ability of the model. At the same time, the model dimension prediction consistency evaluation method based on multi-view learning proposed by the present application can effectively and quantitatively evaluate the consistency between different views in the model expressed by a multi-dimensional index system.

[0098] Embodiment 4: Consistency evaluation of centrifugal compressor mechanical seal failure prediction in a refinery

[0099] Step 1: Multi-dimensional data set construction and preprocessing

[0100] Three years of operation data (total sample size 18,000) were collected from five centrifugal compressors of the same type, and were divided into four dimensions according to the monitoring principle:

[0101] Vibration monitoring dimension (V1): 9 features (such as radial vibration root mean square value, axial vibration harmonic distortion rate) were extracted by a three-axis acceleration sensor

[0102] Temperature monitoring dimension (V2): 7 features (such as sealing end face temperature difference gradient, cooling oil temperature rise rate) were generated by an infrared thermocouple

[0103] Process parameter dimension (V3): 11 parameters (such as inlet pressure fluctuation standard deviation, lubricating oil viscosity) were obtained by a SCADA system

[0104] Acoustic emission dimension (V4): 5 features (such as 150-200 kHz event count rate) were extracted by a high-frequency acoustic sensor

[0105] Preprocessing: wavelet packet decomposition was performed on V1 to reduce noise, and environmental temperature compensation was performed on V2.

[0106] Step 2: Base prediction model training

[0107] XGBoost classifier is chosen as the base model, input 32-dimensional feature vector, mechanical seal failure as the prediction target (label from maintenance records). Training set 12,000 samples (positive samples: negative samples = 1.5:1), parameter settings: tree depth 8, learning rate 0.05, the optimal number of iterations 120 times determined by 10-fold cross-validation.

[0108] Step 3: Build multi-view learning model (corresponding to S2-S3)

[0109] S2: Two-by-two combination of view constraints

[0110] Two-by-two combination of four-dimensional data to build six sets of constraints (V1-V2, V1-V3, V1-V4, V2-V3, V2-V4, V3-V4), based on Euclidean distance to build consistency constraint function as a penalty term added to the base model loss function. Use Adam algorithm optimization, hyperparameter λ m Determined by grid search (initial learning rate 0.002, decay 10% every 50 iterations), stop condition: continuous 10 iterations weight change <10^-6 or total iterations 400 times, obtain two-by-two combination of multi-view model.

[0111] S3: Single view / multi-view selective constraints

[0112] Single view constraint example: select V1 dimension alone to apply constraints, only build V1 internal feature consistency constraints, optimize parameters according to S2 method

[0113] Three-view combination constraint example: select V1+V2+V4 dimension to apply constraints, build three sets of two-by-two constraints (V1-V2, V1-V4, V2-V4), generate multi-view model according to S2 method

[0114] Step 4: Model performance evaluation

[0115] Cross-scenario verification

[0116] Design three types of test sets:

[0117] 1. Test set 1: same factory area, untrained equipment (time span 14 months, simulate aging)

[0118] 2. Test set 2: high sulfur working condition equipment (feature covariance offset 31.2°, test corrosion adaptability)

[0119] 3. Test set 3: overload operation data (recently collected, verify sudden working condition change response)

[0120] Dynamic weight distribution: time decay function, 3-month data weight raised to 0.72, calculate weighted average indicators (precision, recall, F1 score).

[0121] Adversarial robustness test

[0122] FGSM algorithm is used to generate perturbation test set (perturbation amplitude = feature standard deviation x 17%), gradient backtracking is triggered when single indicator deviation rate > 11% or comprehensive deviation rate > 8%. Example: recall rate decreased by 14.7% under V4 dimension perturbation, located to the second convolution layer gradient saturated neuron.

[0123] Step 5: output evaluation results

[0124] Constraint value and performance indicator extraction

[0125] Two constraints are combined: V1-V4 constraint value 0.128, V2-V3 constraint value 0.201

[0126] Multi-view model performance: weighted average F1 of full-dimensional model = 0.812 ± 0.021

[0127] Deep analysis and rectification

[0128] 1. Bootstrap confidence interval: stratified resampling 1000 times to generate 95% confidence interval (e.g. V1-V2 constraint interval [0.116, 0.153])

[0129] 2. Distribution alignment rectification: F1 of test set 3 = 0.748 deviates from the historical interval [0.795, 0.832], start Bayesian posterior optimization (Gaussian process regression), after 19 iterations, KL divergence is reduced from 0.071 to 0.043

[0130] Final evaluation conclusion

[0131] Dimension consistency ranking (penalty term mean): V1-V4 (0.128) > V2-V3 (0.201) > V1-V3 (0.235)

[0132] Model optimization suggestion: after removing the weakest V3 dimension, F1 is improved to 0.837 ± 0.017, output composite evaluation document containing gradient backtracking report and confidence interval.

[0133] Although the embodiments of the present application have been disclosed as above, they are not limited to the use listed in the specification and embodiments, and can be fully applied to various fields suitable for the present application, and additional modifications can be easily realized by those skilled in the art, therefore the present application is not limited to specific details and examples shown and described herein, without departing from the general concept defined by the claims and equivalent scope.

Claims

1.A method for evaluating consistency of model dimension prediction based on multi-view learning, characterized in that, The method comprises the following steps: S1: obtaining a basic prediction model and multi-dimensional feature data; S2: combining the multi-dimensional feature data two by two in turn to construct a consistency constraint function based on the Euclidean distance of the two feature data, adding all the constraint functions obtained to the loss function of the basic prediction model as a penalty term, and performing parameter optimization to obtain a multi-view learning model; S3: selecting two or more dimensional feature data to apply constraints, and constructing a multi-view learning model according to the method of S2; S4: evaluating the prediction performance indicators of the multi-view learning models obtained in S2 and S3 on the test set, and the prediction performance indicators include accuracy, precision, recall and F1 score; S5: extracting the constraint values of the multi-view learning model in S2 and the prediction performance indicators of the multi-view learning model in S4, and outputting the evaluation results. 2.The multi-view learning based model dimension prediction consistency evaluation method of claim 1, wherein, In S2, the calculation formula of the consistency constraint function g(w) is: k represents the number of dimensions, is the combination number, and are the feature data of different dimensions, respectively, and are the coefficients corresponding to the feature data, respectively, and λ m is a hyperparameter used to make the consistency constraint function and the loss function of the base prediction model comparable in order of magnitude. 3.The method of claim 2, wherein, In S2, the target function J(w) is constructed for parameter optimization, and the specific calculation formula is: Where h(w) is the loss function of the basic prediction model. 4.The method of claim 3, wherein, The gradient descent method is used to solve the target function, and in the solving process, when the weight change of adjacent two iterations is less than a predetermined threshold, or the maximum iteration number is reached, the iteration is stopped. 5.The method of claim 1, wherein, In S4, it further comprises: At least three independent test sets with significant distribution difference are used to verify the same multi-view learning model across scenes, and the distribution difference is quantitatively confirmed by a data acquisition time span threshold or a feature covariance offset angle; The weighted average value and the standard deviation of each performance indicator on multiple test sets are calculated, wherein the weight of the test set is dynamically adjusted according to the data freshness, and the recent data obtains a higher evaluation weight. 6.The method of claim 5, wherein, In S4, the distribution difference of the independent test set is quantitatively confirmed by a preset data acquisition time span threshold or a feature covariance offset angle, wherein the time span threshold is greater than or equal to 6 months, and the covariance offset angle is obtained by calculating the similarity of the feature covariance matrix; The weight distribution of the weighted average value adopts a time decay function, and the test set weight is negatively correlated with the data acquisition time, and the specific weight value is an exponential decreasing function of the difference value between the current time and the data acquisition time. 7.The multi-view learning based model dimension prediction consistency evaluation method of claim 1, wherein, In S5, it further comprises: FGSM fast gradient symbol attack or PGD projection gradient descent adversarial algorithm is used to generate a perturbation test set, wherein the perturbation amplitude is controlled in the interval of 15%-20% of the feature value standard deviation, the absolute deviation mean of the accuracy, precision, recall and F1 score of the original test set and the perturbation test set is calculated, a double trigger mechanism of a single indicator deviation rate threshold of 10% and a comprehensive deviation rate threshold of 8% is set, when any indicator deviation rate breaks through 10% or the comprehensive deviation rate exceeds 8%, the gradient distribution data in the historical training log of the last three months is automatically called, the gradient backtracking analysis based on the reverse tracking of the activation function is performed, the sensitive neuron layer causing the robustness defect is located, and an evaluation review report containing the adversarial sample vulnerability distribution map is generated in parallel. 8.The method of claim 7, wherein, S5 further comprises: Bootstrap resampling 1000 times to construct the 95% confidence interval of the constraint value, the sampling process introduces stratified sampling strategy to ensure the uniformity of each feature dimension coverage, when the F1 score optimal value deviates from the confidence interval in cross-scene verification, automatically extract the constraint value sequence, performance index matrix and feature covariance tensor in the last three historical evaluations, construct the Bayesian posterior distribution model with Gaussian process regression as the core, perform distribution alignment and correction by calculating the KL divergence between the current evaluation result and the posterior distribution, where the KL divergence threshold is set to 0.05 and the maximum iteration number is 50, the correction process synchronously optimizes the L2 norm stability of the feature weight vector, finally outputs the corrected constraint value distribution curve and the weight vector convergence diagnostic report. 9.The method of claim 8, wherein, The evaluation result output by S5 is input into the adversarial sample generation process, when the comprehensive deviation rate of the original test set and the perturbed test set does not exceed 8%, the evaluation result is directly output, if the deviation rate exceeds the threshold, gradient backtracking analysis is started; The results of gradient backtracking analysis are input into the Bootstrap confidence interval construction process to perform correction in the form of KL divergence distribution alignment, and the corrected constraint value covers the corresponding value in the original evaluation result of S5; Finally, a composite evaluation report containing the gradient backtracking analysis report and the Bootstrap confidence interval data is generated, and the corrected constraint value and the cross-scene performance index are marked.

Images