User financing behavior analysis method based on artificial intelligence and financial big data

By applying artificial intelligence technologies such as correlation measurement quantum state analytical network, dynamic niche neural optimization and manifold learning in financial data processing, the problems of poor generalization ability and information loss in traditional methods when processing complex financial data are solved, and the problem of the traditional method is achieved. Higher data feature capture and risk assessment accuracy.

CN118964933AInactive Publication Date: 2024-11-15GUANGZHOU DEXU TECHNOLOGY CO LTD

Patent Information

Application Number
CN202411003472.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

When traditional financial data processing methods process complex, high-dimensional and unstructured financial data, there are problems such as poor generalization capabilities, low processing efficiency, and difficulty in capturing deep-level data relationships, resulting in inaccurate risk assessment or high risk of missed detection. Trading behavior.

Method used

Using artificial intelligence-based methods, including generative adversarial networks with correlation metric quantum state analysis, dynamic niche neural optimization algorithm, manifold learning auto-coded neural networks and mixed arrangement entropy random forest algorithms to enhance data feature capture capabilities and adjust Neural network parameters, preserve data intrinsic structure and improve the accuracy and robustness of the model in complex data processing.

Benefits of technology

It improves the generalization ability and accuracy of the model, avoids the problem of gradient disappearance and explosion, effectively retains the key information of the data, enhances the processing ability of complex financial data, and improves the accuracy of risk assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004962309230000041
    Figure BDA0004962309230000041
  • Figure BDA0004962309230000043
    Figure BDA0004962309230000043
  • Figure BDA0004962309230000051
    Figure BDA0004962309230000051
Patent Text Reader

Abstract

A user financial management behavior analysis method based on artificial intelligence and financial big data belongs to the field of financial management, and comprises the following steps: S1, data acquisition and manual annotation; s2, performing sample generation by adopting a generative adversarial network algorithm based on correlation quantum state analysis, and further realizing data expansion; s3, inputting the expanded data into a feature extraction model to train the feature extraction model; s4, inputting the data subjected to feature extraction into a feature dimension reduction model for training of the feature dimension reduction model; s5, inputting the data after dimension reduction into a classifier model for classifier model training; and S6, analyzing the financial management behavior of the user. According to the method, a data sample with higher quality can be generated, the generalization ability and accuracy of the model are improved, the training stability and the model performance are also improved, the accuracy and efficiency of a subsequent classification model are improved, and the processing ability of the model to complex financial data is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of financial management, and specifically relates to a user financial management behavior analysis method based on artificial intelligence and financial big data. Background Art

[0002] With the expansion of financial markets and the increase in trading activities, financial institutions are facing increasingly complex data processing and analysis challenges. Especially in the analysis of financial behavior, traditional data processing methods are often unable to effectively process and analyze large-scale, high-dimensional and unstructured financial data due to technical and methodological limitations. In addition, accuracy and predictability in financial behavior analysis are crucial and directly related to risk control and decision support. Traditional methods are often prone to problems such as poor model generalization, low processing efficiency, and difficulty in capturing deep data relationships when processing complex data. These problems may lead to inaccurate risk assessment or missed detection of high-risk trading behaviors.

[0003] The Chinese invention patent with application number CN202410249459.9 proposes a method, processor and storage medium for recommending financial content. The method includes: obtaining the user's buried data in the financial scenario; inputting the buried data into a preset funnel analysis model to obtain the user interest corresponding to multiple financial content categories; determining the target financial content category in multiple financial content categories based on multiple user interests; determining the target financial content related to the target financial content category in the pre-stored financial content library; and recommending the target financial content to the user. This application scheme determines the user's interest in multiple financial categories in the financial scenario based on the buried data and the funnel analysis model, and determines the recommended content, so as to realize accurate recommendation of financial content to users under the financial vertical category. The Chinese invention patent with application number CN202311776426.1 proposes a method and system for user transaction feature analysis and early warning, the method comprising obtaining the target user's financial loan agreement to determine the target user's loan amount and interest and fee settlement rules; obtaining the target user's transaction details information, calculating the loan utilization rate, preferential merchant consumption rate, average interest-free period rate, and interest payment fund rate for each settlement cycle; obtaining the target user's income information and asset declaration information to analyze the target user's reasonable consumption amount and reasonable preferential merchant consumption amount, and then calculating the reasonable expenditure correction rate and reasonable preferential correction rate; calculating the transaction risk characteristic value Z=KA(AA-0)+KB(BB-0)+KC(CC-0)-KD(DD-0)+N; when Z>Z0, generating transaction risk early warning information; this application has the effect of reducing the economic losses caused by abnormal transaction behavior. The Chinese invention patent with application number CN201811558485.0 proposes a method, device, storage medium and electronic device for processing the automated test results of a financial data analysis system. The method includes: obtaining the test index value before the construction code of the financial data analysis system is changed, and storing the test index value in a buffer; obtaining the business data processing result obtained by the financial data analysis system; in response to the user's test request after the construction code of the financial data analysis system is changed, the financial data analysis system is automatically tested to obtain the automated test result, and the automated test result is stored in the buffer; when it is detected that the business data processing result is sent by email, the test index value and the automated test result are withdrawn from the buffer and displayed in a manner that is only visible to the user. The technical solution of the present invention can improve the security of test result processing.

[0004] The above technical solution still has the following problems that need to be further solved:

[0005] 1. Traditional generative adversarial networks may not be able to fully capture and reproduce complex data features in specific financial data scenarios.

[0006] 2. In feature extraction and classification models, common neural network training is susceptible to gradient vanishing or exploding, which limits the effectiveness and reliability of the model.

[0007] 3. Existing dimensionality reduction techniques may fail to effectively preserve the intrinsic structure of the data, resulting in information loss and affecting the final analysis accuracy.

[0008] 4. Traditional classification models may lack sufficient accuracy and robustness when dealing with high-dimensional and complex data. Summary of the invention

[0009] The purpose of the present invention is to propose a user financial management behavior analysis method based on artificial intelligence and financial big data. The generative adversarial network based on association metric quantum state analysis adopts quantum state to enhance the data feature capture ability of the generator; the dynamic niche neural optimization algorithm dynamically adjusts the parameter configuration of the neural network to find the optimal solution; the autoencoder neural network based on manifold learning adopts the manifold learning principle to enable the network to better retain the intrinsic structure and important characteristics of the data during dimensionality reduction; the random forest algorithm based on mixed permutation entropy adopts mixed permutation entropy to optimize the split point selection of the decision tree, thereby enhancing the performance and accuracy of the model when processing complex and nonlinear data.

[0010] The technical solution adopted by the present invention is:

[0011] The user financial management behavior analysis method based on artificial intelligence and financial big data includes the following steps:

[0012] S1. Data collection and annotation

[0013] Data is collected from financial databases, including financial transaction records, user profiles, credit reports, and historical financial management behavior data. The collected data is stored in a structured data format, specifically JSON format; the collected data is manually annotated;

[0014] S2. Data expansion

[0015] A generative adversarial network algorithm based on correlation metric substate analysis is used to generate samples, thereby realizing data expansion; the generative adversarial network based on correlation metric substate analysis includes two components: a generator (G) and a discriminator (D); the generator is responsible for creating new data that is almost indistinguishable from real data, and the discriminator's task is to distinguish whether the input data comes from the original data set or is generated by the generator. The present invention adopts a correlation metric substate analysis strategy, and by simulating the interaction between the essential correlation of data and quantum states, the generator's ability to capture data features is enhanced.

[0016] Specifically, the training process of the generative adversarial network algorithm based on association metric sub-state parsing is as follows:

[0017] S201, initialize the network parameters of the generator and the discriminator, specifically, the network parameters θ of the generator G and the discriminator D g and θ d Initialize randomly from the prior distribution, expressed as:

[0018]

[0019] In the formula, represents a normal distribution with a mean of 0 and a covariance of the unit matrix; θ g and θ d are the parameter vectors of the generator and the discriminator respectively;

[0020] S202, before starting adversarial training, pre-train the discriminator separately, and the discriminator D uses the real data set {x (i)} is trained, the goal is to maximize the probability of correctly classifying real data and generated data, and the loss function adopts the cross entropy loss function, which is expressed as:

[0021]

[0022] Where, L d is the loss function of the discriminator, n is the number of samples of real data, x (i) is the i-th real data sample, D(x (i) θ d ) is the output of the discriminator for sample i;

[0023] S203. In the adversarial training phase, the generator attempts to generate new data samples and adjusts parameters through the feedback of the discriminator to optimize the generated data samples so that the discriminator cannot distinguish between real and generated data. The loss function of the generator is defined as:

[0024]

[0025] Where, L g is the loss function of the generator, m is the number of generated samples, z (j) is the jth random noise vector, G(z (j) θ g ) is the parameter θ g The generator at the time is based on the noise z (j) The generated samples, G(z (j) θ g ) is calculated as:

[0026] G(z (j) θ g )=tanh(W g ·z (j)+b g )

[0027] Where W g is the weight matrix of the generator, b g is the bias term, z (j) is the input random noise vector, tanh() is the Tanh activation function;

[0028] S204. In each optimization update of the generator and the discriminator, the update of the generator and the discriminator will take into account the correlation metric quantum state of the data, and further enhance the expression of data features by simulating the interaction of quantum states; the correlation metric quantum state analytical function is used to simulate the interaction of quantum states, which is expressed as:

[0029]

[0030] Where Q(G,D) is the correlation metric substate analysis function, λ k is the weight of the kth quantum state, z k is the noise represented by the corresponding quantum state; Sig(D,G(z k θ g )) is the Sigmoid activation function;

[0031] Sig(D,G(z k θ g )) is calculated as:

[0032]

[0033] S205. In the process of optimizing the discriminator, the confidence estimation loss function C(D) is used to evaluate the confidence of the discriminator for each sample, so as to adjust the learning strategy of the discriminator, which is expressed as:

[0034]

[0035] Where C(D) is the confidence estimation loss function, γ bg is the adjustment factor, is the confidence weight for each sample, indicating the confidence of the discriminator in the jth generated sample;

[0036] S206, repeating iterative steps S201-S205 until a preset stop iteration condition is met, indicating that the model training is completed; after the data expansion model training is completed, the trained data expansion model is used to increase the number of samples;

[0037] S3. Feature extraction model training

[0038] The expanded data is input into the feature extraction model for training, and a 3-layer fully connected neural network is used for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion, or falling into the local optimal solution may occur, affecting the stability of training and the performance of the model. Specifically, a neural network model based on the dynamic niche neural optimization algorithm is used. Inspired by the niche theory in ecology, each species seeks its optimal position in the ecosystem to achieve an advantage in the competition for survival. Similarly, the dynamic niche neural optimization algorithm aims to find the optimal configuration of neural network parameters so that the network performs best on a given task.

[0039] Specifically, the training process of the neural network algorithm based on the dynamic niche neural optimization algorithm is as follows:

[0040] S301. In the initialization phase, a diverse primary population is generated. Each individual represents a set of neural network parameters. Let P (0) is the initial population, each individual p i is a parameter vector containing the weights and biases of the neural network, expressed as:

[0041]

[0042] Where Nu is the population size, W i is the weight matrix, b i is the bias vector, Represents direct concatenation of vectors;

[0043] S302, in the dynamic evaluation stage, the performance of each individual is evaluated according to the attractor properties of the dynamic trajectory. The dynamic evaluation method is implemented by using the fitness function f(), which is expressed as:

[0044] f(p i )=α Ts ·Acc(p i )+β Ts Stab(p i )

[0045] In the formula, Acc(p i ) represents the accuracy of the network on the labeled training data, which is calculated by the preset Softmax output; Stab(p i ) represents the dynamic stability of the parameters; α Ts and β Ts They are performance hyperparameters and stability hyperparameters respectively;

[0046] Dynamic stability of parameters Stab(p i ) is calculated as:

[0047]

[0048] In the formula, δ sta is the attenuation coefficient; is the loss function of the network, w ij is an individual p i is the jth weight parameter in , Mu is the total number of weights;

[0049] S303, calculate the loss function of the network for the next iteration, define κ(t) as the regularization strength of the t-th iteration, and κ(t+1) as the regularization strength of the t+1-th iteration. The calculation method is expressed as:

[0050] κ(t+1)=κ(t)·exp(-α pe g(t)+β pe ·(1-L(t)))

[0051] g(t)=(f(p (t) )-f(p (t-1) ))·exp-γ pe ·Ds(p (t) )

[0052]

[0053] In the formula, α pe and β pe They are the fitness gradient hyperparameter and the learning progress hyperparameter respectively; g(t) is the fitness gradient, which measures the rate of change of the fitness function over time; L(t) is the learning progress, which represents the improvement of the performance of the network on the training data set at each iteration; γ pe is the fractal dimension adjustment factor; Ds() is the fractal dimension calculation function;

[0054] S304. Individuals move in the ecological niche according to their fitness. Excellent individuals occupy the dominant ecological niche, and weaker individuals explore new ecological niches or are eliminated. The adjustment method is expressed as:

[0055]

[0056] In the formula, f avg is the average fitness of the current population, η vd is the interaction level adjustment factor, p best is the parameter of the current optimal individual, is the loss function About the i-th layer input The partial derivative of γ vd and λ vd They are the update step size hyperparameter and the learning strategy hyperparameter;

[0057] S305, perform crossover and update operations, and the new individual inherits the characteristics of the two excellent individuals, expressed as:

[0058] p new =ζ kg ·p parent1 +(1-ζ kg )·p parent2

[0059] In the formula, ζ kg is a randomly generated crossover factor between 0 and 1; p parent1 and p parent2 are two randomly selected parent individuals;

[0060] S306, performing mutation and mutation operations, using a mutation mechanism based on nonlinear dynamics theory, so that individual parameters can perform random jumps in a larger range, expressed as:

[0061] p′ i =p i +∈ bg ·randn() ·mask()

[0062] mask()=rand(size(p i ))<μ bg

[0063] In the formula, ∈ bg is the mutation intensity, μ bg is the probability of mutation, randn() generates a normally distributed random number; mask() generates a random number with p i A random binary mask of the same size, controlling which parameters will mutate; size(p i ) is the length measurement function;

[0064] S307, perform global dynamic update, the entire population is updated according to the theory of ecological dynamic system, the strategies of excellent individuals will be imitated and spread, and the unsuitable strategies will gradually disappear, expressed as:

[0065] P (t+1) = sort(P (t) ,f)

[0066] In the formula, the sort function sorts the population according to the fitness f(), retains individuals with high fitness, and eliminates individuals with low fitness;

[0067] S308, repeating the iteration steps S301-S307 until the preset stop iteration condition is met, indicating that the model training is completed;

[0068] S4. Feature dimensionality reduction model training

[0069] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model, and the feature dimensionality reduction is performed using an autoencoder neural network algorithm based on manifold learning. The autoencoder neural network based on manifold learning includes two parts: an encoder and a decoder. The function of the encoder is to map the high-dimensional input features to a low-dimensional latent space, and the decoder attempts to reconstruct the original input from this low-dimensional latent space. The present invention adopts the manifold learning principle to enable the network to better understand and retain the intrinsic structure of the data, and at the same time adopts a mask transformation mechanism to enhance the robustness of the model to noise and outliers in the input data.

[0070] The training process of the autoencoder neural network algorithm based on manifold learning is as follows:

[0071] S401, initialize the weights of the encoder and decoder, and the initialization method is random initialization, which is expressed as:

[0072]

[0073] Where W e and W d are the weight matrices of the encoder and decoder respectively; b e and b d is the corresponding bias vector; σ 2 represents the standard deviation of initialization;

[0074] S402: Sampling the manifold structure from the training data to estimate the local manifold structure of the data is performed by the k-nearest neighbor algorithm, that is, using the k-nearest neighbor method to sample each data point x i Sampling is performed and the average distance of its k nearest neighbors is calculated, which is expressed as:

[0075]

[0076] In the formula, x i Represents the current data point; NN k (x i ) is x i The k nearest neighbors of ; ∥x i -x j ∥ 2 is x i and x j The square of the Euclidean distance between them;

[0077] S403: Generate a random mask for each input sample, where the mask is used to simulate the partial loss of input data. Specifically, for each input sample x i , generate a random mask m i The way is expressed as:

[0078] m i = Bernoulli(p)

[0079] In the formula, m i is applied to x i The mask of , each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, p is the probability that the mask element is 1;

[0080] S404: After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i , calculated as:

[0081] z i =Sig(W e ·(1+γ csd β csd )·(x i ⊙m i )+b e )

[0082] Where ⊙ represents the corresponding multiplication of elements; Sig() is the Sigmoid activation function; β csd is the dynamic coding adjustment coefficient; γ csd is the preset scaling factor;

[0083] The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data. Specifically, the skewness S(x i ) and kurtosis K(x i ), expressed as:

[0084]

[0085] In the formula, μ pd and σ pd x i The mean and standard deviation of

[0086] Furthermore, the adjustment coefficient β csd The calculation method is expressed as:

[0087]

[0088] Where tanh() is the hyperbolic tangent function;

[0089] S405. In the latent space, the manifold constraint is used to adjust the low-dimensional representation to ensure that the geometric and topological characteristics of the data are preserved. The calculation method of the manifold constraint loss is expressed as:

[0090]

[0091] In the formula, zi and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss, fidelity() is the quantum state fidelity loss function, λ p is the regularization parameter;

[0092] S406, reconstruct the input data from the low-dimensional representation through the decoder, calculate the reconstruction error and the manifold constraint error, and update the network parameters through back propagation. The total loss function Lur is composed of the reconstruction error and the manifold constraint error, which is expressed as:

[0093]

[0094]

[0095] In the formula, is the reconstructed data point; kc is the regularization parameter of the manifold constraint;

[0096] S407, using the gradient descent method to optimize the network parameters and continuously reduce the total loss, the method of updating the parameters is expressed as:

[0097]

[0098] In the formula, α gk is the learning rate;

[0099] S408, repeating the iteration steps S401-S407 until the preset stop iteration condition is met, indicating that the model training is completed;

[0100] S5. Classifier model training

[0101] The reduced-dimensional data is input into a classifier model for classifier model training, and a random forest based on mixed permutation entropy is used as a classification algorithm, wherein the random forest based on mixed permutation entropy includes a plurality of decision trees, and each tree uses a randomly selected data subset and feature subset during training. The mixed permutation entropy of the present invention is used to evaluate and select the optimal split point, thereby enhancing the robustness and accuracy of the model when processing complex financial behavior data.

[0102] Specifically, the training process of the random forest algorithm based on mixed permutation entropy is as follows:

[0103] S501. For each tree, randomly select a data subset and a feature subset before training to ensure the diversity and robustness of the model. For each tree t, the selected data subset D t and feature subset F tIt is obtained by random sampling and is expressed as:

[0104] D t =sample(D,size=s D ,replace=false)

[0105] F t =sample(F,size=s F ,replace=false)

[0106] In the formula, D represents the complete training data set, F represents the complete feature set, and s D and F Respectively represent the number of extracted data and features, and the sample() function performs a random sampling operation;

[0107] S502: For each decision node, the hybrid permutation entropy H is used to select the best segmentation feature f and segmentation point θ, and then the data is distributed to the left and right child nodes according to the selected segmentation point, which is expressed as:

[0108]

[0109] and,

[0110] D left ={x∈D t |x f ≤θ}

[0111] D right ={x∈D t |x f >θ}

[0112]

[0113] In the formula, x f represents the value of sample x on feature f, p c is the probability of class c in data set D, φ(x i ,x j ) represents the data point x i and x j The relative ranking contribution between ;

[0114] φ(x i ,x j ) is calculated as:

[0115]

[0116] In the formula, ∥x i -x j ∥ 2 Represents x i and xj The square of the Euclidean distance between tr is a hyperparameter that controls the range of similarity influence, λ tr is the weight coefficient that balances the traditional entropy and permutation contributions;

[0117] S503, after the tree is fully grown, an independent validation data set is used to evaluate the performance of the tree, and then the structure of the tree is dynamically adjusted according to the preset performance index, and branches that do not provide additional information gain are removed; specifically, for the t-th tree, its performance Perf(t) on the independent validation set V is evaluated, and pruning is performed; the performance evaluation function Perf(t) is implemented by ensemble learning, and is expressed as:

[0118]

[0119] In the formula, ω(x, t) represents the weight of sample x in tree t, which is dynamically adjusted based on the contribution of the sample; I(t(x)=y x ) is an indicator function, which takes the value of 1 when the tree t predicts the sample x correctly, otherwise it takes the value of 0;

[0120] S504, repeat steps S501-S503, construct multiple decision trees, integrate the multiple trees, and use the majority voting method to make the final classification decision, which is expressed as:

[0121] y pred =mode({t(x)|t∈T})

[0122] Where t(x) is the prediction of tree t for sample x, T is the set of all trees, and mode() is the decision function that takes the majority;

[0123] S6. Analysis of user financial management behavior

[0124] The trained model is used to process new sample data, so as to realize the analysis and prediction of user financial management behavior. Specifically, the new input sample is firstly obtained through the feature extraction model to obtain the extracted features, and then the extracted high-dimensional feature vector is input into the data dimension reduction model to reduce the feature dimension; finally, the feature vector after dimension reduction is input into the classifier model to predict the type of financial management behavior.

[0125] Furthermore, in step S203, the discriminator is trained based on the samples provided by the generator and the real samples to optimize the recognition ability of the discriminator and improve the accuracy of distinguishing the true from the false. The loss function of the discriminator is defined as:

[0126]

[0127] In the formula, n and m represent the number of real samples and generated samples respectively, x (i)represents the i-th true sample.

[0128] Furthermore, in step S203, during the optimization process of the discriminator, the confidence weight β of each sample is j The calculation method is expressed as:

[0129]

[0130] In the formula, ∥D(G(z (j) θ g ))-y real ∥ 2 Represents the generated sample and the real sample label y real The square of the Euclidean distance between them, ∥∥ represents the Euclidean distance, that is, the L2 norm.

[0131] Furthermore, in step S303, the loss function of the network The calculation method is expressed as:

[0132]

[0133] Where R(W) is the L2 norm of the weight vector W, which is used to calculate the sum of the squares of the weights; is the cross entropy loss.

[0134] Further, in step S304, the i-th layer input It is obtained through adaptive normalization, and the calculation method is expressed as:

[0135]

[0136] In the formula, x i is the i-th layer input before normalization, μ B and are the mean and variance of the batch input data respectively; γ zs and β zs is a learnable parameter, specifically the first learnable parameter and the second learnable parameter; ∈ zs is a very small constant;

[0137] Adjust γ based on the best performing individual in the current population zs and β zs , to better adapt to the characteristics of the training data, the adjustment method is expressed as:

[0138] γ zs (t+1) =γ zs (t) ·exp(-η γ ·Var(p (t) ))

[0139]

[0140] In the formula, Var(p (t) )and are the variance and mean of the parameter vector of the current batch, γ zs (t) and β zs (t) are the first learnable parameter and the second learnable parameter of the current batch, γ zs (t+1) and β zs (t+1) are the first learnable parameter and the second learnable parameter for the next batch iteration, μ target is the target mean, η γ and η β is the learning rate.

[0141] Further, in step S403, the Bernoulli distribution is a discrete distribution used to simulate a random variable with one of two possible results. The Bernoulli random variable m ij The generation method is expressed as:

[0142]

[0143] Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij =1; if U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0.

[0144] Furthermore, in step S405, the calculation method of the quantum state fidelity loss function fidelity() is expressed as:

[0145] fidelity p ,q p )=1-| <z p |q p >| 2

[0146] In the formula, | <z p |q p >| 2 Represents the quantum state q p With the original encoding state z p The square of the modulus of the inner product between;

[0147] The original encoding state z p The encoder takes the input feature vector x i Converted to a low-dimensional quantum state representation z through an encoderp , expressed as:

[0148]

[0149] In the formula, x p is the input feature vector, z p is the encoded low-dimensional representation, is the encoder function; are the parameters of the encoder, including W e and b e .

[0150] Further, in step S503, ω(x, t) is adjusted according to the error rate of the sample, which is expressed as:

[0151] ω(x,t)=exp(-α t I(t(x)≠y x ))

[0152] In the formula, α t is the performance parameter of tree t in the current training iteration, which is proportional to the error rate of the tree.

[0153] Beneficial effects of the present invention:

[0154] 1. By generating adversarial networks with associated metric sub-states, higher quality data samples can be generated, thereby improving the generalization ability and accuracy of the model.

[0155] 2. The dynamic niche neural optimization algorithm avoids common problems in neural network training, such as gradient problems, by dynamically evaluating and adjusting parameters, improving training stability and model performance.

[0156] 3. Manifold learning enhances the ability of the autoencoder network, enabling it to effectively retain key information during dimensionality reduction, thereby improving the accuracy and efficiency of subsequent classification models.

[0157] 4. The use of mixed permutation entropy makes the random forest model more accurate in selecting split points and enhances the model's ability to handle complex financial data. DETAILED DESCRIPTION

[0158] The user financial management behavior analysis method based on artificial intelligence and financial big data includes the following steps:

[0159] S1. Data collection and annotation

[0160] Data is collected from a financial database, including financial transaction records, user profiles, credit reports, and historical financial management behavior data. The collected data is stored in a structured data format, specifically in JSON format. In this embodiment, the attributes of the data include:

[0161] ax1: transaction ID;

[0162] ax2: user ID;

[0163] ax3: trading time;

[0164] ax4: transaction amount;

[0165] ax5: account type;

[0166] ax6: account balance;

[0167] ax7: transaction type;

[0168] ax8: user age;

[0169] ax9: user credit score;

[0170] ax10: Number of historical defaults.

[0171] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0172] The collected data is manually labeled. In this embodiment, the labeled categories include: normal transactions, abnormal transactions, and high-risk behaviors, a total of 3 categories.

[0173] S2. Data expansion

[0174] It is understandable that the collection, acquisition, annotation and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention adopts a generative adversarial network algorithm based on correlation metric sub-state analysis to generate samples, thereby achieving data expansion; the generative adversarial network based on correlation metric sub-state analysis includes two components: a generator (G) and a discriminator (D); the generator is responsible for creating new data that is almost indistinguishable from real data, and the discriminator's task is to distinguish whether the input data comes from the original data set or is generated by the generator. The present invention adopts a correlation metric sub-state analysis strategy, which enhances the generator's ability to capture data features by simulating the interaction between the essential correlation of data and quantum states.

[0175] Specifically, the training process of the generative adversarial network algorithm based on association metric sub-state parsing is as follows:

[0176] S201, initialize the network parameters of the generator and the discriminator, specifically, the network parameters θ of the generator G and the discriminator D g and θ d Initialize randomly from the prior distribution, expressed as:

[0177]

[0178] In the formula, represents a normal distribution with a mean of 0 and a covariance of the unit matrix; θ g and θ d are the parameter vectors of the generator and the discriminator respectively;

[0179] S202, before starting adversarial training, pre-train the discriminator separately, and the discriminator D uses the real data set {x (i)} is trained, the goal is to maximize the probability of correctly classifying real data and generated data, and the loss function adopts the cross entropy loss function, which is expressed as:

[0180]

[0181] Where, L d is the loss function of the discriminator, n is the number of samples of real data, x (i) is the i-th real data sample, D(x (i) θ d ) is the output of the discriminator for sample i;

[0182] S203. In the adversarial training phase, the generator attempts to generate new data samples and adjusts parameters through the feedback of the discriminator to optimize the generated data samples so that the discriminator cannot distinguish between real and generated data. The loss function of the generator is defined as:

[0183]

[0184] Where, L g is the loss function of the generator, m is the number of generated samples, z (j) is the jth random noise vector, G(z (j) θ g ) is the parameter θ g The generator at the time is based on the noise z (j) The generated samples, G(z (j) θ g ) is calculated as:

[0185] G(z (j) θ g )=tanh(W g ·z (j) +b g )

[0186] Where W g is the weight matrix of the generator, b g is the bias term, z (j)is the input random noise vector, tanh() is the Tanh activation function;

[0187] The discriminator is trained based on the samples provided by the generator and the real samples to optimize the discriminator's recognition ability and improve the accuracy of distinguishing true from false. The loss function of the discriminator is defined as:

[0188]

[0189] In the formula, n and m represent the number of real samples and generated samples respectively, x (i) represents the i-th true sample.

[0190] During the optimization of the discriminator, the confidence weight β for each sample j The calculation method is expressed as:

[0191]

[0192] In the formula, ∥D(G(z (j) θ g ))-y real ∥ 2 Represents the generated sample and the real sample label y real The square of the Euclidean distance between them, ∥∥ represents the Euclidean distance, that is, the L2 norm.

[0193] S204. In each optimization update of the generator and the discriminator, the update of the generator and the discriminator will take into account the correlation metric quantum state of the data, and further enhance the expression of data features by simulating the interaction of quantum states; the correlation metric quantum state analytical function is used to simulate the interaction of quantum states, which is expressed as:

[0194]

[0195] Where Q(G,D) is the correlation metric substate analysis function, λ k is the weight of the kth quantum state, z k is the noise represented by the corresponding quantum state; Sig(D,G(z k θ g )) is the Sigmoid activation function;

[0196] Sig(D,G(z k θ g )) is calculated as:

[0197]

[0198] S205. In the process of optimizing the discriminator, the confidence estimation loss function C(D) is used to evaluate the confidence of the discriminator for each sample, so as to adjust the learning strategy of the discriminator, which is expressed as:

[0199]

[0200] Where C(D) is the confidence estimation loss function, γ bg is the adjustment factor, is the confidence weight of each sample, indicating the confidence of the discriminator on the jth generated sample; in this embodiment, if the confidence is less than a preset threshold, the discriminator considers it to be false data, and the preset threshold is set to 0.75.

[0201] S206. Repeat iterative steps S201-S205 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0202] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples; in this embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0203] S3. Feature extraction model training

[0204] The expanded data is input into the feature extraction model for training, and a 3-layer fully connected neural network is used for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion, or falling into the local optimal solution may occur, affecting the stability of training and the performance of the model. Specifically, a neural network model based on the dynamic niche neural optimization algorithm is used. Inspired by the niche theory in ecology, each species seeks its optimal position in the ecosystem to achieve an advantage in the competition for survival. Similarly, the dynamic niche neural optimization algorithm aims to find the optimal configuration of neural network parameters so that the network performs best on a given task.

[0205] Specifically, the training process of the neural network algorithm based on the dynamic niche neural optimization algorithm is as follows:

[0206] S301. In the initialization phase, a diverse primary population is generated. Each individual represents a set of neural network parameters. Let P (0) is the initial population, each individual p i is a parameter vector containing the weights and biases of the neural network, expressed as:

[0207]

[0208] Where Nu is the population size, W i is the weight matrix, bi is the bias vector, represents the direct concatenation of vectors; in this embodiment, P (0) The initialization method is random initialization.

[0209] S302, in the dynamic evaluation stage, the performance of each individual is evaluated according to the attractor properties of the dynamic trajectory. The dynamic evaluation method is implemented by using the fitness function f(), which is expressed as:

[0210] f(p i )=α Ts ·Acc(p i )+β Ts Stab(p i )

[0211] In the formula, Acc(p i ) represents the accuracy of the network on the labeled training data, which is calculated by the preset Softmax output; Stab(p i ) represents the dynamic stability of the parameters; α Ts and β Ts are performance hyperparameters and stability hyperparameters respectively; preferably, α Ts and β Ts Set to 0.7 and 0.3 respectively.

[0212] Dynamic stability of parameters Stab(p i ) is calculated as:

[0213]

[0214] In the formula, δ sta is the attenuation coefficient; is the loss function of the network, w ij is an individual p i is the jth weight parameter in , Mu is the total number of weights; preferably, δ sta Set to 0.05.

[0215] S303, calculate the loss function of the network for the next iteration, define κ(t) as the regularization strength of the t-th iteration, and κ(t+1) as the regularization strength of the t+1-th iteration. The calculation method is expressed as:

[0216] κ(t+1)=κ(t)·exp-α pe g(t)+β pe (1-L(t))

[0217] g(t)=(f(p (t) )-f(p (t-1) ))·exp-γ pe·Ds(p (t) )

[0218]

[0219] In the formula, α pe and β pe They are the fitness gradient hyperparameter and the learning progress hyperparameter respectively; g(t) is the fitness gradient, which measures the rate of change of the fitness function over time; L(t) is the learning progress, which represents the improvement of the performance of the network on the training data set at each iteration; γ pe is the fractal dimension adjustment factor; Ds() is the fractal dimension calculation function; preferably, α pe and β pe Set to 0.3 and 0.7 respectively.

[0220] In this embodiment, Ds(p) is defined as the local fractal dimension at the parameter vector p, and a small radius ∈ is defined around the parameter vector p. br The local area of ​​the sphere, calculate the number of individuals N(∈ br ), then the calculation method of fractal dimension is expressed as:

[0221]

[0222] Preferably, ∈ br Set to 0.0001.

[0223] The loss function of the network The calculation method is expressed as:

[0224]

[0225] Where R(W) is the L2 norm of the weight vector W, which is used to calculate the sum of the squares of the weights; is the cross entropy loss.

[0226] S304. Individuals move in the ecological niche according to their fitness. Excellent individuals occupy the dominant ecological niche, and weaker individuals explore new ecological niches or are eliminated. The adjustment method is expressed as:

[0227]

[0228]

[0229] In the formula, f avg is the average fitness of the current population, η vd is the interaction level adjustment factor, p best is the parameter of the current optimal individual, is the loss function About the i-th layer input The partial derivative of γ vd and λ vd are the update step size hyperparameter and the learning strategy hyperparameter respectively; preferably, γ vd Set to 0.1, λ vd Set to 0.05, η vd Set to 0.1.

[0230] The i-th layer input It is obtained through adaptive normalization, and the calculation method is expressed as:

[0231]

[0232] In the formula, x i is the i-th layer input before normalization, μ B and are the mean and variance of the batch input data respectively; γ zs and β zs is a learnable parameter, specifically the first learnable parameter and the second learnable parameter; ∈ zs is a very small constant; preferably, ∈ zs Set to 0.0001.

[0233] Adjust γ based on the best performing individual in the current population zs and β zs , to better adapt to the characteristics of the training data, the adjustment method is expressed as:

[0234] γ zs (t+1) =γ zs (t) ·exp(-η γ ·Var(p (t) ))

[0235]

[0236] In the formula, Var(p (t) )and are the variance and mean of the parameter vector of the current batch, γ zs (t) and β zs (t) are the first learnable parameter and the second learnable parameter of the current batch, γ zs (t+1) and β zs (t+1) are the first learnable parameter and the second learnable parameter for the next batch iteration, μ target is the target mean, η γ and η β is the learning rate. Preferably, η γ and ηβ Set to 0.1 and 0.05 respectively.

[0237] S305, perform crossover and update operations, and the new individual inherits the characteristics of the two excellent individuals, expressed as:

[0238] p new =ζ kg ·p parent1 +(1-ζ kg )·p parent2

[0239] In the formula, ζ kg is a randomly generated crossover factor between 0 and 1; p parent1 and p parent2 are two randomly selected parent individuals;

[0240] S306, performing mutation and mutation operations, using a mutation mechanism based on nonlinear dynamics theory, so that individual parameters can perform random jumps in a larger range, expressed as:

[0241] p′ i =p i +∈ bg ·randn() ·mask()

[0242] mask()=rand(size(p i ))<μ bg

[0243] In the formula, ∈ bg is the mutation intensity, μ bg is the probability of mutation, randn() generates a normally distributed random number; mask() generates a random number with p i A random binary mask of the same size, controlling which parameters will mutate; size(p i ) is a length measurement function; preferably, ∈ bg Set to 0.01, μ bg Set to 0.1.

[0244] S307, perform global dynamic update, the entire population is updated according to the theory of ecological dynamic system, the strategies of excellent individuals will be imitated and spread, and the unsuitable strategies will gradually disappear, expressed as:

[0245] P (t+1) = sort(P (t) ,f)

[0246] In the formula, the sort function sorts the population according to the fitness f(), retains individuals with high fitness, and eliminates individuals with low fitness;

[0247] S308. Repeat iterative steps S301-S307 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0248] S4. Feature dimensionality reduction model training

[0249] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model, and the feature dimensionality reduction is performed using an autoencoder neural network algorithm based on manifold learning. The autoencoder neural network based on manifold learning includes two parts: an encoder and a decoder. The function of the encoder is to map the high-dimensional input features to a low-dimensional latent space, and the decoder attempts to reconstruct the original input from this low-dimensional latent space. The present invention adopts the manifold learning principle to enable the network to better understand and retain the intrinsic structure of the data, and at the same time adopts a mask transformation mechanism to enhance the robustness of the model to noise and outliers in the input data.

[0250] The training process of the autoencoder neural network algorithm based on manifold learning is as follows:

[0251] S401, initialize the weights of the encoder and decoder, and the initialization method is random initialization, which is expressed as:

[0252]

[0253] Where W e and W d are the weight matrices of the encoder and decoder respectively; b e and b d is the corresponding bias vector; σ 2 represents the initialization standard deviation; preferably, σ 2 Set to 0.01.

[0254] S402: Sampling the manifold structure from the training data to estimate the local manifold structure of the data is performed by the k-nearest neighbor algorithm, that is, using the k-nearest neighbor method to sample each data point x i Sampling is performed and the average distance of its k nearest neighbors is calculated, which is expressed as:

[0255]

[0256] In the formula, x i Represents the current data point; NN k (x i ) is x i The k nearest neighbors of ; ∥x i -x j ∥ 2 is x i and x jThe square of the Euclidean distance between them;

[0257] S403: Generate a random mask for each input sample, where the mask is used to simulate the partial loss of input data. Specifically, for each input sample x i , generate a random mask m i The way is expressed as:

[0258] m i = Bernoulli(p)

[0259] In the formula, m i is applied to x i The mask of , each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, p is the probability that the mask element is 1;

[0260] Bernoulli distribution is a discrete distribution used to simulate a random variable with one of two possible outcomes. The Bernoulli random variable m ij The generation method is expressed as:

[0261]

[0262] Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij =1; if U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0. Preferably, pr is set to 0.45.

[0263] S404: After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i , calculated as:

[0264] z i =Sig(W e ·(1+γ csd β csd )·(x i ⊙m i )+b e )

[0265] Where ⊙ represents the corresponding multiplication of elements; Sig() is the Sigmoid activation function; β csd is the dynamic coding adjustment coefficient; γ csd is a preset scaling factor; preferably, γ csd Set to 0.03.

[0266] The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data. Specifically, the skewness S(x i ) and kurtosis K(x i ), expressed as:

[0267]

[0268]

[0269] In the formula, μ pd and σ pd x i The mean and standard deviation of

[0270] Furthermore, the adjustment coefficient β csd The calculation method is expressed as:

[0271]

[0272] Where tanh() is the hyperbolic tangent function;

[0273] S405. In the latent space, the manifold constraint is used to adjust the low-dimensional representation to ensure that the geometric and topological characteristics of the data are preserved. The calculation method of the manifold constraint loss is expressed as:

[0274]

[0275] In the formula, z i and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss, fidelity() is the quantum state fidelity loss function, λ p is the regularization parameter; preferably, λ p Set to 0.4;

[0276] The calculation method of the quantum state fidelity loss function fidelity() is expressed as:

[0277] fidelity p ,q p )=1-| <z p |q p >| 2

[0278] In the formula, | <z p |q p >| 2 Represents the quantum state q p With the original encoding state z p The square of the modulus of the inner product between;

[0279] The original encoding state z p The encoder takes the input feature vector x i Converted to a low-dimensional quantum state representation z through an encoder p , expressed as:

[0280]

[0281] In the formula, x p is the input feature vector, z p is the encoded low-dimensional representation, is the encoder function; are the parameters of the encoder, including W e and b e .

[0282] S406, reconstruct the input data from the low-dimensional representation through the decoder, calculate the reconstruction error and the manifold constraint error, and update the network parameters through back propagation. The total loss function Lur is composed of the reconstruction error and the manifold constraint error, which is expressed as:

[0283]

[0284] In the formula, is the reconstructed data point; kc is the regularization parameter of the manifold constraint; preferably, λ kc Set to 0.2.

[0285] S407, using the gradient descent method to optimize the network parameters and continuously reduce the total loss, the method of updating the parameters is expressed as:

[0286]

[0287] In the formula, α gk is the learning rate; preferably, α gk Set to 0.04.

[0288] S408. Repeat iterative steps S401-S407 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0289] S5. Classifier model training

[0290] The reduced-dimensional data is input into a classifier model for classifier model training, and a random forest based on mixed permutation entropy is used as a classification algorithm, wherein the random forest based on mixed permutation entropy includes a plurality of decision trees, and each tree uses a randomly selected data subset and feature subset during training. The mixed permutation entropy of the present invention is used to evaluate and select the optimal split point, thereby enhancing the robustness and accuracy of the model when processing complex financial behavior data.

[0291] Specifically, the training process of the random forest algorithm based on mixed permutation entropy is as follows:

[0292] S501. For each tree, randomly select a data subset and a feature subset before training to ensure the diversity and robustness of the model. For each tree t, the selected data subset D t and feature subset F t It is obtained by random sampling and is expressed as:

[0293] D t =sample(D,size=s D ,replace=false)

[0294] F t =sample(F,size=s F ,replace=false)

[0295] In the formula, D represents the complete training data set, F represents the complete feature set, and s D and F Respectively represent the number of extracted data and features, and the sample() function performs a random sampling operation;

[0296] S502: For each decision node, the hybrid permutation entropy H is used to select the best segmentation feature f and segmentation point θ, and then the data is distributed to the left and right child nodes according to the selected segmentation point, which is expressed as:

[0297]

[0298] and,

[0299] D left ={x∈D t |x f ≤θ}

[0300] D right ={x∈D t |x f >θ}

[0301]

[0302] In the formula, x frepresents the value of sample x on feature f, p c is the probability of class c in data set D, φ(x i ,x j ) represents the data point x i and x j The relative ranking contribution between ;

[0303] φ(x i ,x j ) is calculated as:

[0304]

[0305] In the formula, ∥x i -x j ∥ 2 Represents x i and x j The square of the Euclidean distance between tr is a hyperparameter that controls the range of similarity influence, λ tr is a weight coefficient that balances the traditional entropy and permutation contributions; preferably, λ tr Set to 0.2, σ tr Set to 0.1.

[0306] S503, after the tree is fully grown, an independent validation data set is used to evaluate the performance of the tree, and then the structure of the tree is dynamically adjusted according to the preset performance index, and branches that do not provide additional information gain are removed; specifically, for the t-th tree, its performance Perf(t) on the independent validation set V is evaluated, and pruning is performed; the performance evaluation function Perf(t) is implemented by ensemble learning, and is expressed as:

[0307]

[0308] In the formula, ω(x, t) represents the weight of sample x in tree t, which is dynamically adjusted based on the contribution of the sample; I(t(x)=y x ) is an indicator function, which takes the value of 1 when the tree t predicts the sample x correctly, otherwise it takes the value of 0;

[0309] ω(x,t) is adjusted according to the error rate of the sample, expressed as:

[0310] ω(x,t)=exp-α t I(t(x)≠y x )

[0311] In the formula, α t is the performance parameter of tree t in the current training iteration, which is proportional to the error rate of the tree.

[0312] S504, repeat steps S501-S503, construct multiple decision trees, integrate the multiple trees, and use the majority voting method to make the final classification decision, which is expressed as:

[0313] y pred =mode({t(x)|t∈T})

[0314] Where t(x) is the prediction of tree t for sample x, T is the set of all trees, and mode() is the decision function that takes the majority;

[0315] S6. Analysis of user financial management behavior

[0316] The trained model is used to process new sample data, so as to realize the analysis and prediction of user financial management behavior. Specifically, the new input sample first obtains the extracted features through the feature extraction model, and then the extracted high-dimensional feature vector is input into the data dimension reduction model to reduce the feature dimension; finally, the dimension-reduced feature vector is input into the classifier model to predict the type of financial management behavior. In this embodiment, the predicted categories include: normal transactions, abnormal transactions, and high-risk behaviors, a total of 3 categories.

Claims

1. A user financial management behavior analysis method based on artificial intelligence and financial big data, characterized in that: The following steps are involved: S1. Data collection and annotation Data is collected from financial databases, including financial transaction records, user profiles, credit reports, and historical financial management behavior data. The collected data is stored in a structured data format, specifically JSON format; the collected data is manually annotated; S2. Data expansion The sample generation is performed by using a generative adversarial network algorithm based on the analysis of associated metric substates, thereby realizing data expansion. The generative adversarial network based on the analysis of associated metric substates includes two components: a generator and a discriminator. Specifically, the training process of the generative adversarial network algorithm based on the analysis of associated metric substates is as follows: S201, initialize the network parameters of the generator and the discriminator, specifically, the network parameters θ of the generator G and the discriminator D g and θ d Initialize randomly from the prior distribution, expressed as: In the formula, represents a normal distribution with a mean of 0 and a covariance of the unit matrix; θ g and θ d are the parameter vectors of the generator and the discriminator respectively; S202, before starting adversarial training, pre-train the discriminator separately, and the discriminator D uses the real data set {x (i) } is trained, the goal is to maximize the probability of correctly classifying real data and generated data, and the loss function adopts the cross entropy loss function, which is expressed as: Where, L d is the loss function of the discriminator, n is the number of samples of real data, x (i) is the i-th real data sample, D(x (i) ; θ d ) is the output of the discriminator for sample i; S203. In the adversarial training phase, the generator attempts to generate new data samples and adjusts parameters through the feedback of the discriminator to optimize the generated data samples so that the discriminator cannot distinguish between real and generated data. The loss function of the generator is defined as: Where, L g is the loss function of the generator, m is the number of generated samples, z (j) is the jth random noise vector, G(z (j) θ g ) is the parameter θ g The generator at the time is based on the noise z (j) The generated samples, G(z (j) θ g ) is calculated as: G(z (j) ;θ g )=tanh(W g ·z (j) +b g ) Where W g is the weight matrix of the generator, b g is the bias term, z (j) is the input random noise vector, tanh() is the Tanh activation function; S204. In each optimization update of the generator and the discriminator, the update of the generator and the discriminator will take into account the correlation metric quantum state of the data, and further enhance the expression of data features by simulating the interaction of quantum states; the correlation metric quantum state analytical function is used to simulate the interaction of quantum states, which is expressed as: Where Q(G,D) is the correlation metric substate analysis function, λ k is the weight of the kth quantum state, z k is the noise represented by the corresponding quantum state; Sig(D,G(z k θ g )) is the Sigmoid activation function; Sig(D,G(z k θ g )) is calculated as: S205. In the process of optimizing the discriminator, the confidence estimation loss function C(D) is used to evaluate the confidence of the discriminator for each sample, so as to adjust the learning strategy of the discriminator, which is expressed as: Where C(D) is the confidence estimation loss function, γ bg is the adjustment factor, is the confidence weight for each sample, indicating the confidence of the discriminator in the jth generated sample; S206, repeating iterative steps S201-S205 until a preset stop iteration condition is met, indicating that the model training is completed; After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples; S3. Feature extraction model training The expanded data is input into the feature extraction model for training. A 3-layer fully connected neural network is used for feature extraction. Specifically, the neural network algorithm based on the dynamic niche neural optimization algorithm has the following training process: S301. In the initialization phase, a diverse population is generated. Each individual represents a set of neural network parameters. Let P (0) is the initial population, each individual p i is a parameter vector containing the weights and biases of the neural network, expressed as: Where Nu is the population size, W i is the weight matrix, b i is the bias vector, Represents direct concatenation of vectors; S302, in the dynamic evaluation stage, the performance of each individual is evaluated according to the attractor properties of the dynamic trajectory. The dynamic evaluation method is implemented by using the fitness function f(), which is expressed as: f(p i )=α Ts ·Acc(p i )+β Ts ·Stab(p i ) In the formula, Acc(p i ) represents the accuracy of the network on the labeled training data, which is calculated by the preset Softmax output; Stab(p i ) represents the dynamic stability of the parameters; α Ts and β Ts They are performance hyperparameters and stability hyperparameters respectively; Dynamic stability of parameters Stab(p i ) is calculated as: In the formula, δ sta is the attenuation coefficient; is the loss function of the network, w ij is an individual p i is the jth weight parameter in , Mu is the total number of weights; S303, calculate the loss function of the network for the next iteration, define κ(t) as the regularization strength of the t-th iteration, and κ(t+1) as the regularization strength of the t+1-th iteration. The calculation method is expressed as: κ(t+1)=κ(t)·exp(-α pe ·g(t)+β pe ·(1-L(t))) g(t)=(f(p (t) )-f(p (t-1) ))·exp(-γ pe ·Ds(p (t) )) In the formula, α pe and β pe They are the fitness gradient hyperparameter and the learning progress hyperparameter respectively; g(t) is the fitness gradient, which measures the rate of change of the fitness function over time; L(t) is the learning progress, which represents the performance improvement of the network on the training data set at each iteration; γ pe is the fractal dimension adjustment factor; Ds() is the fractal dimension calculation function; S304. Individuals move in ecological niches according to their fitness. Excellent individuals occupy the dominant ecological niches, while weaker individuals explore new ecological niches or are eliminated. The adjustment method is expressed as: In the formula, f avg is the average fitness of the current population, η vd is the interaction level adjustment factor, p best is the parameter of the current optimal individual, is the loss function About the i-th layer input The partial derivative of γ vd and λ vd They are the update step size hyperparameter and the learning strategy hyperparameter; S305, perform crossover and update operations, and the new individual inherits the characteristics of the two excellent individuals, expressed as: p new =ζ kg ·p parent1 +(1-z kg )·p parent2 In the formula, ζ kg is a randomly generated crossover factor between 0 and 1; p parent1 and p parent2 are two randomly selected parent individuals; S306, performing mutation and mutation operations, using a mutation mechanism based on nonlinear dynamics theory, so that individual parameters can perform random jumps in a larger range, expressed as: p′ i =p i +∈ bg ·randn()·mask() mask()=rand(size(p i ))<μ bg In the formula, ∈ bg is the mutation intensity, μ bg is the probability of mutation, randn() generates a normally distributed random number; mask() generates a random number with p i A random binary mask of the same size, controlling which parameters will mutate; size(p i ) is the length measurement function; S307, perform global dynamic update, the entire population is updated according to the theory of ecological dynamic system, the strategies of excellent individuals will be imitated and spread, and the unsuitable strategies will gradually disappear, expressed as: P (t+1) =sort(P (t) ,f) In the formula, the sort function sorts the population according to the fitness f(), retains individuals with high fitness, and eliminates individuals with low fitness; S308, repeating the iteration steps S301-S307 until the preset stop iteration condition is met, indicating that the model training is completed; S4. Feature dimensionality reduction model training The data after feature extraction is input into the feature dimension reduction model for training. The autoencoder neural network algorithm based on manifold learning is used for feature dimension reduction. The training process is as follows: S401, initialize the weights of the encoder and decoder, and the initialization method is random initialization, which is expressed as: Where W e and W d are the weight matrices of the encoder and decoder respectively; b e and b d is the corresponding bias vector; σ 2 represents the standard deviation of initialization; S402: Sampling the manifold structure from the training data to estimate the local manifold structure of the data is performed by the k-nearest neighbor algorithm, that is, using the k-nearest neighbor method to sample each data point x i Sampling is performed and the average distance of its k nearest neighbors is calculated, which is expressed as: In the formula, x i Represents the current data point; NN k (x i ) is x i The k nearest neighbors of ; ∥x i -x j ∥ 2 is x i and x j The square of the Euclidean distance between them; S403: Generate a random mask for each input sample, where the mask is used to simulate the partial loss of input data. Specifically, for each input sample x i , generate a random mask m i The way is expressed as: m i =Bernoulli(p) In the formula, m i is applied to x i The mask of , each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, p is the probability that the mask element is 1; S404: After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i , calculated as: z i =Sig(W e ·(1+γ csd β csd )·(x i ⊙m i )+b e ) Where ⊙ represents the corresponding multiplication of elements; Sig() is the Sigmoid activation function; β csd is the dynamic coding adjustment coefficient; γ csd is the preset scaling factor; The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data. Specifically, the skewness S(x i ) and kurtosis K(x i ), expressed as: In the formula, μ pd and σ pd x i The mean and standard deviation of Furthermore, the adjustment coefficient β csd The calculation method is expressed as: Where tanh() is the hyperbolic tangent function; S405. In the latent space, the manifold constraint is used to adjust the low-dimensional representation to ensure that the geometric and topological characteristics of the data are preserved. The calculation method of the manifold constraint loss is expressed as: In the formula, z i and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss, fidelity() is the quantum state fidelity loss function, λ p is the regularization parameter; S406, reconstruct the input data from the low-dimensional representation through the decoder, calculate the reconstruction error and the manifold constraint error, and update the network parameters through back propagation. The total loss function Lur is composed of the reconstruction error and the manifold constraint error, which is expressed as: In the formula, is the reconstructed data point; λ kc is the regularization parameter of the manifold constraint; S407, using the gradient descent method to optimize the network parameters and continuously reduce the total loss, the method of updating the parameters is expressed as: In the formula, α gk is the learning rate; S408, repeating the iteration steps S401-S407 until the preset stop iteration condition is met, indicating that the model training is completed; S5. Classifier model training The reduced-dimensional data is input into the classifier model for classifier model training. The random forest based on mixed permutation entropy is used as the classification algorithm. The training process is as follows: S501. For each tree, randomly select a data subset and a feature subset before training to ensure the diversity and robustness of the model. For each tree t, the selected data subset D t and feature subset F t It is obtained by random sampling and is expressed as: D t =sample(D,size=s D ,replace=false) F t =sample(F,size=s F ,replace=false) In the formula, D represents the complete training data set, F represents the complete feature set, and s D and F Respectively represent the number of extracted data and features, and the sample() function performs a random sampling operation; S502: For each decision node, the hybrid permutation entropy H is used to select the best segmentation feature f and segmentation point θ, and then the data is distributed to the left and right child nodes according to the selected segmentation point, which is expressed as: and, D left ={x∈D t |x f ≤θ} D right ={x∈D t |x f >θ} In the formula, x f represents the value of sample x on feature f, p c is the probability of class c in data set D, φ(x i ,x j ) represents the data point x i and x j The relative ranking contribution between ; φ(x i ,x j ) is calculated as: In the formula, ∥x i -x j ∥ 2 Represents x i and x j The square of the Euclidean distance between tr is a hyperparameter that controls the range of similarity influence, λ tr is the weight coefficient that balances the traditional entropy and permutation contributions; S503, after the tree is fully grown, an independent validation data set is used to evaluate the performance of the tree, and then the structure of the tree is dynamically adjusted according to the preset performance index, and branches that do not provide additional information gain are removed; specifically, for the t-th tree, its performance Perf(t) on the independent validation set V is evaluated, and pruning is performed; the performance evaluation function Perf(t) is implemented by ensemble learning, and is expressed as: In the formula, ω(x, t) represents the weight of sample x in tree t, which is dynamically adjusted based on the contribution of the sample; I(t(x)=y x ) is an indicator function, which takes the value of 1 when the tree t predicts the sample x correctly, otherwise it takes the value of 0; S504, repeat steps S501-S503, construct multiple decision trees, integrate the multiple trees, and use the majority voting method to make the final classification decision, which is expressed as: y pred =mode({t(x)|t∈T}) Where t(x) is the prediction of tree t for sample x, T is the set of all trees, and mode() is the decision function that takes the majority; S6. Analysis of user financial management behavior The trained model is used to process new sample data to analyze and predict the user's financial management behavior. Specifically, the new input sample is first extracted by the feature extraction model to obtain the features, and then the extracted high-dimensional feature vector is input into the data dimensionality reduction model to reduce the feature dimension; finally, the feature vector after dimensionality reduction is input into the classifier model to predict the type of financial management behavior.

2. The user financial management behavior analysis method based on artificial intelligence and financial big data according to claim 1 is characterized in that: In step S203, the discriminator is trained based on the samples provided by the generator and the real samples to optimize the recognition ability of the discriminator and improve the accuracy of distinguishing the true from the false. The loss function of the discriminator is defined as: In the formula, n and m represent the number of real samples and generated samples respectively, x (i) represents the i-th true sample.

3. The user financial management behavior analysis method based on artificial intelligence and financial big data according to claim 1 is characterized in that: In step S203, during the optimization process of the discriminator, the confidence weight β of each sample is j The calculation method is expressed as: In the formula, ∥D(G(z (j) θ g ))-y real ∥ 2 Represents the generated sample and the real sample label y real The square of the Euclidean distance between them, ∥∥ represents the Euclidean distance, that is, the L2 norm.

4. The user financial management behavior analysis method based on artificial intelligence and financial big data according to claim 1 is characterized in that: In step S303, the network loss function The calculation method is expressed as: Where R(W) is the L2 norm of the weight vector W, which is used to calculate the sum of the squares of the weights; is the cross entropy loss.

5. The user financial management behavior analysis method based on artificial intelligence and financial big data according to claim 1 is characterized in that: In step S304, the i-th layer input It is obtained through adaptive normalization, and the calculation method is expressed as: In the formula, x i is the i-th layer input before normalization, μ B and are the mean and variance of the batch input data respectively; γ zs and β zs is a learnable parameter, specifically the first learnable parameter and the second learnable parameter; ∈ zs is a very small constant; Adjust γ based on the best performing individual in the current population zs and β zs , to better adapt to the characteristics of the training data, the adjustment method is expressed as: c zs (t+1) =c zs (t) ·exp(-η γ ·Var(p (t) )) In the formula, Var(p (t) )and are the variance and mean of the parameter vector of the current batch, γ zs (t) and β zs (t) are the first learnable parameter and the second learnable parameter of the current batch, γ zs (t+1) and β zs (t+1) are the first learnable parameter and the second learnable parameter for the next batch iteration, μ target is the target mean, η γ and η β is the learning rate.

6. The method for analyzing user financial management behavior based on artificial intelligence and financial big data according to claim 1, characterized in that: In step S403, the Bernoulli distribution is a discrete distribution used to simulate a random variable with one of two possible results. The Bernoulli random variable m ij The generation method is expressed as: Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij =1 probability; If U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0.

7. The user financial management behavior analysis method based on artificial intelligence and financial big data according to claim 1 is characterized in that: In step S405, the calculation method of the quantum state fidelity loss function fidelity() is expressed as: fidelity(z p ,q p )=1-|<z p |q p >| 2 In the formula, | <z p |q p >| 2 Represents the quantum state q p With the original encoding state z p The square of the modulus of the inner product between; The original encoding state z p The encoder takes the input feature vector x i Converted to a low-dimensional quantum state representation z through an encoder p , expressed as: In the formula, x p is the input feature vector, z p is the encoded low-dimensional representation, is the encoder function; are the parameters of the encoder, including W e and b e .

8. The method for analyzing user financial management behavior based on artificial intelligence and financial big data according to claim 1, characterized in that: In step S503, ω(x, t) is adjusted according to the error rate of the sample, which is expressed as: ω(x,t)=exp(-α t ·I(t(x)≠y x )) In the formula, α t is the performance parameter of tree t in the current training iteration, which is proportional to the error rate of the tree.

Citation Information

Patent Citations

  • A method and a device for processing an automatic test result of a financial data analysis system

    CN109634868A

  • Method and system for user transaction feature analysis and early warning

    CN117726341A

  • Financial content recommendation method, processor and storage medium

    CN118071426A

  • Strabismus postoperative visual training scheme recommendation method based on artificial intelligence

    CN118629584A

Cited By

  • Intelligent management and control method and system for building water

    CN119228590A

  • Baseline evaluation method for air conditioner load under various meteorological conditions

    CN119903458A

  • Machine room equipment operation and maintenance data classification method and system based on artificial intelligence

    CN120257132A