Multi-source data fusion method and system based on cloud computing

Abstract

The invention relates to the technical field of cloud computing and multi-source data fusion, and discloses a multi-source data fusion method and system based on cloud computing, and the method comprises the steps: obtaining source scene data and target scene data, carrying out the modeling of the uncertainty distribution of a source scene through a Bayesian neural network, and obtaining the target scene data; network parameters with probability distribution are obtained; calculating the domain difference between the source scene and the target scene, and determining the distribution offset degree; uncertainty perception knowledge migration is executed, parameter probability distribution of a source scene model is migrated to a target scene, and the migration intensity is adaptively determined by domain differences; using the migration result to generate a fusion result with a confidence interval, and providing decision reliability evaluation; on the basis of actual feedback of the target scene, uncertainty model parameters are updated, and a migration strategy is optimized; according to the multi-source data fusion method for Bayesian uncertainty migration, knowledge migration can be carried out while uncertainty is reserved and adjusted.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and multi-source data fusion technology, and more specifically, to a cloud computing-based multi-source data fusion method and system. Background Technology

[0002] With the rapid development of cloud computing and big data technologies, enterprises have accumulated a large amount of heterogeneous and multi-source data resources in their actual business operations. In order to improve data utilization efficiency and intelligent decision-making capabilities, enterprises usually need to apply existing data fusion models to constantly changing and expanding new business scenarios.

[0003] However, the data distribution, feature structure, and business semantics often differ significantly across different business scenarios, posing challenges to the adaptability and generalization ability of models during the transfer process. Existing data fusion and transfer learning methods often focus on the direct transfer of model parameters or domain knowledge, paying less attention to the modeling and quantification of uncertainties in new scenarios. This neglect of uncertainty can easily lead to overconfidence in models when facing new scenarios with distributional shifts or insufficient samples, making it impossible to effectively assess the reliability of prediction results and thus increasing the risk of actual decision-making.

[0004] Therefore, how to fully model and utilize uncertain information during multi-source data fusion and transfer learning, and improve the adaptability and reliability of models in new business scenarios, has become an important problem that urgently needs to be solved in the current technical field. Summary of the Invention

[0005] This invention provides a cloud computing-based multi-source data fusion method and system, which solves the technical problem of overconfidence in model migration and difficulty in quantifying reliability under new scenario distribution shifts in related technologies.

[0006] This invention provides a cloud computing-based multi-source data fusion method, including: Acquire source scene data and target scene data, and model the uncertainty distribution of the source scene using a Bayesian neural network to obtain network parameters with probability distribution; Based on network parameters, the domain difference between the source scene and the target scene is calculated to determine the degree of distribution offset; Based on the obtained distribution offset, uncertainty-aware knowledge transfer is performed to transfer the parameter probability distribution of the source scene model to the target scene. The transfer strength is adaptively determined by the domain differences. Using the migration results, fusion results with confidence intervals are generated to provide an assessment of decision reliability; Based on the actual feedback from the target scenario, update the parameters of the uncertainty model and optimize the migration strategy.

[0007] Furthermore, the Bayesian neural network adopts a hierarchical structure design, including an input layer, multiple hidden layers and an output layer, and the weight parameters of each layer are modeled as probability distributions rather than fixed values.

[0008] Furthermore, the calculation of the domain difference is implemented using the maximum average difference principle and kernel method. By quantifying the data distribution of the source scene and the target scene, the degree of distribution offset between the data distribution of the source scene and the target scene is evaluated, which serves as the basis for subsequent knowledge transfer.

[0009] Furthermore, when performing uncertainty-aware knowledge transfer, the parameter adjustment adopts an adaptive transfer intensity parameter, which is dynamically adjusted according to the results of domain difference measurement through a preset decay rate parameter to achieve precise control of the knowledge transfer intensity.

[0010] Furthermore, the confidence interval of the fusion result is generated by a Bayesian prediction algorithm, which uses the Monte Carlo importance sampling method to extract multiple sets of parameters from the parameter distribution and calculate the mean and variance of the predicted values.

[0011] Furthermore, after acquiring the source scene data and the target scene data, the business data is analyzed, and the scene semantic embedding vector is extracted to obtain the conceptual level representation of the scene.

[0012] Furthermore, after performing the uncertainty-aware knowledge transfer, the regularization strength of the model parameter probability distribution is dynamically adjusted based on the quantification results of domain differences, and an adaptive prior regularization method is adopted to prevent the model from overfitting or underfitting.

[0013] Furthermore, the adaptive prior regularization, through a regularization strength parameter automatically adjusted based on domain differences and combined with a regularization function, adaptively constrains the probability distribution of model parameters, thereby optimizing the model's generalization ability.

[0014] Furthermore, updating the uncertainty model parameters based on the actual feedback from the target scenario includes calculating the coverage and width of the prediction confidence interval, and dynamically adjusting the uncertainty parameters of the Bayesian neural network model according to these indicators.

[0015] This invention provides a cloud computing-based multi-source data fusion system for executing the aforementioned cloud computing-based multi-source data fusion method, comprising: The data acquisition module is used to acquire source scene data and target scene data; The uncertainty modeling module is used to model the uncertainty distribution of the source scene using a Bayesian neural network; The domain difference calculation module is used to calculate the degree of distribution offset between the source scene and the target scene; The knowledge transfer module is used to perform uncertainty-aware knowledge transfer, transferring the parameter probability distribution of the source scene model to the target scene. The results generation module is used to generate fusion results with confidence intervals, providing an assessment of decision reliability. The parameter update module is used to update the parameters of the uncertainty model and optimize the migration strategy based on the actual feedback from the target scenario.

[0016] The beneficial effect of this invention is that by modeling the probability distribution of parameters rather than point estimates, it can capture the uncertainty information of model parameters. During transfer learning, this method retains and adjusts this uncertainty information, avoiding the overconfidence problem of traditional transfer methods in new scenarios. A domain difference quantization algorithm is introduced, which can accurately calculate the distribution offset between the source and target scenes. Based on this difference metric, the system can adaptively adjust the strength of the migration strategy, avoiding performance degradation caused by direct migration when scene differences are large. The generated fusion results include confidence intervals, directly quantifying the reliability of the predictions. Compared to traditional methods, this approach not only provides the prediction results but also an assessment of the prediction's uncertainty, enabling decision-makers to gain a more comprehensive understanding of the prediction's reliability and avoid making erroneous decisions under conditions of high uncertainty. Adaptive prior regularization can automatically adjust the regularization strength according to domain differences, helping the model prevent overfitting in target scenarios with scarce data. The model parameter update mechanism based on actual feedback enables the system to continuously optimize uncertainty estimation, making the predicted confidence interval more accurate. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cloud computing-based multi-source data fusion method in this invention; Figure 2 It is an area map comparing the distribution differences between the source scene and the target scene across four key feature dimensions; Figure 3 It is a bar chart comparing the raw offset and weighted offset of the four feature dimensions; Figure 4 It is a composite plot comparing the parameter distribution uncertainty before and after the migration; Figure 5 This is a graph showing the relationship between the confidence interval coverage improvement after model feedback updates. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a multi-source data fusion method based on cloud computing, such as... Figure 1 Shown, including: Step 1: Obtain source scene data and target scene data; model the uncertainty distribution of the source scene using a Bayesian neural network to obtain network parameters with probability distribution. In this step, the system first receives the dataset of the source scene. , which includes One sample. Each Indicates the first in the source scene Features of each sample Indicates the first The system also receives the dataset corresponding to each sample. ,in Indicates the first in the target scene Features of each sample For the first Each sample has a label. It should be noted that this refers to the number of samples in the dataset for the target scene. Typically much smaller than the number of dataset samples in the source scenario. .

[0020] To ensure data quality and model performance, the system preprocesses the received source and target scene data, including: standardizing numerical features to transform features of different dimensions into a distribution with a mean of 0 and a variance of 1, eliminating the impact of dimensional differences on the model; performing one-hot encoding or embedding encoding on categorical features to convert non-numerical data into a numerical representation that the model can process; imputing missing values ​​using the mean / median or model-based prediction; and handling outliers through truncation or special transformations to ensure data stability. The preprocessed feature vectors have the same dimensions and scale, ensuring that computational biases due to dimensional differences do not cause errors in subsequent distance calculations and parameter learning.

[0021] Bayesian neural networks, as a conventional model, take source scene datasets as input. and prior distribution The output is the posterior probability distribution of the network parameters. Variational inference algorithms, as conventional algorithms, take the prior distribution as their input. and likelihood function The output is the posterior distribution. The posterior distribution is an approximation. It contains not only the optimal parameter values ​​but also information about the parameter uncertainty.

[0022] Furthermore, the uncertainty-aware Bayesian neural network in this embodiment employs a hierarchical structure, including an input layer, multiple hidden layers, and an output layer. The weight parameters of each layer are modeled as probability distributions rather than fixed values. Multiple sets of parameter values ​​are sampled from these distributions using Monte Carlo sampling for forward computation, generating a distribution of prediction results, thereby capturing different types of uncertainty information. The network's output layer is designed as either a regression output layer (linear activation function) or a classification output layer (softmax activation function) depending on the task type. The loss function uses a combination of negative log-likelihood and KL divergence. ; in, This is the total loss function; Represents the source scene dataset; Represents the parameters of the neural network; These are variational distribution parameters; Indicates to about Expectations; This is a variational approximation of the posterior distribution; It is the prior distribution; It is the likelihood function; for and Kullback-Leibler divergence between them; This is the natural logarithm operator.

[0023] Furthermore, in this embodiment, the optimization objective function during network training also includes a time dimension design, expressed as: ; in, This is a time-constrained loss function; These are variational distribution parameters; For the training time step, the range of values ​​is... ; This is the maximum number of training iterations, typically 10,000. This is the total loss function; The regularization coefficient varies over time; This is a time-dependent regularization term.

[0024] Equal to the initial regularization coefficient Multiply by a decay factor that decreases with the number of training steps. Specifically, as the training steps... The increase, It will gradually decrease, and the decay rate is determined by the parameter. control. The value range is from 0.1 to 1.0, and the default value is 0.5; The value range is from 0.5 to 2.0, and the default value is 1.0; Indicates the current training steps. This is the maximum number of training steps. This design allows for strong regularization constraints on the model in the early stages of training, with the regularization strength gradually decreasing as training progresses, thereby balancing the model's generalization and fitting abilities.

[0025] Defined as: ; in For time-related regularization terms, These are the parameters of the neural network; These are the parameters of the neural network at the previous time step; The square of the L2 norm; It is an exponential function; This is the time decay coefficient, with a value range of [1.0, 5.0], and a default value of 3.0; This is the current training step; This represents the maximum number of training steps.

[0026] Furthermore, in this embodiment, the prior distribution of the weight parameters is set to a Gaussian distribution, where the prior standard deviation ranges from [0.1, 1.0], with a default value of 0.5. The variational approximation distribution adopts a diagonal Gaussian distribution, where the mean and variance are variational parameters obtained through network learning. To achieve a stable variational inference process, this embodiment uses a reparameterization technique to reduce the variance of gradient estimation by linearly combining the variational parameters with standard normal distribution sample values ​​to obtain the weight parameters. During training, the number of Monte Carlo samplings is set to... Within the range, the default is Next (such as) To balance computational cost and estimation accuracy, the weighting coefficients of the KL divergence term are used. The range of values ​​is The annealing mechanism dynamically adjusts the parameters during training. The initial value is small to encourage the model to explore the parameter space, and it is gradually increased in the later stages to apply appropriate regularization constraints.

[0027] In this embodiment of the application, in order to improve the adaptability of transfer learning, the following steps may also be performed after step 1: Step 1.1: Analyze the business data of the source and target scenarios, extract the scene semantic embedding vectors, and obtain the conceptual level representation of the scenario; In this step, the system processes the business data of the source and target scenarios and extracts the semantic information of the scenarios.

[0028] The semantic embedding model, used in this implementation, takes business data from the source and target scenarios as input and outputs semantic representation vectors of the scenarios. These semantic representation vectors capture high-level semantic information of the scenarios, such as the business background, business rules, and business relationships.

[0029] Furthermore, in this embodiment, the dimension of the semantic embedding vector is set at... Within the range, the default is Dimensions (e.g.) To balance expressive power and computational complexity, semantic similarity is calculated using cosine similarity, with values ​​ranging from [value range missing]. The similarity threshold is 1, where 1 represents perfect similarity, -1 represents perfect opposites, and 0 represents no correlation. In practical applications, we set the similarity threshold to 1. When the cosine similarity of the semantic embedding vectors of two scenes exceeds this threshold, they are considered to have a high semantic correlation and can be subject to stronger knowledge transfer; when the similarity is lower than this threshold, they are considered to have a lower correlation. At that time, it was considered that they were semantically significantly different, and the intensity of knowledge transfer should be reduced or a more conservative transfer strategy should be adopted. Similarity in For scene pairs within the interval, a linear interpolation method is used to determine the appropriate migration intensity.

[0030] Furthermore, the scene semantic embedding model in this embodiment employs a deep bidirectional attention mechanism, comprising two key components: a scene feature extractor and a semantic association encoder. The scene feature extractor uses a multilayer perceptron to extract features from the original business data, while the semantic association encoder captures the semantic relationships between different features through an attention module, generating the final semantic embedding vectors. These semantic embedding vectors are used to guide the adjustment of parameter distribution during transfer learning, making the transfer process more adaptable to the semantic characteristics of different scenarios. The output layer of the scene semantic embedding model is a fully connected layer, generating fixed-dimensional semantic embedding vectors. A contrastive loss function is used during training. ; in, The semantic embedding loss function; Indicates the summation symbol; To take the larger of zero and the value within the parentheses; These are boundary parameters; It is a cosine distance function; For the first Semantic embedding vectors for each scene; For Semantically similar scene embedding vectors; For Embedding vectors for semantically dissimilar scenes.

[0031] Furthermore, in this embodiment, the specific architecture and implementation process of the scene semantic embedding model are as follows: Business data preprocessing: The business data of the source scene and the target scene are converted into a structured feature matrix, including numerical features (through standardization) and categorical features (through embedding representation); Feature extractor structure: It consists of three fully connected layers, containing 128, 256 and 128 neurons respectively. Each layer uses the LeakyReLU activation function (negative slope of 0.01) and batch normalization. A Dropout layer (dropout rate of 0.3) is added after the second layer to prevent overfitting; Attention mechanism: A multi-head self-attention mechanism is adopted, with the number of heads set to The dimension of each attention head is The formula for calculating attention is: ; in, For attention output; For query matrix; The key matrix; It is a value matrix; for and The product of the transposes of matrices; The dimension of the key; This is the softmax normalization function.

[0032] Semantic association encoding: The self-attention output is combined with the original feature representation, processed through residual connections and layer normalization, and then passed through a fully connected layer containing 64 neurons to generate the final embedding vector; Model training: Boundary parameters in the loss function are compared. The learning rate is set to 0.2. Positive sample pairs are constructed from similar business scenarios, while negative sample pairs are constructed from randomly sampled unrelated scenarios. The Adam optimizer is used with a learning rate of 0.0005 and a batch size of 32. Training is conducted for 200 epochs or until convergence. After training, the model can map any business scenario data to a semantic space, which can be used to calculate the similarity between scenarios and guide knowledge transfer.

[0033] Step 2: Based on network parameters, calculate the domain difference between the source scene and the target scene to determine the degree of distribution offset; In this step, the system applies a domain difference quantification algorithm to calculate the difference between the data distribution of the source scene and the target scene.

[0034] The domain difference quantization algorithm, an innovative algorithm in this implementation, takes the source scene feature distribution and the target scene feature distribution as inputs and outputs the distribution offset degree. The domain difference quantization algorithm employs the maximum average difference principle, and its mathematical expression is as follows: ; in and These represent the feature distributions of the source scene and the target scene, respectively. As a measure of domain differences; The supremum; For the hypothesis function, It is a set of hypothesis functions; Represents the expectation operator; The symbol representing "obedience" , These represent the first and second input business data samples, respectively. This is for absolute value operations; and These represent the feature distributions of the source scene and the target scene, respectively.

[0035] Inner function It is a hypothesis function used to transform input data Mapping to the feature space to maximize the distributional difference between the source and target domains. It is a set of hypothesis functions, usually a neural network or a family of kernel functions. The output values ​​are normalized to ensure that their range is within a certain range. Within the range.

[0036] Furthermore, in this embodiment, the calculation of the domain difference metric is implemented using a kernel method, specifically defined as: ; in This indicates a domain difference measure implemented using a kernel method. The feature mapping function maps the original data to the reproducing kernel Hilbert space. ; Indicates the summation symbol; Let be the norm in the RKHS space; The number of samples in the source scene; The number of samples in the target scene; For the first in the source scene One sample; For the target scenario One sample.

[0037] In actual calculations, this can be simplified using kernel tricks as follows: ; in This represents a simplified measure of domain difference using kernel techniques. For kernel functions, the Gaussian kernel is commonly used; Indicates the summation symbol; The number of samples in the source scene; The number of samples in the target scene; , The first in the source scene The and the first One sample; , For the target scenario The and the first One sample; , They represent the first in the source scene respectively. The and the first One sample.

[0038] The Gaussian kernel function measures the similarity between two input sample vectors. Specifically, it calculates the similarity based on the Euclidean distance between the two vectors; the closer the distance, the higher the similarity. This function includes a kernel width parameter. Its value ranges from 0.1 to 10, with a default value of 1. Used to control the kernel function's sensitivity to distance. The larger the value, the less sensitive the kernel function is to changes in distance. The output value of the Gaussian kernel function is transformed by an exponential function to finally obtain a numerical value that reflects the similarity of samples.

[0039] Domain Difference Measurement The value range is [0, 2]. It is normalized to the interval [0, 1] by linear transformation, that is, the original difference measure value is divided by 2 for standardization.

[0040] Furthermore, the domain difference quantification algorithm in this embodiment is implemented through a domain adversarial network, comprising two key components: a feature extractor and a domain discriminator. The feature extractor transforms the raw data into a high-dimensional feature representation, while the domain discriminator attempts to distinguish whether the data originates from the source domain or the target domain. Higher classification accuracy of the discriminator indicates a greater difference between the two domains, thus quantitatively assessing the degree of domain difference. The discriminator output layer of the domain adversarial network employs a binary classification design (sigmoid activation function) to distinguish whether features originate from the source or target domain. The network training uses an adversarial loss function. ; in, For adversarial loss function; Represents the expectation operator; Symbol representing "obedience"; The discriminator outputs the probability; Output features for the feature extractor; The natural logarithm operator; and These represent the feature distributions of the source scene and the target scene, respectively. and Representing the feature distribution from the source scene respectively and target scene feature distribution Business data samples obtained from sampling.

[0041] Furthermore, in this embodiment, the domain adversarial network is specifically implemented using a gradient reversal layer, enabling the feature extractor and the domain discriminator to be trained simultaneously but with opposite optimization objectives. Feature Extractor The structure consists of three fully connected layers, each containing 512, 256, and 128 neurons respectively, with ReLU activation function; a domain discriminator. It contains two fully connected layers, with 64 and 1 neurons respectively. The classification layer uses the sigmoid activation function. Training employs mini-batch stochastic gradient descent with a batch size of 64, a learning rate of 0.001, and the Adam optimizer. Domain adversarial training is performed through the following steps: First, the feature extractor is fixed. Parameters, optimize the discriminator Parameters are set to maximize the resistance to loss; then, the discriminator is fixed. The parameters are updated in the feature extractor through the gradient inversion layer. The parameters are optimized to minimize the adversarial loss, thereby achieving an accurate measurement of distributional dissimilarity. This alternating optimization process iterates until the network converges, typically requiring 50–100 epochs. After training, the domain dissimilarity metric is calculated as follows: the discriminator's classification accuracy is measured on a balanced test set, compared to 0.5, and the absolute value of the difference is multiplied by 2 to obtain the final domain dissimilarity metric. Parameters This represents the classification accuracy of the discriminator, and the method can effectively reflect the degree of distribution difference between the source domain and the target domain.

[0042] like Figure 3 As shown, the results of domain difference quantification are presented, including the original and weighted offsets of each feature dimension. It can be intuitively seen that the overall offset is mainly contributed by "market competition" and "seasonal factors," supporting the measurement and normalization of domain differences in the paper.

[0043] Step 3: Based on the obtained distribution offset, perform uncertainty-aware knowledge transfer to transfer the parameter probability distribution of the source scene model to the target scene. The transfer strength is adaptively determined by the domain differences. In this step, the system adjusts and migrates the parameter probability distribution of the source scene model based on the domain difference calculation results from step 200.

[0044] The uncertainty-aware knowledge transfer algorithm, as an innovative algorithm in this implementation method, takes the parameter probability distribution of the source scene model as its input. Target scene data and domain difference measurement The output is the probability distribution of the target scene model parameters after the transfer. Unlike traditional transfer learning, which only transfers point estimates of parameters, uncertainty-aware knowledge transfer algorithms transfer the complete parameter probability distribution, preserving uncertainty information. The specific formula is as follows: ; in Represents the model parameters in the target scene. The posterior probability distribution; This means "proportional to", that is, the left side is equal to the right side multiplied by the normalization constant; Indicates in the parameter Below, target scene data The likelihood function; This represents the distribution of source scene parameters adjusted according to domain differences. This is the source scene data.

[0045] Adjusted source scene parameter distribution It is to divide the original parameters. Perform a power transformation (i.e., raise to) (The integral is raised to the power of 1), and then normalized to make its integral equal to 1. This allows for adaptive adjustment of the influence of source scene knowledge on the target scene based on domain differences.

[0046] in It is based on the differences in fields An adaptively adjustable parameter controls the intensity of knowledge transfer. The value of this parameter dynamically adjusts as the domain differences between the source and target scenarios change: when the domain differences are small, When the value of is close to 1, the transfer process will retain more knowledge from the source scenario; however, when the domain differences are significant, The value of will decrease, approaching 0, indicating that the dependence on source scene knowledge is weakened and the transfer strength is reduced. The adaptive adjustment function changes monotonically, and its specific rate of change can be determined by the parameter. Adjustments should be made. The larger the value, the better. The faster the rate of knowledge transfer decreases as the differences between domains diminish. In this way, the system can flexibly adjust the extent of knowledge transfer based on the similarity of the actual scenario, ensuring the rationality and effectiveness of the transfer process.

[0047] Furthermore, in this embodiment, The parameter value range is [1, 5], with a default value of 2.5. When the source scene and the target scene are highly similar, a smaller value can be selected. A value close to 1 indicates strong knowledge transfer; when the difference is significant, a larger value can be selected. A value close to 5 quickly reduces the reliance on knowledge of the source scene. It's important to note that in the calculation... Previously, there were differences in the fields. It has been standardized to dimensionless values ​​in the interval [0, 1], ensuring Consistency in calculation. This ensures that when comparing and transferring knowledge between different scenarios, the calculation of transfer strength is not affected by the original dimensions of the difference measure, thus ensuring the rationality and consistency of the transfer process.

[0048] Furthermore, the uncertainty-aware knowledge transfer algorithm of this embodiment has the following unique steps: First, the adaptation factor is calculated based on the domain difference metric. Secondly, the parameter distribution of the source scene is adjusted using a power transform. Then, the adjusted distribution is combined with the likelihood function of the target scene, and variational inference techniques are used to calculate the final posterior distribution. This method allows the model to adaptively retain or adjust uncertainty during the transfer process.

[0049] Furthermore, in this embodiment, the specific implementation process of uncertainty-aware knowledge transfer is as follows: Parameter distribution representation: The weight parameters of each layer in a Bayesian neural network... Modeled as a diagonal Gaussian distribution, recording the mean vector and variance ; Power transformation calculation: When performing a power transformation on the parameter distribution of the source scene, the original variance is increased (i.e., scaled using a factor related to domain differences), while the mean remains unchanged. After this processing, the entropy of the parameter distribution increases, and the distribution becomes more dispersed and uncertain, thus reflecting the impact of domain differences on the knowledge transfer process.

[0050] Target scene parameter learning: Combining the target scene data and the adjusted prior distribution, a variational objective function is constructed: ; in, Let be the variational objective function for the target scenario; represents the parameters of the variational distribution; Indicates the parameter Regarding variational distribution Expectations; For parameters Variational distribution; The natural logarithm operator; For parameters Next target scene data The likelihood function; For target scene data; These are model parameters; As an adaptive adjustment factor for domain differences, As a measure of domain differences, it has been normalized to interval; The Kullback-Leibler divergence; The distribution of source scene parameters adjusted according to domain differences. This is the source scene data.

[0051] Optimization process: The objective function is optimized using a stochastic variational inference algorithm with the Adam optimizer. The initial learning rate is set to 0.001. Iterative training is performed on the target scene data, and the number of iterations is dynamically adjusted according to the amount of data, usually 500 to 2000 times. Adam is an adaptive moment estimation optimization algorithm.

[0052] Calculate the posterior distribution: After optimization, obtain the variational posterior distribution. Distribution of parameters of the target scene model after migration ;in For data in the target scene and source scene data The posterior distribution of model parameters under given conditions. The entire process is implemented within a computational graph framework, ensuring computational efficiency and numerical stability.

[0053] Furthermore, in this embodiment, the variational objective function for learning the target scene parameters includes time-dimensional constraints, which are fully expressed as follows: ; in, The time-constrained variational objective function for the target scenario; These are variational distribution parameters; The time step for transfer learning, with a range of ; This is the maximum number of iterations for transfer learning; the default value is 1000. For time-dependent constraint weighting functions, For time-dependent constraint functions; This is the natural logarithm operator.

[0054] The calculation method is as follows: in the initial stage of transfer learning (i.e. Less than or equal to the preheating period hour), It will progress over time The increase is linear until the maximum constraint weight is reached. ;when Exceeding the preheating period back, Stay No more additions will be made. This represents the maximum constraint weight, with a value ranging from 0.1 to 2.0 and a default value of 0.5. This refers to the number of steps during the warm-up period, typically 100. 10% to 30%, with 20% as the default.

[0055] It is a time-dependent constraint function, defined as: ; in, For time-dependent constraint functions; For the current time step Model parameters; For the previous time step Model parameters; It is an L2 norm; For parameters Data item loss function The gradient; This represents the maximum number of transition iterations. This is the current time step.

[0056] This time-related constraint is in the early stages of transfer learning ( (Smaller) focuses on smooth parameter changes to avoid knowledge mutations; in the later stages ( near Emphasis is placed on gradient constraints to ensure model convergence and stability. To prevent over-constraint during optimization, an early stopping condition is introduced: the objective function change is less than a threshold for five consecutive time steps. If necessary, the transfer learning process should be ended prematurely.

[0057] like Figure 4As shown, the changes in the parameter distribution of key layers in a Bayesian neural network before and after the transfer are illustrated. The comparison of the "mean and standard deviation" of different layers indicates that the parameter uncertainty generally increases after the transfer, which is consistent with the premise of increasing variance by adaptively adjusting the prior and using power transformation when the target data is small.

[0058] Step 3.1: Dynamically adjust the regularization strength of the probability distribution of model parameters according to domain differences to prevent overfitting or underfitting.

[0059] In this step, the system dynamically adjusts the regularization strength of the Bayesian neural network model based on domain differences.

[0060] The adaptive prior regularization algorithm, as an innovative algorithm in this implementation, takes the domain difference as its input. and the parameter distribution after migration The output is the parameter distribution of the target scene model after adding the regularization term. Adaptive prior regularization algorithm based on domain differences Design an adaptive regularization term. An adaptive regularization term is then applied to the parameter distribution of the Bayesian neural network model for the target scene. The strength of the adaptive regularization term is proportional to the magnitude of the domain difference; that is, the greater the domain difference, the stronger the regularization, thus preventing the Bayesian neural network model from over-relying on the knowledge of the source scene. The modified formula for the parameter distribution of the target scene model is as follows: ; in, For data in the target scene and source scene data Posterior distribution of model parameters under given conditions; This means "proportional to", that is, the left side is equal to the right side multiplied by the normalization constant; For parameters Next target scene data The likelihood function; The distribution of source scene parameters adjusted according to domain differences; It is an exponential function; Based on domain differences The automatically adjusted regularization intensity parameter has been normalized to a dimensionless value; For regularization functions, the comprehensive parameters are... and field differences Calculate the regularization penalty term.

[0061] Furthermore, in this embodiment, the regularization function Specifically defined as: ; in, This is a regularization penalty term; This is the L2 regularization coefficient, with a value range of [0.001, 0.1], and a default value of 0.01; For parameters The square of the L2 norm; This is the L1 regularization coefficient, with a value range of [0.05, 0.5], and a default value of 0.1; As a measure of domain difference, it has been normalized to [0, 1]; It is an L1 norm; These are prior parameters.

[0062] Regularization strength parameter Defined as: ; in, This is the regularization strength parameter; The base regularization strength, with a value range of [0.01, 0.1], and a default of 0.05; This is an adjustment factor, with a value range of [1, 10], and a default value of 5; This is a measure of domain differences.

[0063] To ensure dimensional consistency in the calculation of regularization terms, By employing a specific mapping mechanism, parameters with different dimensions and domain differences are converted into dimensionless values, thus enabling... It is dimensionless and can be directly used calculate.

[0064] This implementation also employs a customized regularization function sensitive to domain differences. In this approach, another method for calculating the regularization strength parameter is to multiply the base regularization strength by an exponential function of the domain difference measure. Specifically, the regularization strength increases exponentially with increasing domain differences, where the base regularization strength serves as the initial value, and the base of the exponential function is determined by both the adjustment factor and the domain difference measure. This allows for enhanced regularization strength when domain differences are significant, thereby more effectively suppressing the model's over-reliance on source scene knowledge.

[0065] Furthermore, in this embodiment, the adaptive prior regularization process is affected by time constraints, and the complete temporal optimization objective function is: ; in, The objective function for the time-series optimization of the regularization process; These are model parameters; As a measure of domain differences; For optimized time steps; The basic objective function; The time weighting function; For time-series smoothing constraints.

[0066] ; in, For time-series smoothing constraints; The number of layers in the neural network; For the first Layer in time step Parameters; For the first Layer in time step Parameters; It is the Frobenius norm; It is an exponential function; This is the time decay coefficient, with a value range of [0.01, 0.1], and a default value of 0.05; This is the current time step; The summation symbol is used.

[0067] This timing constraint ensures that the parameters change smoothly during the optimization process, especially in the early stages. When the parameters are relatively small, it provides strong smoothing constraints to avoid drastic parameter fluctuations; while in the later stages (… near (At time) the constraints are gradually relaxed, allowing the parameters to fully adapt to the target scenario data. To ensure computational efficiency, when the objective function changes less than 1 / 3 of the time step over three consecutive time steps... The early stopping mechanism is triggered in time to end the regularization optimization process ahead of schedule.

[0068] Step 4: Utilize the transfer results to generate fusion results with confidence intervals, providing an assessment of decision reliability; In this step, the system uses the Bayesian neural network model transferred to the target scene in step 3 to predict the target scene data.

[0069] The Bayesian prediction algorithm used in this embodiment takes target scene data samples as input. and target scene model parameter distribution The output is the probability distribution of the prediction results. and confidence interval Unlike traditional models that only output point predictions, Bayesian prediction algorithms output a complete prediction distribution and can calculate various uncertainty measures. The specific prediction process is as follows: ; in, For data samples in a given target scene Target scene data and source scene data The probability distribution of prediction results under the given conditions; For parameters Enter below The conditional probability distribution; For data in the target scene and source scene data Posterior distribution of model parameters under given conditions; This is the set of parameters for a Bayesian neural network. Indicates the parameter Integration operations.

[0070] Based on the distribution of model parameters for the target scene, the system generates the average value of the predictions. and confidence interval ,in This represents the mean of the predicted results; The standard deviation of the prediction results; This is the confidence level parameter, determined based on the required confidence level (e.g., at a 95% confidence level). ).

[0071] Furthermore, in this embodiment, the confidence level parameter The value of is directly related to the required confidence level. For a 95% confidence level, the confidence level parameter is... The value is 1.96; for a 99% confidence level, the confidence level parameter is... The value is 2.576; for a 90% confidence level, the confidence level parameter is... The value is 1.645. By default, this system uses a 95% confidence level (confidence level parameter). =1.96) generates confidence intervals for the prediction results to balance the reliability and practicality of the confidence intervals. Users can also adjust the confidence level parameter according to specific business needs. For high-risk decision-making scenarios, a higher confidence level (such as 99%) can be selected, while for scenarios with lower accuracy requirements, a lower confidence level (such as 90%) can be selected.

[0072] The predictions generated by Bayesian neural network models need to be decoded and transformed in practical applications to enable actionable decision-making. The system provides a set of decoding and transformation functions to convert the output of the Bayesian neural network model into a business-understandable form. These functions include: for regression tasks, inversely transforming the standardized output back to its original units and dimensions; for classification tasks, converting the probability distribution into decision levels or specific classification labels; and converting the quantified uncertainty results into visualized risk levels. These transformations ensure that the output of the Bayesian neural network model can be directly understood and manipulated by business personnel, thereby supporting practical business decisions.

[0073] It should be noted that the uncertainty of the model comes from two sources: cognitive uncertainty (uncertainty of model parameters) and data uncertainty (inherent variation caused by data noise). This method considers both types of uncertainty simultaneously.

[0074] Furthermore, the prediction confidence interval generation algorithm in this embodiment employs the Monte Carlo importance sampling method for effective probability integral calculation, avoiding the computational complexity problem caused by directly calculating multiple integrals. The algorithm samples multiple sets of parameters from the parameter distribution, calculates the output of a Bayesian neural network model for each set of parameters, and then calculates the mean and variance of these outputs to obtain the confidence interval of the prediction result. This method makes the predicted confidence interval more comprehensive and accurate.

[0075] Furthermore, in this embodiment, the specific execution process of Monte Carlo importance sampling is as follows: Parameter sampling: from variational posterior distribution Extraction Group network parameter samples ,in The value is typically set between 50 and 200 to balance computational cost and estimation accuracy. Indicates the number of parameter groups sampled. , , They represent group 1, group 2, and group 3 respectively. Group network parameter samples; Forward prediction: For each set of sampled parameters and input data, the system will perform forward inference of the model to obtain the corresponding prediction output.

[0076] Statistical summary: The system will perform statistical analysis on all the predicted results obtained from sampling, and calculate the mean and variance of these results for subsequent uncertainty analysis.

[0077] Uncertainty decomposition: The system divides the total uncertainty into two parts: cognitive uncertainty (i.e., the uncertainty of the model parameters themselves) and data uncertainty (i.e., the uncertainty caused by noise in the data itself). For regression tasks, data uncertainty can be estimated using the additional output of the Bayesian neural network; for classification tasks, the system calculates the model's reliability score based on the entropy value of the predicted probability distribution. The higher the entropy value, the more uncertain the model is.

[0078] Confidence interval generation: Based on the predicted mean and standard deviation obtained from the aforementioned statistics, and combined with the user-defined confidence level, the system automatically generates confidence intervals for the predicted results. The entire sampling and calculation process is implemented using an efficient vectorization method, improving computational efficiency.

[0079] Step 5: Based on the actual feedback from the target scenario, update the parameters of the uncertainty model and optimize the migration strategy; In this step, the system collects actual feedback data of the Bayesian neural network model running in the target scene and updates the parameters of the uncertain model.

[0080] The feedback-based model parameter update algorithm used in this implementation takes as input the comparison between actual observed data and predicted values, and outputs the updated Bayesian neural network model parameter distribution. The system primarily collects two types of feedback data: the accuracy of the prediction results and the validity of the confidence intervals. Based on this feedback data, the system further adjusts the parameters and transfer strategy of the Bayesian neural network model.

[0081] Furthermore, the feedback update mechanism in this embodiment sets two key indicator thresholds: a prediction accuracy threshold and a confidence interval coverage threshold. The prediction accuracy threshold is set according to the task type. For regression tasks, the mean absolute error (MAE) is used as the indicator, with a threshold range of [0.5, 1.5] times the acceptable error, and a default of 1. For classification tasks, the F1 score is used as the indicator, with a threshold range of [0.7, 0.9], and a default of 0.8. The confidence interval coverage threshold corresponds to the confidence level. For a 95% confidence level, the coverage threshold range is [0.93, 0.97], and a default of 0.95. When the actual coverage is lower than the lower limit of the threshold, the system increases the uncertainty estimation; when the actual coverage is higher than the upper limit of the threshold, the system decreases the uncertainty estimation. These threshold settings ensure the quantifiability and controllability of the model update process.

[0082] Furthermore, the feedback-based model parameter update algorithm in this embodiment uses a calibration scoring metric to evaluate the quality of the uncertainty estimation. Specifically, the system calculates the coverage (the proportion of actual observations falling within the confidence interval) and the interval width of the predicted confidence interval, and dynamically adjusts the uncertainty parameters of the Bayesian neural network model based on these metrics. If the confidence interval is too wide, it indicates that the Bayesian neural network model has overestimated the uncertainty, and the system will shrink the uncertainty estimation; conversely, if the actual observations frequently fall outside the confidence interval, it indicates that the Bayesian neural network model has underestimated the uncertainty, and the system will expand the uncertainty estimation. This adaptive update method ensures that the uncertainty estimation of the Bayesian neural network model always remains consistent with the actual situation.

[0083] like Figure 5 As shown, the changes in confidence interval coverage for each product category before and after model updates based on actual feedback are illustrated. The model converges towards the ideal coverage overall, indicating that uncertainty calibration and parameter updates improve the effectiveness of the intervals.

[0084] A cloud-based multi-source data fusion system, used to execute the aforementioned cloud-based multi-source data fusion method, includes: The data acquisition module is used to acquire source scene data and target scene data; The uncertainty modeling module is used to model the uncertainty distribution of the source scene using a Bayesian neural network; The domain difference calculation module is used to calculate the degree of distribution offset between the source scene and the target scene; The knowledge transfer module is used to perform uncertainty-aware knowledge transfer, transferring the parameter probability distribution of the source scene model to the target scene. The results generation module is used to generate fusion results with confidence intervals, providing an assessment of decision reliability. The parameter update module is used to update the parameters of the uncertainty model and optimize the migration strategy based on the actual feedback from the target scenario.

[0085] Here, the present invention provides an implementation example: A large e-commerce platform needs to quickly establish intelligent pricing models for new markets in different regions. The platform already has a mature pricing model in market A (the source scenario), and now needs to migrate it to the newly developed market B (the target scenario). However, consumer behavior, competitive environment, and seasonal factors in market B are significantly different from those in market A, and historical data for market B is limited.

[0086] The system first acquires product pricing data from market A (source scenario) and market B (target scenario). Market A contains 200,000 product pricing records, while market B contains only 5,000 records. The table below shows a sample of some source scenario data after preprocessing: Table 1: Sample of Product Pricing Data from the Source Scenario (Market A)

[0087] The table below shows a partial data sample of the target scene: Table 2: Sample of Product Pricing Data for the Target Scenario (Market B)

[0088] The system uses a Bayesian neural network to model the uncertainty distribution of the source scene. The Bayesian network uses a variational inference algorithm to output the posterior distribution, which includes not only the optimal parameter values ​​but also the uncertainty information of the parameters. The probability distribution is calculated through Monte Carlo sampling (sampling time set to 10 times).

[0089] Subsequently, the system applies a domain difference quantization algorithm to calculate the distribution offset between the source and target scenes. The system implements domain difference calculation using the maximum average difference principle and kernel method. The domain difference quantization results are shown in the table below: Table 3: Quantitative Results of Domain Differences

[0090] The distribution offset was calculated using a Gaussian kernel function with a kernel width parameter set to 1.0. According to the calculation results, the overall offset between the source and target scenes is 0.2442 (standardized by dividing the original difference measure by 2), indicating a moderate degree of difference between the two scenes. The main differences lie in two dimensions: market competition and seasonality.

[0091] like Figure 2 As shown, this diagram compares the feature distributions of the source scenario (Market A) and the target scenario (Market B) across four dimensions: market competition, inventory levels, seasonality, and promotional activities. The differences in area and line breaks reflect the distribution shift, providing a basis for subsequent calculations of domain difference measurements and adjustments to the migration intensity.

[0092] Based on the domain difference calculation results, the system first analyzes the business data of the source and target scenarios, extracts the scene semantic embedding vectors, and obtains the conceptual level representation of the scenario. Through a scene semantic embedding model with a deep bidirectional attention mechanism, the semantic similarity between the two scenarios is calculated to be 0.68, which is between the similarity thresholds of 0.5 and 0.75. Linear interpolation is then used to determine an appropriate transfer strength.

[0093] Then, the system sets the adaptive parameter to 0.53 and performs uncertainty-aware knowledge transfer, migrating the parameter probability distribution of the source scene model to the target scene. Based on domain differences, the system also dynamically adjusts the regularization strength of the model parameter probability distribution, setting the regularization strength parameter to 0.111. The table below shows a comparison of the parameter distribution of key layers in the network before and after the migration: Table 4: Comparison of parameter probability distributions before and after migration

[0094] It can be observed that the standard deviation of the parameter distribution generally increases after migration, which reflects the increased uncertainty of parameter estimation when there is less data in the target scenario.

[0095] The system utilizes a transferred Bayesian neural network model to predict the pricing of new products in a target scenario. Using the Monte Carlo importance sampling method (drawing 100 sets of network parameter samples from the variational posterior distribution), the forward propagation output of each parameter sample is calculated. Then, the mean and variance of the predicted values ​​are calculated, and confidence intervals are generated based on a 95% confidence level. The table below shows some of the prediction results: Table 5: Pricing Forecast Results for the Target Scenario (Market B)

[0096] By using confidence intervals, merchandise operators can understand the degree of pricing uncertainty and adopt more flexible pricing strategies for products with high uncertainty, such as more frequent price adjustments or further market research.

[0097] Compared with the traditional transfer learning model, the transferred uncertainty-aware pricing model reduced the average pricing error by 18.5% on the test set in market B, and the coverage of the prediction confidence interval reached 94.3%, close to the theoretical value of 95%. More importantly, the system can accurately identify products with high uncertainty, helping the operations team avoid multiple potential pricing risks and reducing sales and profit losses caused by improper pricing.

[0098] The system collected one month's worth of operational data from the target scenario as actual feedback, including actual pricing and sales data, and compared it with model predictions. Based on this feedback data, the system calculated an actual coverage rate of 94.3% for the predicted confidence interval, which falls within the preset 95% confidence level threshold, indicating that the uncertainty estimation is generally reasonable. However, for some product categories (such as clothing), the coverage rate was too low (91.8%), and the system accordingly increased the uncertainty estimation for these products. Meanwhile, for household goods and food categories, the confidence interval was too wide (coverage rate reached 98.2%), and the system appropriately narrowed the uncertainty estimation for these categories.

[0099] Through this feedback-based model parameter update mechanism, the system continuously optimized the uncertainty estimation, making the predicted confidence interval more accurate. In the following three months of operation, the coverage rate of each category remained stable between 94.5% and 95.5%, indicating that the system's estimation of uncertainty was more accurate.

[0100] Practice has shown that the uncertainty perception transfer method of this implementation effectively solves the model transfer problem under the scarcity of new market data in the actual business scenario of e-commerce pricing, and provides reliable uncertainty assessment, providing a more comprehensive basis for business decision-making.

[0101] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-source data fusion method based on cloud computing, characterized in that, include: Acquire source scene data and target scene data, and model the uncertainty distribution of the source scene using a Bayesian neural network to obtain network parameters with probability distribution; Based on network parameters, the domain difference between the source scene and the target scene is calculated to determine the degree of distribution offset; Based on the obtained distribution offset, uncertainty-aware knowledge transfer is performed to transfer the parameter probability distribution of the source scene model to the target scene. The transfer strength is adaptively determined by the domain differences. Using the migration results, fusion results with confidence intervals are generated to provide an assessment of decision reliability; Based on the actual feedback from the target scenario, update the parameters of the uncertainty model and optimize the migration strategy.

2. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, The Bayesian neural network adopts a hierarchical structure design, including an input layer, multiple hidden layers and an output layer. The weight parameters of each layer are modeled as probability distributions rather than fixed values.

3. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, The domain difference is calculated using the maximum average difference principle and kernel method. By quantifying the data distribution of the source scene and the target scene, the degree of distribution offset between the data distribution of the source scene and the target scene is evaluated, which serves as the basis for subsequent knowledge transfer.

4. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, When performing uncertainty-aware knowledge transfer, the parameter adjustment adopts an adaptive transfer intensity parameter. Based on the results of domain difference measurement, the parameter is dynamically adjusted through a preset decay rate parameter to achieve precise control of the knowledge transfer intensity.

5. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, The confidence interval of the fusion result is generated by a Bayesian prediction algorithm. The Monte Carlo importance sampling method is used to extract multiple sets of parameters from the parameter distribution and calculate the mean and variance of the predicted values.

6. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, After acquiring source scene data and target scene data, the business data is analyzed, scene semantic embedding vectors are extracted, and a conceptual representation of the scene is obtained.

7. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, After performing the knowledge transfer with uncertainty perception, the regularization strength of the model parameter probability distribution is dynamically adjusted based on the quantification results of domain differences. An adaptive prior regularization method is adopted to prevent the model from overfitting or underfitting.

8. The multi-source data fusion method based on cloud computing according to claim 7, characterized in that, The adaptive prior regularization optimizes the model's generalization ability by automatically adjusting the regularization strength parameter based on domain differences and combining it with the regularization function to adaptively constrain the probability distribution of model parameters.

9. The multi-source data fusion method based on cloud computing according to claim 1, characterized in that, The process of updating the uncertainty model parameters based on actual feedback from the target scenario includes calculating the coverage and width of the prediction confidence interval, and dynamically adjusting the uncertainty parameters of the Bayesian neural network model based on these indicators.

10. A cloud computing-based multi-source data fusion system, characterized in that, The method for performing the cloud computing-based multi-source data fusion method according to any one of claims 1-9 includes: The data acquisition module is used to acquire source scene data and target scene data; The uncertainty modeling module is used to model the uncertainty distribution of the source scene using a Bayesian neural network; The domain difference calculation module is used to calculate the degree of distribution offset between the source scene and the target scene; The knowledge transfer module is used to perform uncertainty-aware knowledge transfer, transferring the parameter probability distribution of the source scene model to the target scene. The results generation module is used to generate fusion results with confidence intervals, providing an assessment of decision reliability. The parameter update module is used to update the parameters of the uncertainty model and optimize the migration strategy based on the actual feedback from the target scenario.

Images