Data asset valuation method based on machine learning fusion of benefits and market factors

Abstract

The application relates to the technical field of data asset valuation, and discloses a data asset valuation method based on machine learning fusion of benefits and market factors. The method collects operation benefit indexes and market fluctuation indexes of a target data asset, the operation benefit indexes include data usage frequency, completeness, update cycle and call cost saving value, and the market fluctuation indexes include industry demand growth rate, competitive product substitution rate, policy influence coefficient and technical iteration risk value. The operation benefit indexes are input into a pre-trained random forest regression model to obtain an initial benefit valuation vector, and the market fluctuation indexes are input into a pre-trained time sequence convolutional neural network to generate a dynamic discount rate matrix. Multidimensional tensor interaction calculation is performed on the two, a fusion value feature tensor is output, nonlinear transformation processing is performed, a final valuation result is generated, and the final valuation result is stored in a blockchain storage system. The method fuses operation benefits and market factors, improves valuation accuracy and dynamic adaptability, and guarantees result reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data asset valuation, in particular to a data asset valuation method based on machine learning fusion of benefits and market factors. BACKGROUND

[0002] With the rapid development of digital economy, the value of data as a new type of production factor is increasingly prominent, and the valuation of data assets has become an important link in market transactions, investment and financing activities, and enterprise financial management. Currently, data asset valuation methods mainly include cost method, income method and market method, but these traditional methods have obvious limitations in practical application.

[0003] The cost method takes the construction cost of data assets as the core for valuation, only considering the input of data collection, storage and processing, but ignoring the dynamic benefits generated during the use of data, which makes it difficult to reflect the actual application value of data assets. Although the income method focuses on the future income brought by data assets, it usually relies on static discount rate calculation, which cannot cope with the rapid changes in market environment, resulting in a large deviation between the valuation result and the actual market value. The market method takes the transaction price of similar data assets as a reference, but data assets have unique and non-standardized characteristics, and the number of transaction cases that can be referred to is limited, and it is difficult to quantify the quality differences between different data assets, which greatly reduces the accuracy of valuation.

[0004] Existing valuation methods do not consider the multi-dimensional factors that affect the value of data assets. At the operation level, the use frequency and completeness of data directly determine its actual utility, but traditional methods often simplify the weight allocation of such indicators; at the market level, external factors such as industry demand fluctuations, policy adjustments and technology iterations will significantly affect the market acceptance and value stability of data assets, but existing models lack the ability to capture and integrate these dynamic factors in real time. In addition, most valuation methods lack effective result storage mechanisms, and the transparency of the valuation process and the credibility of the results are difficult to guarantee, which further restricts the circulation and application of data assets in the market. SUMMARY

[0005] The purpose of the present application is to provide a data asset valuation method based on machine learning fusion of benefits and market factors to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides a data asset valuation method based on machine learning fusion of benefits and market factors, which comprises:

[0007] Collect operation benefit indicators and market fluctuation indicators of the target data asset, the operation benefit indicators including data usage frequency, data completeness, data update period, data call cost saving value, the market fluctuation indicators including industry demand growth rate, competitive product substitution rate, policy influence coefficient, and technology iteration risk value;

[0008] Input the operation benefit indicators into a pre-trained random forest regression model to perform benefit benchmark valuation, to obtain an initial benefit valuation vector; input the market fluctuation indicators into a pre-trained time series convolutional neural network to perform market discount rate prediction, to generate a dynamic discount rate matrix;

[0009] Perform multi-dimensional tensor interaction calculation on the initial benefit valuation vector and the dynamic discount rate matrix through a feature cross fusion module, to output a fusion value feature tensor;

[0010] Perform nonlinear transformation processing on the fusion value feature tensor by using a fully connected valuation network, to generate a final data asset valuation result and store the result in a blockchain storage system.

[0011] Preferably, the construction process of the pre-trained random forest regression model includes: obtaining operation benefit indicator samples and corresponding valuation labels of historical data assets, the operation benefit indicator samples including data call delay time, data error rate, data correlation dimension number, and data service availability;

[0012] Construct a decision tree forest structure and set a tree depth constraint condition, the tree depth constraint condition including maximum split number limit, node purity threshold, and feature sampling ratio;

[0013] Perform a recursive feature splitting operation based on a Gini impurity minimization principle, to generate a plurality of decision tree models;

[0014] Calculate weight coefficients of the decision tree models by using an out-of-bag error estimation method;

[0015] Weight-aggregate output results of the plurality of decision tree models based on the weight coefficients, to complete construction of the pre-trained random forest regression model.

[0016] Preferably, the training process of the pre-trained time series convolutional neural network includes: collecting historical market fluctuation indicator sequences and corresponding discount rate labels, the historical market fluctuation indicator sequences including industry concentration variation curves, price elasticity coefficient sequences, and capital inflow rate time series;

[0017] Construct a multi-scale dilated convolution layer structure, the multi-scale dilated convolution layer structure containing convolution kernel groups with different dilation rates;

[0018] Capture long-term time series dependency features by using a gated recurrent unit, to output a time series feature encoding vector;

[0019] The attention mechanism is used to weight the feature importance of the time sequence feature encoding vector;

[0020] The network parameters are optimized by using the mean square error loss function until convergence, and the pre-trained time sequence convolutional neural network is generated.

[0021] Preferably, the operation process of the feature cross fusion module includes: mapping the initial benefit estimate vector to a first embedding matrix in a high-dimensional feature space through an embedding layer;

[0022] The dynamic discount rate matrix is converted into a second embedding matrix through orthogonal transformation;

[0023] The Kronecker product of the first embedding matrix and the second embedding matrix is calculated to generate an interaction feature cube;

[0024] The tensor folding operation is performed on the interaction feature cube to output a reduced fusion feature matrix;

[0025] The fusion feature matrix is processed through a batch normalization layer to generate the fusion value feature tensor and passed to the fully connected estimate network.

[0026] Preferably, the operation mechanism of the fully connected estimate network includes: receiving the fusion value feature tensor from the feature cross fusion module;

[0027] A deep neural network architecture including residual connections is constructed, and the deep neural network architecture has multiple hidden layers;

[0028] Random dropout regularization strategy is implemented between hidden layers;

[0029] Nonlinear feature transformation is performed by an exponential linear unit activation function;

[0030] The output layer uses a linear activation function to generate the final data asset estimate result.

[0031] Preferably, the method further includes real-time updating of market fluctuation indicators: establishing an industry monitoring data stream pipeline to continuously collect real-time market fluctuation indicators;

[0032] The real-time market fluctuation indicators are input into an incremental learning engine, and the incremental learning engine includes an online sequential minimum optimization algorithm;

[0033] The convolution kernel weight parameters of the pre-trained time sequence convolutional neural network are dynamically updated;

[0034] The updated dynamic discount rate matrix is transmitted to the feature cross fusion module.

[0035] Preferably, the workflow of the incremental learning engine comprises: setting a sliding time window to capture real-time market fluctuation indicator streams;

[0036] Calculating the cosine similarity between newly arrived market fluctuation indicators and historical features;

[0037] Triggering model update when the cosine similarity is below an adaptive threshold;

[0038] Adjusting the convolution kernel weight parameters using incremental gradient descent method;

[0039] Synchronizing the updated network parameters to the pre-trained time series convolutional neural network.

[0040] Preferably, the method further comprises a verification link of the valuation result: obtaining benchmark valuation data provided by a third-party audit platform;

[0041] Inputting the final data asset valuation result and the benchmark valuation data into a difference analysis module;

[0042] Calculating the relative error percentage between the two;

[0043] Triggering model retraining when the relative error percentage exceeds a tolerance threshold;

[0044] Sending retraining instructions to the pre-trained random forest regression model and the pre-trained time series convolutional neural network.

[0045] Preferably, the execution process of the model retraining comprises: collecting a new data asset sample set, which contains new operating benefit indicators and new market fluctuation indicators;

[0046] Extending the historical training data set to the new data asset sample set;

[0047] Reinitializing the decision tree forest structure;

[0048] Performing feature importance sorting and removing redundant features;

[0049] Accelerating the model training process through parallel computing clusters.

[0050] Preferably, the method further comprises a valuation traceability mechanism: creating a valuation transaction log in a blockchain notarization system, which contains input feature hash values, model version identification, and timestamp information;

[0051] Generating tamper-proof valuation notarization credentials;

[0052] Associating the valuation notarization credentials to the metadata directory of the target data asset;

[0053] Outputting the complete valuation process chain in response to an audit request.

[0054] Compared with the prior art, the present application has the beneficial effects that:

[0055] By fusing the operation benefits and market factors, combined with machine learning technology, the overall and dynamic evaluation of the value of data assets is realized. In terms of operation benefit evaluation, by collecting specific indicators such as data usage frequency, completeness, update cycle and call cost saving value, and using a pre-trained random forest regression model for processing, the utility and value contribution of data assets in actual application can be more accurately reflected, avoiding the estimation deviation caused by the simplified processing of operation indicators in traditional methods.

[0056] For the consideration of market factors, the time series convolutional neural network is used to analyze market fluctuation indicators such as industry demand growth rate, competitor substitution rate, policy influence coefficient and technology iteration risk value, to generate a dynamic discount rate matrix, breaking through the limitations of traditional static discount rate, and can capture the influence of market environment changes on the value of data assets in real time, making the estimation results more in line with the actual market situation.

[0057] Through the multi-dimensional tensor interaction calculation of the initial benefit estimation vector and the dynamic discount rate matrix by the feature cross fusion module, the internal relationship between operation benefits and market factors can be deeply mined, and the fusion value feature tensor formed covers richer value information, providing a comprehensive feature basis for subsequent nonlinear transformation processing. The application of fully connected estimation network further realizes the nonlinear mapping of complex value features, improves the processing capacity of the estimation model for multi-factor interaction, and makes the final estimation result more accurate and reasonable.

[0058] The estimation results are stored in the blockchain storage system, and the non-tamperable characteristics of the blockchain are used to enhance the transparency of the estimation process and the credibility of the results, providing a reliable value basis for subsequent links such as the transaction and audit of data assets, which helps to promote the standardized circulation and market recognition of data assets. Overall, this method integrates multi-dimensional value factors, combines advanced machine learning technology and storage mechanism, and forms a more scientific, dynamic and credible estimation system, which adapts to the complexity and dynamicity of data asset value evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The working principle diagram of the data asset valuation method based on machine learning fusion of benefits and market factors described in the present application;

[0060] Figure 2 The flowchart for constructing a random forest regression model;

[0061] Figure 3 The flowchart for training a time series convolutional neural network;

[0062] Figure 4A flowchart for the mechanism of the fully connected valuation network;

[0063] Figure 5 A flowchart for the operation of the incremental learning engine. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0065] Please refer to Figure 1 The present application provides a data asset valuation method based on machine learning fusion of benefits and market factors, which comprises:

[0066] The dynamic value evaluation of the data asset is realized by constructing a multi-modal machine learning model. The method first collects the operation benefit indicators and market fluctuation indicators of the target data asset, wherein the operation benefit indicators include data usage frequency, data integrity, data update cycle and data call cost saving value, and the market fluctuation indicators include industry demand growth rate, competitive product substitution rate, policy influence coefficient and technical iteration risk value. The operation benefit indicators are input into the pre-trained random forest regression model for benefit benchmark valuation, generating an initial benefit evaluation vector; the market fluctuation indicators are input into the pre-trained time series convolutional neural network for market discount rate prediction, outputting a dynamic discount rate matrix. The feature cross fusion module adopts tensor operation to perform multi-dimensional interaction calculation on the initial benefit evaluation vector and the dynamic discount rate matrix, generating a fusion value feature tensor. The fully connected valuation network performs nonlinear transformation processing on the fusion features, outputs the final data asset valuation result and stores it in the blockchain storage system. The method realizes the double consideration of benefits and market factors through the machine learning model, and adopts the blockchain technology to ensure the traceability of the valuation process.

[0067] Embodiment 1: refer to Figure 2The construction process of the pre-trained random forest regression model involves the coordinated operation of multiple technical links. The sample collection of the operational benefit indicators of historical data assets uses distributed data crawling technology to aggregate data from multiple sources such as enterprise internal databases, cloud storage platforms, and third-party data markets. The sample data undergoes a strict cleaning process, including missing value filling, outlier removal, and data standardization. The evaluation label is obtained by cross-validation of the pricing report of professional evaluation agencies and market transaction records to ensure the authenticity and reliability of the label data. The data call delay time indicator is obtained through API call log analysis, with millisecond-level timestamp records. The data error rate calculation uses a sampling detection method to verify the integrity, consistency, and accuracy of the data set from multiple dimensions. The data correlation dimension number is based on the field mapping relationship in the metadata management system to identify the connection attributes across data tables. The data service availability indicator is extracted from the operation and maintenance monitoring system to comprehensively evaluate the response time, success rate, and stability of the service.

[0068] The construction of the decision tree forest structure uses a parallel algorithm to generate multiple decision trees simultaneously on a distributed computing framework. The maximum split number limit parameter is dynamically adjusted according to the dimensionality of the feature space to avoid overfitting while maintaining sufficient feature expression capability. The node purity threshold setting uses an adaptive algorithm to dynamically optimize based on the performance of the validation set during model training. The feature sampling ratio control introduces a random seed mechanism to ensure the reproducibility of each sample. In the recursive feature splitting phase, the algorithm maintains a priority queue to manage the nodes to be split, prioritizing the feature combinations with the highest information gain. The Gini impurity calculation uses an optimized numerical calculation method to reduce the precision loss caused by floating-point operations.

[0069] The generation process of the decision tree model uses a depth-first search strategy, with each internal node storing the index of the split feature and the split threshold. The leaf nodes store the statistical distribution information of the predicted values, including the mean, variance, and sample size. The model training process implements an early stopping mechanism that terminates the growth of the current tree when the validation set error does not decrease for several consecutive rounds. In the out-of-bag error estimation phase, the training samples of each tree are labeled in reverse, and the prediction error of the samples not selected on the tree is calculated. The weight coefficient calculation introduces a temperature parameter to adjust the output distribution of the softmax function, balancing the contribution differences between different decision trees.

[0070] The model parameters are saved in binary serialization format, including the tree structure array, feature index table, and split threshold list. The feature importance ranking is based on the frequency of feature appearance in the split nodes and the information gain brought by the feature. The model explanation module generates a visual report to show the marginal contribution of each feature to the final predicted value. The model update mechanism is designed in an incremental learning mode, where only part of the decision trees are retrained when new data arrives, maintaining the stability of the overall model structure.

[0071] During the model inference phase, the input features are standardized and then used in parallel to predict outcomes through all decision trees. The predictions from each tree are weighted and summed according to their respective weight coefficients, ultimately outputting an initial benefit estimation vector. The vector dimension matches the output requirements of the target estimation task and includes information such as point estimates and confidence intervals. The model performance monitoring system tracks the prediction error distribution in real time and triggers an alarm mechanism when performance degradation is detected.

[0072] The hyperparameter tuning of the random forest regression model employs Bayesian optimization to find the optimal parameter combination under given computational resource constraints. Stratified sampling is used for cross-validation to maintain consistency in the data distribution between the training and validation sets. The model deployment environment is configured as a containerized service, supporting horizontal scaling to handle high-concurrency evaluation requests. The service interface design follows RESTful specifications, and input / output data is encapsulated in JSON format.

[0073] The model's security measures include input data validation, model fingerprint signing, and access control lists. Audit logs record all model call requests, including timestamps, request parameters, and response results. Semantic version control is used for model version management, ensuring backward compatibility for major updates. Performance benchmarks cover datasets of varying sizes and hardware configurations, establishing a standardized performance evaluation system.

[0074] Model interpretability enhancement techniques include locally interpretable models and feature contribution analysis. The decision path tracing function can reconstruct the inference chain corresponding to any prediction result. The uncertainty quantification module estimates the reliability of predicted values, assisting decision-makers in assessing risk. Integration with upstream and downstream systems achieves asynchronous communication through message queues, ensuring system decoupling and high availability.

[0075] The model's monitoring metrics include operational metrics such as real-time throughput, response latency, and error rate, as well as business metrics such as feature distribution offset and prediction deviation. Alarm rules are configured with multi-level thresholds, triggering different levels of processing based on the severity of the problem. Disaster recovery solutions include regular model snapshot backups and a rapid recovery mechanism to ensure business continuity. Performance optimization measures include prediction result caching, batch request processing, and computation graph optimization.

[0076] The continuous improvement process establishes an automated data loop, incorporating prediction results and actual performance feedback from the production environment into the retraining data pool. A data quality detection module identifies annotation noise and distribution shift issues. A model degradation detection algorithm analyzes time-series patterns in prediction errors to identify potential concept drift. An A / B testing framework supports comparing the performance of new and old models, employing statistical testing methods to evaluate the improvement.

[0077] The compliance design of the model considers data privacy protection requirements, implements data desensitization and access audit. The verifiability of the computing process is realized by generating proof documents, recording the intermediate results of key computing steps. The fairness of the model detects the prediction differences between different groups, identifies potential discriminatory biases. Transparency enhancement measures include publishing model card documents, disclosing training data composition and model limitations.

[0078] The service governance architecture of the model adopts a microservice design pattern, achieving loose coupling integration with other components. The load balancing strategy dynamically adjusts resource allocation according to real-time traffic. The fault-tolerant mechanism handles the case of partial decision tree prediction failure, and continues to serve through the consensus result of the remaining trees. The gray release strategy controls the exposure range of new model versions, gradually verifying their stability.

[0079] Resource utilization optimization of the model includes memory pool management, computing task scheduling and heterogeneous device acceleration. Energy-saving design considers the energy efficiency ratio of computing-intensive stages, seeking a balance between performance and power consumption. Hardware-aware optimization generates specific instruction sets for different processor architectures, improving computing efficiency. Edge computing supports offloading part of the inference task to terminal devices, reducing network transmission overhead.

[0080] The ecosystem integration of the model includes metadata synchronization with the data catalog system, realizing unified management of feature definitions. Integration with workflow engines supports visual modeling and automated scheduling. Interface with monitoring and alarm platforms enables cross-system fault localization. Connection with BI tools supports visual analysis and report generation of prediction results.

[0081] Embodiment 2: Referring to Figure 3 The construction process of the time series convolutional neural network starts from the collection and preprocessing of historical market volatility indicator sequences. The industry concentration change curve is derived from the industry association's regular reports and listed company public financial data, and through time alignment and interpolation processing, a continuous time series is formed. The price elasticity coefficient sequence is calculated based on the historical records of commodity pricing and sales quantity change data, using a sliding window statistical method to generate. The capital inflow rate time series extracts the original settlement records from the financial transaction system, and aggregates the fund flow according to the preset time granularity. These time series data are normalized to eliminate dimensional differences, and abnormal detection algorithms are used to identify and correct noise and outliers in the data collection process.

[0082] The design of the multi-scale dilated convolution layer structure adopts a hierarchical feature extraction strategy. Different groups of convolution kernels with different expansion rates process the input time series data in parallel, capturing short-term fluctuations, medium-term trends, and long-term periodic characteristics. The orthogonal matrix method is used for parameter initialization of each group of convolution kernels to maintain independence between different convolution kernels. The convolution operation uses a causal padding method to avoid future information leakage into the prediction of the current time point. The design of the gated recurrent unit layer adopts a full gating structure, and the weight matrices of the reset gate and the update gate are initialized separately. The dimension of the unit state vector matches the output features of the convolution layer. During the generation of the time series feature encoding vector, the gated mechanism dynamically adjusts the retention ratio of historical information, and the importance of the current input features determines the degree of state update.

[0083] The implementation of the attention mechanism adopts a multi-head self-attention architecture. The time series feature encoding vector is first transformed into a query vector, a key vector, and a value vector through linear transformation, and the transformation matrices of different attention heads are independently parameterized. The attention weight calculation uses the scaled dot-product method, and the temperature coefficient is used to adjust the sharpness of the weight distribution. The aggregation process of feature importance weighting preserves the output results of each attention head, and the multi-perspective attention features are integrated through concatenation and linear transformation. The optimization process during network training uses the adaptive moment estimation algorithm, and the learning rate is dynamically adjusted according to the gradient changes. When calculating the loss function, a sequence weighting strategy is introduced for long sequence prediction tasks, and recent prediction errors are given higher penalty weights.

[0084] The embedding layer design of the feature cross-fusion module uses a learnable projection matrix. The initial benefit estimate vector is mapped to a high-dimensional space through a multi-layer perceptron, and the activation function selects a smooth nonlinear transformation. The orthogonal transformation of the dynamic discount rate matrix is implemented using the QR decomposition algorithm to maintain the numerical stability of the matrix. Before computing the Kronecker product, the two embedding matrices are aligned in dimension to ensure the compatibility of matrix operations through padding or truncation. The generation process of the interaction feature cube uses block computation to optimize memory usage, and large tensor operations are optimized through a delay execution strategy to improve computational efficiency. The tensor folding operation unfolds the high-dimensional tensor along a specific pattern, and reduces the feature dimension through matrix decomposition techniques.

[0085] The implementation of the batch normalization layer considers the computational efficiency of small batch statistics. The update of the moving average statistics uses a mixed precision algorithm to reduce the rounding error of floating point operations. The scaling and translation parameters in the normalization process are independently initialized to avoid optimization stagnation caused by symmetry. The transfer of the fusion value feature tensor uses a memory sharing mechanism to reduce the copying overhead of large-scale feature matrices. The network parameters are saved using a hierarchical serialization strategy, and the parameters of different components are stored separately for partial updates.

[0086] The inference process of the temporal convolutional neural network adopts an incremental calculation mode. Newly arrived market volatility data is input into the network through a sliding window, and only the affected time step output is calculated. The cache mechanism of the historical hidden state avoids repeated calculation and improves real-time prediction performance through state reuse. The dynamic discount rate matrix output by the network is attached with uncertainty estimation, and the confidence interval of the predicted value is calculated through the Monte Carlo sampling method. The hot update mechanism of the model supports dynamic loading of parameters, and the model version switching can be completed without interrupting the service.

[0087] The visualization analysis tool of the network architecture generates a computation graph topology, showing the data flow of feature transformation in each layer. The feature importance analysis module identifies time points and index dimensions that have a significant impact on the prediction result. The model explanation report contains decision path analysis of typical prediction cases, showing how key features affect the final output through network layers. The network robustness test evaluates the stability of the model under data perturbation by injecting noise and adversarial samples.

[0088] Data interaction with upstream and downstream systems adopts a unified time alignment protocol. The timestamp of input time series data is accurate to the millisecond level, and the consistency in distributed environments is maintained through a global clock service. The version control of output data uses a logical timestamp marker, supporting the tracking and reasoning of causal relationships. The error handling mechanism detects the integrity of input data and initiates a repair process for missing or abnormal time steps.

[0089] Dynamic allocation of computing resources automatically adjusts according to the length of the input sequence. Long sequence processing enables memory optimization mode, reducing the storage requirements of intermediate results through calculation reordering. Hardware accelerators automatically detect and utilize available GPU resources, enabling deep optimization kernels for convolution operations. The energy management module monitors the power consumption of the calculation process, dynamically balancing between energy efficiency and performance.

[0090] The monitoring indicator system of the model covers multiple dimensions. Data quality monitoring tracks changes in the statistical properties of input time series data, detecting concept drift. Computing performance monitoring records the execution time distribution of each layer, identifying performance bottlenecks. Business indicator monitoring analyzes the degree of agreement between prediction results and actual market changes. Alarm rules are triggered according to abnormal patterns of monitoring indicators, supporting multi-level response strategies.

[0091] The maintenance process of the model establishes standardized operation specifications. Regular health checks verify the numerical stability of each layer of the network, detecting signs of gradient vanishing or explosion. Parameter distribution analysis tools visualize the statistical properties of weights and biases, identifying abnormal parameter updates. Model pruning and quantization tools support network size compression, adapting to resource constraints in edge computing scenarios.

[0092] Security measures of the model include parameter encryption and integrity check. Digital signature verification of the model file prevents unauthorized modification. Access control of the inference service is based on role-based permission management, and operation logs record a complete prediction request history. Data de-identification processing removes sensitive information in input time series data to meet privacy protection requirements.

[0093] The deployment architecture of the model supports multiple service modes. Embedded deployment compiles network parameters into a platform-independent intermediate representation, adapting to different hardware environments. Cloud service deployment uses an automatically scalable container group to adjust the number of instances according to load fluctuations. Hybrid deployment mode supports offloading part of the computing task to edge nodes, reducing the pressure on the central server.

[0094] Integration with business systems uses standardized interfaces. Prediction results are transmitted asynchronously through message queues, supporting parallel processing by multiple consumers. Data caching mechanism stores recent prediction results, reducing resource consumption for repeated calculations. Transaction management ensures the atomicity and consistency of prediction requests, automatically rolling back partially completed operations in abnormal situations.

[0095] The documentation system of the model includes detailed technical specifications. Architecture documents describe the functional design and interface conventions of each layer. Training documents record data preparation and hyperparameter tuning processes. Deployment documents provide environment configuration and performance tuning guidelines. Maintenance documents explain the daily inspection and fault handling processes. User manuals explain input and output data formats and error code meanings.

[0096] Example 3: refer to Figure 4 The architecture design of the fully connected valuation network uses a deep residual structure, with skip connections established between network layers to facilitate gradient flow. The input layer receives a 512-dimensional fusion value feature tensor from the feature cross-fusion module, and the tensor elements are standardized before entering the first hidden layer. The neuron configuration of the hidden layer follows the pyramid decreasing principle, with the number of neurons in each layer reduced in an equal ratio sequence. The implementation of residual connection uses feature map addition, and the skip path introduces a 1x1 convolution kernel to adjust the dimension matching. Network parameter initialization uses the He normal distribution method, and the bias term is initialized to a small positive value to avoid dead neuron phenomenon.

[0097] The regularization strategy of the network integrates multiple technical means. The random dropout layer randomly masks neuron outputs with a preset probability during the training phase, and uses scaling compensation to maintain the expected output unchanged during the testing phase. The dropout probability is set using a hierarchical differentiated strategy, with a lower dropout rate near the input layer and a higher dropout rate in deep network. The L2 regularization term is applied to the weight matrix of the fully connected layer, and the regularization coefficient is dynamically adjusted according to the layer depth. The early stopping mechanism monitors the validation set loss curve, and terminates the training process when the loss plateau persists for more than a predetermined period.

[0098] The choice of activation function considers numerical stability and computational efficiency. The computational expression of exponential linear unit is is:

[0099]

[0100] where, represents the neuron input signal, controls the saturation value of the negative half-axis, set as a learnable parameter automatically optimized during network training. The gradient calculation of the activation function is implemented using automatic differentiation, and a smoothing approximation is introduced to handle numerical underflow problems. The linear activation of the output layer maintains the continuous scale characteristics of the evaluation results, avoiding the distortion of the prediction range caused by nonlinear transformation.

[0101] The real-time updating system of market volatility indicators builds a distributed data processing pipeline. Industry monitoring data streams are distributed through message broker middleware, and a publish-subscribe mode is used to realize multi-consumer parallel processing. The division of data stream windows is based on event time semantics, and the handling of late data uses watermarking mechanism and side output stream. The online sequence minimization algorithm of the incremental learning engine maintains the dynamic working set, supporting incremental updating of parameters. The adjustment of convolution kernel weights uses stochastic gradient descent with momentum, and historical gradient information is decayed through exponential moving average.

[0102] The dynamic updating mechanism of the time series convolutional neural network designs a double verification process. Newly arrived market data is first filtered through an outlier detection filter, using the local outlier factor algorithm to identify abnormal fluctuations. The data batch verified triggers the model update signal, and the update strength is adaptively adjusted according to the degree of data distribution deviation. The adjustment of convolution kernel parameters uses a hierarchical learning rate strategy, with a lower learning rate for shallow networks to maintain basic feature extraction capabilities, and a higher learning rate for deep networks to adapt to rapidly changing patterns.

[0103] The fault-tolerant design of network services implements a multi-level protection mechanism. Range checks on input data verify that feature values are within the training data distribution range, and abnormal values trigger data repair processes. Numerical stability monitoring of network calculations detects signs of gradient explosion and disappearance, and automatically adjusts the step size parameter. Service degradation strategies automatically switch to simplified models when the system is overloaded, ensuring basic service quality. Disaster recovery schemes maintain multiple model replicas, seamlessly switching to backup replicas in the event of primary replica failure.

[0104] Performance optimization techniques run throughout the entire network architecture. Matrix multiplication operations use block algorithms to improve cache hit rates, and large tensor operations use memory prefetching to reduce latency. Parallel computing tasks are automatically scheduled based on hardware resources, with CPU-intensive operations overlapping with IO operations. Memory management uses an object pool mode to reuse intermediate result buffers, reducing dynamic memory allocation overhead. The computation graph optimizer analyzes data dependencies to eliminate redundant calculations and combine similar operations.

[0105] The explainability enhancement techniques of the network provide decision insights. Feature importance analysis employs the integral gradient method to quantify the contribution of each input feature to the final estimate. Counterfactual explanation generators construct contrastive samples close to the decision boundary, demonstrating how changes in key features affect the prediction results. Attention visualization tools mark the key time points and feature dimensions that the network focuses on, assisting manual review of model behavior.

[0106] The monitoring system implements full-link observability. Performance indicator collection includes system-level indicators such as computing delay, memory occupation, and CPU utilization at each layer. Business indicator tracking predicts the statistical distribution, volatility, and abnormal prediction proportion of values. The log system records the complete request processing pipeline, supporting end-to-end transaction tracking. The dashboard aggregates multi-dimensional monitoring data, providing a real-time system health status view.

[0107] The deployment architecture supports elastic expansion and rolling upgrade. Containerization encapsulates the network model and dependent environment as an independent unit, ensuring consistency of the running environment. The automatic expansion controller dynamically adjusts the number of service instances according to request load, maintaining resource utilization within the target interval. Version release adopts the blue-green deployment strategy, with new and old versions running in parallel and gradually switching traffic. The configuration management center centrally manages hyperparameters and switch settings, supporting dynamic adjustment at runtime.

[0108] Security measures cover the entire lifecycle of data and processing. Data in transit is encrypted using the TLS protocol, and static model parameters are stored using the AES algorithm. Access control implements an attribute-based permission model, providing fine-grained control over the invocation permissions of the prediction service. Audit logs record all sensitive operations, including model updates, parameter modifications, and data access. Anti-injection attack filters clean input data, blocking maliciously constructed feature vectors.

[0109] Integration with external systems uses standardized interfaces. Data input and output formats comply with industry-standard specifications, supporting multiple serialization methods such as JSON and ProtocolBuffers. Service registration and discovery mechanisms automatically maintain a list of available instances, optimizing request distribution through client-side load balancing. Asynchronous communication is decoupled through message queues, ensuring eventual consistency for important operations. The cache layer stores hot prediction results, reducing the computational load on the backend.

[0110] Maintenance procedures establish systematic operation specifications. Daily inspection checks model performance indicators and service availability status. Regular maintenance windows perform parameter tuning and garbage collection. Change management processes control configuration modifications and version upgrades, implementing impact assessments and rollback plans for major changes. Capacity planning predicts resource requirements based on historical growth trends, allowing for early expansion of key components.

[0111] The documentation system provides comprehensive technical reference. The architecture design document describes system components and interaction relationships. The API reference manual details service interfaces and error codes. The operation manual contains daily operation and fault handling guidelines. The model card document discloses training data, intended use, and limitations. The knowledge base accumulates common problem solutions and best practice cases.

[0112] Example 4: Refer to Figure 5 The workflow of the incremental learning engine is specifically demonstrated by the case of financial industry data asset valuation. A commercial bank needs to dynamically value its customer credit scoring data asset, which is in a market environment that is constantly changing. The system establishes a data stream processing pipeline to collect market fluctuation indicators, including inter-bank product pricing changes, regulatory policy adjustment signals, and market demand fluctuations, with a 10-minute window. The configuration parameters of the sliding time window are determined according to business characteristics, with a 24-hour window reflecting daily trends and a 1-hour sliding step capturing recent changes. The raw data in the window is cleaned and converted into structured records, which are stored in a time series database for subsequent analysis.

[0113] The similarity calculation section uses feature engineering methods to process market fluctuation indicators. The following table shows an example of market fluctuation indicators in a typical time window.

[0114] Table 1: Shows an example of market fluctuation indicators in a typical time window

[0115]

[0116] In the feature vector construction process, the indicator sequence of each time window is aggregated by statistics to generate a 128-dimensional feature encoding. The feature vector is L2 normalized before cosine similarity calculation to eliminate the bias caused by dimension difference. The adaptive threshold mechanism is initially set to 0.85, and the system maintains a queue of the last 20 similarity calculation results. When the proportion of results below the threshold in the queue exceeds 60%, the model is updated. This double verification mechanism avoids frequent model updates caused by temporary fluctuations.

[0117] The model update execution process uses an online learning mode. The incremental gradient descent method is configured with a Nesterov momentum term, with a momentum coefficient of 0.9 and an initial learning rate of 0.001 that decays with the number of iterations. The convolution kernel weight update uses a hierarchical adjustment strategy, with the learning rate of the first layer of convolution kernels set to 0.3 times the baseline value to maintain basic feature extraction capability, and the learning rate of the high-level convolution kernels set to 1.2 times the baseline value to adapt to rapid change patterns. The parameter update process implements gradient clipping, scaling the gradient vector with an L2 norm exceeding 1.0 by a certain proportion to prevent model instability caused by parameter mutations.

[0118] The third-party audit verification link is embodied in the case of the government data opening platform. When the platform estimates the value of its traffic flow data assets, the benchmark valuation provided by the auditors includes three dimensions: the median of historical transaction prices, market quotes for similar data, and valuation by professional assessment institutions. The difference analysis module uses a weighted average method when calculating the relative error percentage, with the weight of recent transaction records set at 1.5 times that of historical data. The tolerance threshold is dynamically adjusted according to the type of data assets, set at 20% for innovative data assets with high volatility and 10% for stable basic data assets. The error analysis report generates a multi-dimensional deviation distribution chart to show the prediction deviation pattern under different characteristic combinations.

[0119] The model retraining trigger mechanism is designed with a hierarchical response strategy. When the relative error percentage is between 1 and 1.5 times the threshold, the system only fine-tunes the last two layers of the time series convolutional neural network; when it exceeds the 1.5 times threshold, it starts the complete model retraining process. The retraining instructions are distributed to the computing cluster through the message queue, and the message header contains the task priority and the estimated resource demand. The task scheduler dynamically allocates computing resources according to the current system load, and high-priority tasks can preempt the resource quota of low-priority tasks.

[0120] The implementation of distributed task queues is based on open-source message middleware, and message persistence storage ensures that tasks are not lost due to system failure. The task status monitoring interface displays the load and task progress of each computing node in real time, and supports filtering by data asset type, task type, and time consumption range. The resource allocation algorithm considers data locality optimization, preferentially scheduling tasks to computing nodes that store the required training data. The automatic retry mechanism for task failure adopts an exponential backoff strategy, with an initial retry interval of 5 minutes and a maximum of 3 retries.

[0121] Version control manages the complete tracking of the model iteration process. Each model update generates a unique version identifier, including a timestamp, data fingerprint, and parameter summary information. The model repository stores binary files and metadata for different versions, supports version difference comparison and specific version rollback. Version compatibility checks are automatically performed before model deployment, verifying the matching degree of input and output interfaces with upstream and downstream systems. Model performance degradation detection compares the difference in indicators between new and old versions on the validation set, and triggers an alarm when key indicators drop by more than a preset percentage.

[0122] The real-time monitoring system builds a complete observability system. The log collection agent collects running logs from various processing links, including data reception, feature extraction, model inference, and result output stages. The index collector regularly captures system performance data, such as CPU utilization, memory occupancy, and network throughput. The visualization dashboard integrates multi-source monitoring data and displays task load distribution in different time periods through a heat map. Custom alarm rules support triggering based on complex conditions, such as sending an urgent notification when the prediction deviation exceeds the threshold for three consecutive time windows and the system load is higher than 80%.

[0123] The disaster recovery backup scheme ensures system high availability. Model parameters are regularly snapshot saved to a distributed file system, and snapshot frequency is automatically adjusted according to model update frequency. Standby computing nodes deployed across machine rooms maintain heartbeat detection, and service switching is completed within 30 seconds when the master node fails. The data synchronization mechanism uses an asynchronous replication mode, and important operations are guaranteed to be recoverable through pre-write logs. Disaster recovery drills are performed once every quarter to test the process and time consumption of restoring complete services from backups.

[0124] Security control measures run through the entire processing flow. The data transmission channel uses two-way SSL authentication encryption, and a secure tunnel is established between the terminal and the server. Model parameter files are digitally signed, and integrity and source authenticity are verified when loaded. Access control lists limit the execution permissions of sensitive operations, and audit logs record all management operations. Input data is strictly verified and cleaned to prevent maliciously constructed feature vectors from causing model misjudgment.

[0125] Embodiment 5: The execution process of model retraining starts from the collection and management of new data asset sample sets. The data collection system continuously monitors the updates of various data transaction platforms, enterprise internal data warehouses, and third-party data markets, and obtains the latest data assets through distributed crawler technology. Each sample contains complete operational benefit indicators and market fluctuation indicators, accompanied by data quality reports and source authentication information. After desensitization, the sample data is stored in a versioned storage system, and content addressing is used to identify data blocks, ensuring data tamper resistance. The data version control system records the change history of each sample set, including the number of new samples, data distribution changes, and abnormal sample exclusion records.

[0126] The expansion of the historical training dataset adopts an incremental merging strategy, preserving representative samples of the original dataset as a benchmark reference. The data selection algorithm determines the retention ratio based on feature space coverage and sample weight, avoiding dilution of important historical samples. The dynamic weighting method is used to balance the new samples and historical data, giving recent samples higher weights without excessively weakening historical patterns. Feature consistency checks are performed during data merging to ensure that all samples have the same feature dimensions and value ranges. The data sharding strategy automatically adjusts based on the number of computing nodes, ensuring that each computing unit has approximately equal workloads.

[0127] The re-initialization of the decision tree forest structure uses an improved parallel construction algorithm. The forest size is dynamically adjusted based on the amount of new data, keeping the number of trees proportional to the data complexity. The adaptive mechanism is introduced by setting the feature sampling ratio, which dynamically allocates sampling probabilities based on feature importance analysis results. The maximum depth parameter of the tree is calculated based on data distribution characteristics to avoid overgrowth and model overfitting. The node splitting criterion combines information gain and variance reduction criteria to adapt to different types of target variable predictions. Memory mapping technology is implemented during the construction of decision trees to handle large datasets that exceed the physical memory capacity.

[0128] The feature importance ranking uses a robust calculation method based on permutation importance. The importance score of each feature is determined by the incremental prediction error after multiple random permutations, and the calculation results are subjected to statistical significance tests. The redundant feature recognition algorithm analyzes the mutual information and conditional independence between features to establish a feature correlation map. The feature selection process uses the stepwise regression idea, iteratively removing features with contributions below a threshold. The retained feature set is subjected to orthogonalization processing to reduce the influence of multicollinearity between features. The feature engineering log records the decision basis and transformation history of each feature selection, supporting subsequent audit tracking.

[0129] The resource configuration of the parallel computing cluster considers task characteristics and hardware architecture. Computing nodes are divided into different instance types based on memory capacity and processor core count, and the task scheduler matches the optimal instance based on computing requirements. Data parallel processing evenly distributes training samples to each work node, and model parallel processing splits large decision trees into multiple sub-trees for distributed computation. Communication optimization uses gradient compression techniques to reduce the amount of parameter synchronization data between nodes. The fault-tolerant mechanism saves intermediate states through checkpoints, and resumes computation from the latest checkpoint in the event of node failure. The resource monitoring system tracks the computing load and memory usage of each node in real time, dynamically adjusting task allocation.

[0130] The implementation of the blockchain storage system is based on a consortium chain architecture, and the participating nodes include evaluation agencies, auditors, and data owners. The structural design of the evaluation transaction log contains multiple layers of nested hash values, with the hash values of the previous records embedded in the subsequent records to form an unalterable chain. The input feature hash values are calculated using a Merkle tree structure, supporting efficient existence proof. The model version identifier includes the model architecture hash, training data fingerprint, and hyperparameter summary, fully defining the identity information of the model. The timestamp service uses a multi-party signature mechanism, with multiple authoritative time sources jointly authenticating the authenticity of the time information.

[0131] The generation process of the evaluation storage certificate implements multi-party collaborative verification. After the evaluation node generates the preliminary evaluation result, the verification node performs a quick review through a lightweight model. The consensus mechanism requires more than half of the verification nodes to confirm the evaluation logic compliance before generating the official storage. The storage certificate contains the digital fingerprint of the evaluation result, the list of participating verification nodes, and the verification time window. The encryption signature of the certificate uses a threshold signature scheme, requiring multiple key fragments to cooperate to generate a valid signature.

[0132] The association management of the metadata directory uses distributed identifier technology. Each data asset is assigned a unique global identifier, and parsing the identifier can obtain the complete metadata and related storage links of the asset. The directory service supports advanced queries based on attributes, allowing filtering of target assets by industry classification, data type, or evaluation range. The version control system records the change history of the metadata, preserving snapshots of all historical versions for auditing. The access control policy defines the operation permissions of different roles on the metadata, enabling fine-grained data governance.

[0133] The processing flow of the audit request is designed as a multi-party collaborative mechanism. When receiving an audit request, the system first verifies the access permissions and audit scope authorization of the requester. The process chain retrieval service collects all related logs and intermediate results from the blockchain and distributed storage. The zero-knowledge proof generator creates a cryptographic proof of the compliance of the model behavior, without revealing the original data and model parameters. The audit report contains a complete evidence chain and verifiable reference identifier, supporting independent verification by third-party institutions. The dispute handling process establishes a multi-level mediation mechanism, including technical review, expert evaluation, and arbitration resolution.

[0134] The system operation and management establishes standardized operation procedures. Daily maintenance includes node health checks, storage space monitoring, and network connection testing. Regular maintenance tasks perform database optimization, log archiving, and backup verification. The security update mechanism applies patches to fix known vulnerabilities in a timely manner, and major updates are preceded by impact assessment and compatibility testing. Capacity planning is based on historical growth trends to predict future resource needs, and necessary hardware infrastructure is deployed in advance. The emergency plan document defines the handling steps and recovery time objectives for various types of failure scenarios in detail.

[0135] The document and knowledge management system covers all aspects of technical implementation and business application. The system architecture document describes the interaction interfaces and data flow diagrams of each component. The API reference manual details the service endpoints, parameter specifications, and error codes. The deployment guide includes environment preparation, configuration tuning, and performance testing methods. The operation manual provides operation steps and troubleshooting methods for common tasks. The training materials are designed in a hierarchical structure to meet the learning needs of users with different roles. The case library collects typical application scenarios and solutions to help new users quickly understand the system capabilities.

[0136] Example 6: A medical data service provider holds a set of pathological data for AI diagnosis assistance, which covers pathological section annotation data of different departments, patient clinical related information, etc. Now the data set needs to be valued to support the purchase cooperation negotiation with multiple hospitals. In the valuation process, a data value refinement evaluation link based on the Sharpe value is introduced to realize the complete link from single data point utility analysis to overall data set valuation.

[0137] Carry out the index collection work of the target data asset. In terms of operational benefit indicators, through the calling log of the hospital AI diagnosis system, the use frequency of the pathological data set (the number of AI model calls per day), the data completeness (the proportion of samples without missing pathological annotation fields), the update cycle (the average update interval of new pathological samples), and the data calling cost saving value (the reduced manual annotation cost and model training time cost conversion value after using the data set) are extracted. Market fluctuation indicators are obtained through industry medical data reports, policy databases, and competitor monitoring platforms, including industry demand growth rate (the demand growth rate of pathological data in the medical AI diagnosis field in the past year), competitor substitution rate (the market share proportion of other service providers providing similar pathological data), policy influence coefficient (the quantitative value of the restriction degree of data circulation based on the adjustment of medical data privacy protection policy), and technology iteration risk value (the evaluation value of the difficulty of adapting existing pathological data due to AI diagnosis model algorithm updates).

[0138] In the model selection and construction phase, the effect of data valuation, training cost and deployment convenience are considered, and at the same time, the random forest model in the scikit-learn library is selected as the basic utility evaluation model to adapt to the calculation demand of Sharpe value. The model parameter adjustment is relatively simple, and it is easy to run efficiently in CPU environment, and can quickly output the utility contribution of data samples to the model prediction result. At the same time, cross-validation is carried out with the logistic regression model, and the utility results of the two models are compared to reduce the influence of single model bias on subsequent calculation. In addition, the pre-trained random forest regression model and time series convolutional neural network are retained, the former is used to process the operation benefit index to generate the initial benefit valuation vector, and the latter is used to analyze the market fluctuation index to generate the dynamic discount rate matrix, while the newly added random forest and logistic regression model are specially used for Sharpe value calculation, forming a multi-model collaborative valuation architecture.

[0139] In the single data point utility evaluation link, the Sharpe value in the field of cooperative game is used for calculation. The Sharpe value quantifies the value of a single data point by considering the marginal utility of the data in all possible subsets or arrangements. Its combination-based calculation formula is:

[0140]

[0141] Among them, represents the Sharpe value of the th data sample (such as a single pathological sample in the pathological data set), which quantifies the marginal value of the sample under a specific model and utility standard; is any subset of the data sample set , and does not contain the th sample ( ); refers to the complete sample set of the target data asset, such as the pathological data set containing all pathological section labeling data and clinical correlation information; is the total number of samples in the complete sample set ; is the number of samples in the subset ; , , respectively represent the factorial of the corresponding numerical value, used to calculate the subset weight; is the weight coefficient of the subset , which measures the importance of the subset in all possible subsets; is the utility value of the subset , that is, the performance indicator (such as prediction accuracy) of the model on the standard test set (such as the standard pathological diagnosis test set) after the subset inputs the basic evaluation model (such as the random forest model); is the utility value of the subset Utility value of new subset after adding the th sample; The marginal utility increment brought by adding the th sample to the subset is the core basis for calculating the Shapley value. At the same time, the Shapley value can also be calculated by the permutation-based formula:

[0142]

[0143] wherein, has the same meaning as the combination-based Shapley value formula, i.e., the Shapley value of the th data sample; is the factorial of , representing the total number of permutations of all samples in the complete sample set , which is used to average the marginal utility of all permutations; is any permutation of the complete sample set , such as a random permutation sequence of pathological samples; is the set of all possible permutations of the complete sample set , containing permutations; is a subset composed of the first samples in permutation , and is a subset composed of the first samples in permutation ; and are the utility values of the corresponding subsets, and the calculation logic is consistent with ; is the marginal utility increment of the th sample when it is added as the th element in permutation ; is the sum of the marginal utility increments of the th sample in all permutations; is the average coefficient, which divides the total marginal utility by the total number of permutations to obtain the average marginal utility of the th sample, i.e., the Shapley value. However, both the combination-based and permutation-based calculation methods face significant computational challenges: when the pathological data set contains 500 samples, the combination-based calculation has a computational complexity of , and the permutation-based calculation has a computational complexity of , which is obviously impossible to achieve directly through strict calculation. Therefore, the Monte Carlo approximation method is used to reduce the computational complexity.

[0144] ​In the Monte Carlo approximation calculation process, first set the random sampling number to 10000 times, and each sampling generates a random arrangement of 500 samples of the pathological data set (allowing repeated arrangement, that is, using the sampling with replacement method). For each generated arrangement, start with the first sample to gradually build a sample subset: take the first 1 sample in the arrangement to form a subset , calculate its utility on the basis of the random forest model ; take the first 2 samples to form a subset , calculate the utility , and the difference between the two is the marginal utility of the second sample under the arrangement. In this way, the marginal utility of all samples in the arrangement is calculated. After 10000 samplings, the marginal utility of each sample in all samplings is summarized, and the average value is taken as the approximate Shapley value of the sample. In the calculation process, it is found that when the prediction results of the basic random forest model for some pathological samples fluctuate greatly, the sampling number needs to be increased to 20000 times to reduce the approximation error, but at the same time the calculation time of each sampling needs to be controlled to avoid too long overall evaluation period.

[0145] From the calculation of Shapley value to data valuation, first sum up the approximate Shapley values of all pathological samples to get the total Shapley value of the data set. Since a certain hospital participating in the negotiation has previously purchased a set of similar pathological data set (containing 100 samples), and the total Shapley value and purchase cost of the data set under the same basic random forest model are known, the preliminary valuation range of the new data set can be calculated by the ratio of the total Shapley values of the two. For example, the total Shapley value of the hospital purchased data set is 850, the purchase cost is 120,000 yuan, and the total Shapley value of the data set to be valued is 1700, so the preliminary estimate of the data set is about 240,000 yuan. In addition, by comparing the Shapley values of individual samples, it is found that the Shapley values of 20 samples are close to zero or negative, and further analysis shows that these samples have problems such as label ambiguity and poor adaptability to existing AI diagnosis models, and are redundant samples with low utility, so in the final valuation, this part of the sample is excluded, and only the Shapley value sum of the remaining 480 effective samples is adjusted to make the valuation result more in line with the actual application value.

[0146] The Shapley value of each valid pathological sample is added as a new feature and concatenated with the initial benefit valuation vector (generated by inputting operational benefit indicators into a pre-trained random forest regression model) to form an expanded benefit feature vector. This vector is mapped to the first embedding matrix in a high-dimensional feature space through an embedding layer. Then, it is orthogonally transformed with the second embedding matrix obtained from the dynamic discount rate matrix (generated by inputting market volatility indicators into a pre-trained temporal convolutional neural network, which captures the long-term temporal dependence of indicators such as industry demand growth rate and policy impact coefficient through multi-scale dilated convolutional layers and combines attention mechanisms to weight key features). These two vectors participate in the feature cross-fusion calculation. During the multi-dimensional tensor interaction process, in addition to calculating the Kronecker product of the initial benefit valuation vector and the dynamic discount rate matrix, the embedding matrix corresponding to the Shapley value feature is also included in the calculation, generating an interactive feature cube containing information on single-sample utility, operational benefit, and market volatility. After tensor folding and batch normalization, the fused value feature tensor is output.

[0147] The fusion tensor is then fed into a fully connected estimation network, which employs a deep architecture with residual connections and five hidden layers. A random dropout regularization strategy (dropout probability set to 0.3) is implemented between the hidden layers. A nonlinear feature transformation is performed using an exponential linear unit activation function, the formula of which is:

[0148]

[0149] in, It is the output value of the exponential linear unit (ELU), that is, the activation function's response to the input signal. The result after nonlinear transformation is used for feature transformation of the hidden layer of the fully connected estimation network. It is the input signal of the activation function, that is, the input value of the hidden layer neuron of the fully connected estimation network, which is obtained by adding the product of the previous layer's output and the corresponding weight to the bias term; These are learnable parameters that are dynamically optimized during model training based on the distribution of training data and model error, and are used to control... The degree of saturation in the output of the activation function; It is a natural exponential function, expressed as a natural constant. (Approximately 2.718) is the base. The exponent is used for negative input signals. Perform nonlinear transformation; For negative input signals The initial transformation results ensure As the value approaches negative infinity, the activation output approaches... ;when At that time, the activation function directly outputs To maintain the linear characteristics of the positive input signal; when Output The feature transformation under negative input is made smoother, and the model's ability to capture negative value features is enhanced. The generated results are stored in the blockchain notarization system together with the Sharpe value of each sample, the number of Monte Carlo samples, the parameter identification of the base model version, the pre-trained random forest regression model, and the time series convolutional neural network. A corresponding valuation transaction log is created in the blockchain, generating an unalterable notarization certificate, and is associated with the metadata directory of the pathology dataset. When the hospital requests an audit, the complete valuation process chain including Sharpe value calculation process, index collection source, and model operation link can be output.

[0150] In subsequent valuation dynamic adjustment, when the industry monitoring data stream pipeline captures real-time market fluctuation indicators such as accelerated iteration speed of medical AI diagnosis model (increased technical iteration risk value), introduction of new medical data privacy policy (change in policy impact coefficient), etc., the incremental learning engine (including online sequential minimal optimization algorithm) will dynamically update the convolution kernel weight parameters of the time series convolutional neural network, generating a new dynamic discount rate matrix; at the same time, the Sharpe value calculation module will also adjust the utility calculation standard of the base random forest model according to the updated market environment (such as calculating the Sharpe value of the updated AI diagnosis model test set ), and re-execute Monte Carlo sampling calculation to ensure that single-sample utility evaluation keeps pace with market changes. Finally, through the re-operation of feature cross-fusion and full-connection network, the adjusted valuation results are output, continuously adapting to the dynamic demand in procurement negotiations.

[0151] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0152] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A data asset valuation method based on machine learning fusion of benefits and market factors, characterized by, The method comprises the following steps: Collecting operation benefit indicators and market fluctuation indicators of the target data asset, the operation benefit indicators including data usage frequency, data completeness, data update period, and data call cost saving value, and the market fluctuation indicators including industry demand growth rate, competitive product substitution rate, policy influence coefficient, and technology iteration risk value; Inputting the operation benefit indicators into a pre-trained random forest regression model to perform benefit benchmark valuation and obtain an initial benefit valuation vector; inputting the market fluctuation indicators into a pre-trained time series convolutional neural network to perform market discount rate prediction and generate a dynamic discount rate matrix; Performing multi-dimensional tensor interaction calculation on the initial benefit valuation vector and the dynamic discount rate matrix through a feature cross fusion module to output a fusion value feature tensor; Performing nonlinear transformation processing on the fusion value feature tensor by using a fully connected valuation network to generate a final data asset valuation result and store the result in a blockchain storage system; The operation process of the feature cross fusion module comprises the following steps: Mapping the initial benefit valuation vector to a first embedding matrix in a high-dimensional feature space through an embedding layer; Converting the dynamic discount rate matrix into a second embedding matrix through orthogonal transformation; Calculating the Kronecker product of the first embedding matrix and the second embedding matrix to generate an interaction feature cube; Performing tensor folding operation on the interaction feature cube to output a reduced fusion feature matrix; 2. The method for data asset valuation based on machine learning fusion of benefits and market factors according to claim 1, characterized in that, Processing the fusion feature matrix through a batch normalization layer to generate the fusion value feature tensor and pass the tensor to the fully connected valuation network. The construction process of the pre-trained random forest regression model comprises the following steps: Obtaining operation benefit indicator samples and corresponding valuation labels of historical data assets, the operation benefit indicator samples including data call delay time, data error rate, data correlation dimension number, and data service availability; Constructing a decision tree forest structure and setting a tree depth constraint condition, the tree depth constraint condition including maximum split number limit, node purity threshold, and feature sampling ratio; Performing recursive feature splitting operation through a Gini impurity minimization principle to generate a plurality of decision tree models; 3. The method for data asset valuation based on machine learning fusion of benefits and market factors according to claim 2, characterized in that, Calculating weight coefficients of the decision tree models by using an out-of-bag error estimation method; Performing weighted aggregation on output results of the plurality of decision tree models based on the weight coefficients to complete the construction of the pre-trained random forest regression model. The training process of the pre-trained time series convolutional neural network comprises the following steps: Collecting historical market fluctuation indicator sequences and corresponding discount rate labels, the historical market fluctuation indicator sequences including industry concentration change curve, price elasticity coefficient sequence, and capital inflow rate time sequence; Constructing a multi-scale dilated convolution layer structure, the multi-scale dilated convolution layer structure comprising convolution kernel groups with different dilation rates; 4. The method for data asset valuation based on machine learning fusion of benefits and market factors according to claim 1, characterized in that, Capturing long-term time series dependence features through a gated recurrent unit to output a time series feature encoding vector; Performing feature importance weighting on the time series feature encoding vector by using an attention mechanism; Optimizing network parameters by using a mean square error loss function until convergence to generate the pre-trained time series convolutional neural network. The running mechanism of the fully connected valuation network comprises the following steps: Receiving the fusion value feature tensor from the feature cross fusion module; A deep neural network architecture comprising residual connections is constructed, the deep neural network architecture having a plurality of hidden layers; A random dropout regularization strategy is implemented between the hidden layers; A nonlinear feature transformation is performed by an exponential linear unit activation function; An output layer employs a linear activation function to generate the final data asset valuation result.

5. The method for data asset valuation based on machine learning fusion of benefits and market factors according to claim 4, characterized in that, Real-time updating of market volatility indicators is also included: an industry monitoring data stream pipeline is established to continuously collect real-time market volatility indicators; The real-time market volatility indicators are input into an incremental learning engine, which includes an online sequential minimal optimization algorithm; The convolution kernel weight parameters of the pre-trained time series convolutional neural network are dynamically updated; The updated dynamic discount rate matrix is transmitted to the feature cross-fusion module.

6. The method for data asset valuation based on machine learning fusion of benefits and market factors according to claim 5, characterized in that, The workflow of the incremental learning engine includes: setting a sliding time window to capture real-time market volatility indicator streams; Calculating the cosine similarity between newly arrived market volatility indicators and historical features; Triggering model updates when the cosine similarity is below the adaptive threshold; Adjusting the convolution kernel weight parameters using incremental gradient descent; Synchronizing the updated network parameters to the pre-trained time series convolutional neural network.

7. The method of claim 6, wherein the method further comprises: Validation of the valuation result is also included: benchmark valuation data provided by a third-party audit platform is obtained; The final data asset valuation result and the benchmark valuation data are input into a difference analysis module; The relative error percentage between the two is calculated; Triggering model retraining when the relative error percentage exceeds the tolerance threshold; Sending retraining instructions to the pre-trained random forest regression model and the pre-trained time series convolutional neural network.

8. The method of claim 7, wherein the method further comprises: The execution process of the model retraining includes: collecting a new data asset sample set, which contains new operating efficiency indicators and new market volatility indicators; Extending the historical training data set to the new data asset sample set; Reinitializing the decision tree forest structure; Performing feature importance sorting and removing redundant features; Accelerating the model training process through parallel computing clusters.

9. The method of claim 8, wherein the method further comprises: The valuation traceability mechanism is also included: an evaluation transaction log is created in a blockchain notarization system, which contains input feature hash values, model version identification, and timestamp information; Generating tamper-proof valuation notarization credentials; Associating the valuation notarization credentials to the metadata directory of the target data asset; Outputting the complete valuation process chain in response to audit requests.

Images