A method for constructing a distributed machine learning optimization model based on financial big data under a resource-constrained network
By partitioning and preprocessing financial big data under resource-constrained networks, a distributed machine learning model is constructed and optimized, solving the problems of high communication overhead, low training efficiency, and poor model accuracy, and achieving efficient training and stability in resource-constrained environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2025-05-15
- Publication Date
- 2026-05-29
Smart Images
Figure CN120597994B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of financial big data processing and distributed machine learning technology, and in particular to a method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks. Background Technology
[0002] In the financial industry, with the rapid development of information technology, the scale of financial big data is growing exponentially. Distributed machine learning, as an effective means of processing large-scale financial data, can train models collaboratively through multiple computing nodes without centralizing data, ensuring data privacy while making full use of distributed computing resources. However, in practical applications, many financial scenarios rely on resource-constrained network environments, such as networks composed of mobile terminal devices and edge computing node networks. These networks typically have characteristics such as limited bandwidth, uneven computing power, high communication latency, and limited energy supply.
[0003] Existing distributed machine learning methods suffer from numerous problems in resource-constrained networks. On one hand, traditional distributed training algorithms, such as federated learning, often employ simple averaging strategies when aggregating model parameters, failing to adequately consider the resource constraints of each computing node. This leads to excessive network communication overhead and low training efficiency. On the other hand, due to resource constraints, computing nodes may be unable to fully store and process large-scale financial datasets. Traditional methods do not optimize for resource constraints during data preprocessing and model training, easily resulting in the loss of data features or model training non-convergence. Furthermore, in resource-constrained networks, computing nodes are less stable and prone to failure or communication interruptions. Traditional methods lack effective fault-tolerance mechanisms and adaptive adjustment strategies, impacting the reliability and accuracy of model construction.
[0004] In summary, the technical problem actually solved by this invention is how to efficiently construct a distributed machine learning model for financial big data in a resource-constrained network environment, and solve the problems of high communication overhead, low training efficiency, poor model accuracy and insufficient fault tolerance in existing methods. Summary of the Invention
[0005] To overcome the shortcomings of existing methods, such as high communication overhead, low training efficiency, poor model accuracy, and insufficient fault tolerance, this invention aims to provide a method for constructing an optimized distributed machine learning model for financial big data under resource-constrained networks. By optimizing the process of constructing a distributed machine learning model for financial big data under resource-constrained networks, this invention effectively solves the problems of high communication overhead, low training efficiency, poor model accuracy, and insufficient fault tolerance in existing methods, thereby improving the training efficiency and accuracy of the model.
[0006] This invention discloses a method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks, comprising the following steps:
[0007] Based on the central server, financial big data is divided into several data subsets, and these subsets are preprocessed to obtain a set of feature vectors.
[0008] An initial distributed machine learning model is built on the central server. This initial distributed machine learning model includes a multi-layered neural network structure, and its loss function is:
[0009]
[0010] in, Represented as the sum of the data samples, The input feature vector, For the labels corresponding to the data samples, This represents the predicted output of the initial distributed machine learning model. The loss function;
[0011] The central server performs real-time assessments of the resource status of each computing node to obtain resource parameter information for each computing node, including computing capacity indicators, communication bandwidth indicators, and remaining energy indicators.
[0012] The resource weight of each computing node is calculated based on the resource parameter information. Specifically, the resource weight of each computing node is calculated based on the computing power index, communication bandwidth index, and remaining energy index of the computing node.
[0013] Based on the resource weight of each computing node Each computing node is assigned a set of feature vectors, and then trains its local model on a subset of the assigned data to update the parameters of that local model. And during training, the stochastic gradient descent algorithm is used:
[0014]
[0015] in, For learning rate, This represents the number of training iterations.
[0016] The central server collects parameters from the local model of each computing node. And based on the resource weight of each computing node Perform weighted aggregation to obtain global model parameters:
[0017]
[0018] in, To calculate the number of nodes;
[0019] The global model is evaluated based on the test dataset to obtain evaluation parameter values. The evaluation parameter values are then compared with preset parameter thresholds. If the evaluation parameter value is less than the preset parameter threshold, the above steps are repeated. When the evaluation parameter value is greater than or equal to the preset parameter threshold, the distributed machine learning optimization model is obtained.
[0020] Preferably, the preprocessing includes feature filtering and normalization of each data subset to obtain a set of feature vectors;
[0021] Feature selection is based on the mutual information method to calculate the mutual information value between each feature and label of each big financial data. The central server selects features with mutual information values greater than a preset mutual information threshold to divide them into data subsets.
[0022] Preferably, computing power indicators It is obtained by calculating the CPU clock frequency, memory capacity, and number of processor cores of the computing node;
[0023] Communication bandwidth indicators It is obtained by measuring the communication rate between the computing nodes and the central server in real time;
[0024] Remaining energy index Data is obtained by monitoring the battery level of the computing nodes.
[0025] Preferably, the global model parameters are constrained based on a regularization term, and the regularization loss function is:
[0026]
[0027] in, For regularization parameters, This refers to the dimension of the model parameters.
[0028] Preferably, the test dataset is selected from financial big data using stratified sampling.
[0029] Preferably, based on the resource weight of each computing node Resource weights for each computing node when allocating feature vector sets. It is proportional to the allocated data subset.
[0030] Preferably, the learning rate in the stochastic gradient descent algorithm is calculated as follows: An adaptive adjustment strategy is adopted to dynamically adjust the loss function based on the rate of change during training.
[0031] Preferably, the learning rate Adaptive adjustment of resource weights of computing nodes The relevant, adaptive adjustment calculation formula is:
[0032]
[0033] in, To adjust the coefficient, Resource weights for all compute nodes The average value, To determine the number of training iterations, Indicates the first The learning rate at the next iteration Indicates the first The learning rate is updated in the next iteration.
[0034] Preferably, the central server uploads parameters of the local model to the computing nodes. When performing weighted aggregation, a parameter difference coefficient is introduced. Parameter difference coefficient Based on the parameters of the local model Compared with the previous round of global model parameters European distance The calculated parameter difference coefficient The specific calculation formula is as follows:
[0035]
[0036] The aggregation formula for global model parameters is:
[0037]
[0038] in, This is used to calculate the resource weights of the nodes.
[0039] By adopting the above technical solution, compared with the existing technology, the present invention effectively solves the problems of large communication overhead, low training efficiency, poor model accuracy and insufficient fault tolerance in the existing methods by optimizing the construction process of distributed machine learning models for financial big data under resource-constrained networks, thereby improving the training efficiency and accuracy of the model.
[0040] This invention uses a central server to evaluate the status of computing nodes in real time and based on resource weights. By allocating feature vector sets, the amount of data transmitted and the number of communications on low-resource computing nodes are reduced, thereby lowering network communication overhead and making full use of the resource weights of computing nodes. It improved training efficiency.
[0041] This invention divides the data into subsets by selecting features with mutual information values greater than the mutual information threshold, thereby removing redundant and irrelevant features, reducing data dimensionality while retaining key information; and introduces a regularized loss function during training to prevent overfitting and improve the model's generalization ability. Attached Figure Description
[0042] Figure 1 This is a schematic diagram illustrating the steps of a method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks, as described in this invention. Detailed Implementation
[0043] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.
[0044] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.
[0045] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.
[0046] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0047] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0048] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.
[0049] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.
[0050] This embodiment discloses a method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks, including the following steps: dividing the financial big data into several data subsets based on a central server, and preprocessing the several data subsets to obtain a set of feature vectors; constructing an initial distributed machine learning model on the central server, the initial distributed machine learning model including a multi-layered neural network structure, and the loss function of the initial distributed machine learning model is: ,in, Represented as the sum of the data samples, The input feature vector, For the labels corresponding to the data samples, This represents the predicted output of the initial distributed machine learning model. The loss function is defined as follows: the central server performs real-time evaluation of the resource status of each computing node to obtain resource parameter information for each computing node. Each computing node is assigned a set of feature vectors, and then trains its local model on a subset of the assigned data to update the parameters of that local model. And during training, the stochastic gradient descent algorithm is used: ,in, For learning rate, The central server collects the parameters of the local model from each computing node to determine the number of training iterations. And based on the resource weight of each computing node Perform weighted aggregation to obtain global model parameters: ,in, The number of computing nodes is used; the performance of the global model is evaluated based on the test dataset to obtain evaluation parameter values, and the evaluation parameter values are compared with preset parameter thresholds. If the evaluation parameter value is less than the preset parameter threshold, the above steps are repeated. When the evaluation parameter value is greater than or equal to the preset parameter threshold, the distributed machine learning optimization model is obtained.
[0051] In this embodiment, refer to Figure 1 As shown, a detailed description of a method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks will be provided, specifically including the following steps:
[0052] Step S100: Financial big data typically contains a large amount of raw data, which may contain redundant, irrelevant, or multidimensional features. The central server divides the financial big data into several subsets and preprocesses them. Preprocessing includes feature selection and normalization. Feature selection removes useless information, reduces data dimensionality, and improves training efficiency. Normalization maps feature values to a uniform range, eliminating the influence of different feature dimensions, allowing the distributed machine learning optimization model to learn more stably. The preprocessed result is a set of feature vectors.
[0053] Step S200: Build an initial distributed machine learning model on the central server. The initial distributed machine learning model includes a multi-layered neural network structure, including an input layer, a hidden layer, and an output layer.
[0054] Input Layer: The number of neurons in the input layer is consistent with the dimension of the preprocessed feature vectors. This is because the preprocessed feature vector set, where each feature vector represents a set of feature information from financial big data, requires the input layer to receive and transmit this feature information completely.
[0055] Hidden layers: Hidden layers play a crucial role in feature extraction and transformation within the model. The number of hidden layers and neurons can be flexibly adjusted based on the complexity of the actual financial task, the dimensionality of the data features, and the computational power of the computing nodes.
[0056] Output layer: The number of neurons in the output layer is determined based on the specific financial task.
[0057] The initial distributed machine learning model includes weights and biases between layers. These parameters are continuously adjusted and optimized during training to make the predicted output closer to the true label. During the forward propagation of the neural network, the input data is weighted and summed with the weights, then the bias is added, and the data is transformed through an activation function before being passed to the next layer.
[0058] The loss function for this initial distributed machine learning model is: In this formula, Represented as the sum of the data samples, The input feature vector, The labels corresponding to the data samples, This is the predicted output of the initial distributed machine learning model. This is the loss function.
[0059] Step S300: Since the resource status of different computing nodes may vary greatly in a resource-constrained network, the central server performs real-time evaluation of the resource status of each computing node to obtain the resource parameter information of each computing node, and then calculates the resource weight of each computing node based on the resource parameter information of each computing node. .
[0060] The central server assesses the resource status of each computing node, obtaining resource parameter information including computing power, communication bandwidth, and remaining energy, and determines the resource weight of each computing node accordingly. The resource weight is used to characterize the overall resource status of each computing node and serves as the basis for feature vector set allocation and weighted aggregation of local model parameters.
[0061] Step S400: Then, based on the resource weight of each computing node... Each computing node is assigned a set of feature vectors and trains its local model on its assigned subset of data independently. During training, a stochastic gradient descent algorithm is used. The stochastic gradient descent algorithm used in this step is... ,in, For learning rate, This represents the number of training iterations.
[0062] It should also be noted that this is based on the resource weight of each computing node. Resource weights when allocating feature vector sets High-performance computing nodes will be allocated to more subsets of data to fully utilize their superior computing and communication capabilities.
[0063] Step S500: In this step, the central server will collect the parameters of the local model from each computing node. The global model parameters are obtained by weighting and aggregating the data based on the resource weights of the computing nodes. .in, The number of computing nodes is denoted by . This weighted aggregation strategy fully considers the resource differences among computing nodes; resource-rich computing nodes contribute more to the global model, improving aggregation efficiency and model accuracy. Simultaneously, a regularized loss function is introduced to constrain the parameters of the global model. The regularized loss function is . This prevents the model from overfitting, thereby improving the model's generalization ability.
[0064] Step S600: Evaluate the performance of the global model and obtain evaluation parameters using the test dataset. The test dataset is selected from the original financial big data set using stratified sampling to ensure that the distribution of the test data is consistent with the training data, avoiding evaluation bias caused by differences in data distribution. The obtained evaluation parameter values are then compared with preset parameter thresholds. If the evaluation parameter value is less than the preset threshold, the above steps are repeated to continue training until the evaluation parameter value is greater than or equal to the preset parameter threshold, at which point a distributed machine learning optimization model is obtained.
[0065] Furthermore, the preprocessing includes feature filtering and normalization for each data subset to obtain a set of feature vectors; feature filtering is based on the mutual information method to calculate the mutual information value between each feature and label of each financial big data, and the central server selects features with mutual information values greater than a preset mutual information threshold to divide them into data subsets.
[0066] In this embodiment, the preprocessing in step S100 will be described in detail. Feature selection uses the mutual information method to calculate the mutual information value between each feature and label of each big financial data and compares it with a preset mutual information threshold. The central server selects features with mutual information values greater than the preset mutual information value to divide the data subset.
[0067] Furthermore, computing power indicators The communication bandwidth index is calculated based on the CPU clock frequency, memory capacity, and number of processor cores of the computing node. Remaining energy indicators are obtained by measuring the communication rate between computing nodes and the central server in real time. Data is obtained by monitoring the battery level of the computing nodes.
[0068] It should be noted that this part has already been described in detail above, and will not be described in detail again here. Please refer to the above text.
[0069] It should be noted that when the remaining energy index of the computing node... When the energy level falls below a preset threshold (e.g., 0.2), the number of training iterations is automatically reduced or the computational precision is lowered. In a financial big data collection network using a large number of smart meters as computing nodes, the power of the smart meters is limited. When the remaining power is insufficient, the planned 100 training iterations can be reduced to 20, or the data computation precision can be reduced from 32-bit floating-point numbers to 16-bit floating-point numbers. This ensures a certain training effect while extending the working time of the computing nodes, ensuring that the computing nodes continuously participate in training and maintaining the stability of distributed training.
[0070] Furthermore, the global model parameters are constrained based on the regularization term, and the regularization loss function is: ,in, For regularization parameters, This refers to the dimension of the model parameters.
[0071] In this embodiment, the regularization loss function is described in detail. By constraining the global model through the regularization loss function, the size of the model parameters is limited, the model is prevented from becoming too complex, and thus the generalization ability of the model is improved.
[0072] Furthermore, the test dataset was selected from financial big data using stratified sampling.
[0073] In this embodiment, the test dataset is selected from financial big data using stratified sampling to ensure consistency between the distribution of test data and training data. For example, in some embodiments, the test dataset is composed of a 7:3 ratio of data from customers with good credit and data from customers with poor credit in the financial big data. This avoids evaluation bias caused by differences in data distribution and allows performance metrics such as accuracy, recall, and F1 score to more accurately reflect the model's performance in real-world applications.
[0074] Furthermore, based on the resource weight of each computing node Resource weights for each computing node when allocating feature vector sets. It is proportional to the allocated data subset.
[0075] In this embodiment, when the central server allocates the feature vector set, the resource weight of each computing node is... The resource weight is proportional to the allocated data subset. Higher-ranking computing nodes are allocated larger subsets of data, resulting in higher resource weight. Lower-level computing nodes are allocated fewer data subsets to fully utilize the resource advantages of each computing node, thereby improving overall training efficiency.
[0076] Furthermore, the learning rate in the stochastic gradient descent algorithm is calculated... An adaptive adjustment strategy is adopted to dynamically adjust the loss function based on the rate of change during training.
[0077] In this embodiment, the stochastic gradient descent algorithm will be described in detail, including the learning rate in the stochastic gradient descent algorithm calculation formula. An adaptive adjustment strategy is adopted. This means that in the early stages of training, the loss function decreases rapidly while maintaining the learning rate. Keep the learning rate unchanged; increase the learning rate when the loss function decreases by less than 0.01 over 10 consecutive iterations. The convergence speed; if the loss function fluctuates, reduce the learning rate. Stabilize the training process. This adaptive adjustment allows the model to maintain a good learning state at different training stages, thereby improving training efficiency and model accuracy.
[0078] Furthermore, learning rate Adaptive adjustment of resource weights of computing nodes The relevant, adaptive adjustment calculation formula is: ,in, To adjust the coefficient, Resource weights for all compute nodes The average value, To the number of training iterations, Indicates the first Learning rate at the next iteration Indicates the first The learning rate is updated in the next iteration.
[0079] The purpose of this embodiment is to make the learning rate The adjustment is more closely aligned with the actual resource situation of computing nodes. In traditional training, the learning rate... Adjustments are typically based solely on the model's training state, but this embodiment introduces resource weights for computation nodes. To correlate it with the learning rate This is related to the adaptive adjustment of the computing node's resource weight. Higher than the resource weight of all computing nodes average When this occurs, it means that the computing node has abundant resources, which can increase the learning rate. Accelerate training speed; conversely, if resource weighting Below the average of all computing nodes When that happens, reduce the learning rate. This ensures the stability of training and avoids deviations in training due to insufficient resources.
[0080] It should be noted that the calculation formula for the above adaptive adjustment is: middle, Indicates the first The learning rate at the next iteration It is in the first The learning rate is updated in the next iteration. The adjustment factor is a hyperparameter that can be set according to the actual situation, used to control the learning rate. The extent of the adjustment The larger the value, the higher the learning rate. Subject to resource weight The greater the magnitude of the change in impact, Used to measure the resource weight of a computing node Resource weights of all computing nodes The relative difference of the average value, when When the result of the formula is positive, the learning rate is... It will be increased on the original basis; when When the result is negative, the learning rate is... It will decrease.
[0081] In some embodiments, a resource-constrained network consisting of 5 computing nodes is assumed to be used to train a model predicting customer credit risk. In a given training iteration, the resource weights of each computing node are... They are respectively , , , , It can calculate the resource weights of all computing nodes. average Set adjustment coefficient .
[0082] For compute node 1, its resource weight , ,but Learning rate It remains unchanged.
[0083] For compute node 2, its resource weight , ,but Learning rate The weight has been reduced to 0.975 times its original value due to the resource weight of this computing node. If the learning rate is below average, reduce the learning rate. This ensures that its computing nodes maintain a certain level of stability.
[0084] For compute node 3, its resource weight , ,but Learning rate The number of nodes was increased to 1.05 times the original size, and the training progress was accelerated by utilizing computing nodes.
[0085] The same applies to compute node 4 and compute node 5, and will not be described in detail here.
[0086] Furthermore, the central server uploads parameters of the local model to the computing nodes. When performing weighted aggregation, a parameter difference coefficient is introduced. Parameter difference coefficient Based on the parameters of the local model Compared with the previous round of global model parameters European distance The calculated parameter difference coefficient The specific calculation formula is as follows: The aggregation formula for the global model parameters is: ,in, This is used to calculate the resource weights of the nodes.
[0087] In this embodiment, due to differences in data distribution, computing power, and other factors, the parameters of the local model vary among different computing nodes during the model training process. The update scenarios vary. Some computing nodes may have significantly different global model parameters due to data anomalies or computational biases. Directly aggregating these parameters using traditional methods would interfere with the stabilization and convergence speed of the global model. This embodiment introduces a parameter difference coefficient. The aim is to adjust these parameters with significant differences by weighting them, so that the aggregated results of the global model are more robust and reliable.
[0088] Parameter difference coefficient Its core function is to quantify the parameters of the local model. Compared with the previous round of global model parameters The influence of the degree of difference on the polymerization process. This is determined by Euclidean distance. To measure the differences between parameters The larger the value, the higher the parameters of the local model of the computing node. With global model parameters The greater the deviation, the more likely there are data or calculation anomalies. Based on this, the formula is used... Calculate the coefficient of difference. The larger, The smaller the value, the lower the weight of the computation node parameter in the aggregation, and vice versa. The smaller, The closer it is to 1, the higher the proportion of its parameter in the aggregation.
[0089] Aggregate formula , Introduction Parameters of the local model for each compute node Weighting is applied. (Molecular weight) Parameters of the local model for all compute nodes Perform a weighted summation, denominator It plays a normalization role and ensures The parameters are set reasonably so that the aggregation results are more consistent with the actual situation and avoid interference from abnormal parameters.
[0090] For example, in some specific embodiments, in a certain iteration, the resource weight of node 1 is calculated. In its local model parameters Compared with the previous round of global model parameters European distance Then the parameter difference coefficient ; Calculate the resource weight of node 2 Then the parameter difference coefficient , ; Calculate the resource weight of node 3 , , If we do not consider Aggregation is performed solely based on resource weights, and the computation nodes evaluate the global model parameters. The impact is significant, but in the introduction Then, the weighted parameters for node 1 are calculated as follows: The weighted parameters for computing node 2 are: The weighted parameters for computing node 3 are: It is evident that due to the significant parameter differences, the actual contribution of computing node 3 to aggregation is greatly reduced, resulting in a decrease in the global model parameters. The updates are more stable.
[0091] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.
Claims
1. A method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks, characterized in that, Including the following steps: The financial big data is divided into several data subsets based on the central server, and the data subsets are preprocessed to obtain a set of feature vectors. An initial distributed machine learning model is constructed on the central server. This initial distributed machine learning model includes a multi-layered neural network structure, and its loss function is: in, Represented as the sum of the data samples, The input feature vector, The labels corresponding to the data samples, This is the predicted output of the initial distributed machine learning model. The loss function; The central server performs real-time assessment of the resource status of each computing node to obtain resource parameter information for each computing node. The resource parameter information includes the computing power index, communication bandwidth index, and remaining energy index of the computing node. The resource weight of each computing node is calculated based on the resource parameter information. Specifically, the resource weight of each computing node is calculated based on the computing power index, communication bandwidth index, and remaining energy index of the computing node. Based on the resource weight of each computing node Each computing node assigns the set of feature vectors and trains the assigned subset of data using a local model to update the parameters of the local model. And during training, the stochastic gradient descent algorithm is used: in, For learning rate, This represents the number of training iterations. The learning rate in the formula for the stochastic gradient descent algorithm An adaptive adjustment strategy is adopted to dynamically adjust the loss function based on the rate of change during training. The learning rate Adaptive adjustment of the resource weights of the computing nodes The relevant, adaptive adjustment calculation formula is: in, To adjust the coefficient, Resource weights for all the aforementioned computing nodes The average value, To determine the number of training iterations, Indicates the first The learning rate at the next iteration Indicates the first The learning rate updated in the next iteration; The central server collects parameters from the local model of each computing node. And according to the resource weight of each of the computing nodes Perform weighted aggregation to obtain global model parameters: in, The number of computing nodes; The global model is evaluated based on the test dataset to obtain evaluation parameter values. The evaluation parameter values are then compared with preset parameter thresholds. If the evaluation parameter values are less than the preset parameter thresholds, the above steps are repeated. When the evaluation parameter values are greater than or equal to the preset parameter thresholds, the distributed machine learning optimization model is obtained. The central server uploads the parameters of the local model to the computing node. When performing weighted aggregation, a parameter difference coefficient is introduced. The parameter difference coefficient Based on the parameters of the local model Compared with the global model parameters described in the previous round European distance The calculated parameter difference coefficient The specific calculation formula is as follows: The aggregation formula for the global model parameters is: Among them, the The resource weight of the computing node.
2. The method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks according to claim 1, characterized in that, The preprocessing includes feature filtering and normalization for each subset of data to obtain a set of feature vectors; The feature selection is based on the mutual information method to calculate the mutual information value between each feature and label of each financial big data. The central server selects features with mutual information values greater than a preset mutual information threshold to divide them into the data subset.
3. The method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks according to claim 1, characterized in that, The computing power index It is calculated using the CPU clock frequency, memory capacity, and number of processor cores of the computing node; The communication bandwidth indicator The communication rate between the computing node and the central server is obtained by measuring it in real time. The remaining energy index The battery level of the computing node is monitored.
4. The method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks according to claim 1, characterized in that, The global model parameters are based on the regularization term. The constraints are applied, and the regularization loss function is: in, For regularization parameters, This refers to the dimension of the model parameters.
5. The method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks according to claim 1, characterized in that, The test dataset was selected from the financial big data using stratified sampling.
6. The method for constructing a distributed machine learning optimization model for financial big data under resource-constrained networks according to claim 1, characterized in that, Based on the resource weight of each computing node When allocating the feature vector set, the resource weight of each computing node It is proportional to the data subset that is assigned.