Supply chain data collaboration platform and method based on federated learning

Abstract

The application provides a supply chain data collaboration platform and method based on federated learning, belonging to the technical field of data collaboration, which distributes an initial global model to each supply chain participant, trains a local teacher model according to local private data, and then generates soft labels. After preprocessing the soft labels, the similarity between the soft labels of each two supply chain participants is calculated to obtain a soft label difference matrix. Then, all participants are divided into several knowledge clusters according to the soft label difference matrix, and the soft labels in each knowledge cluster are weighted and fused to obtain cluster-level fused soft labels. All cluster-level fused soft labels and each sample in a public data set form a multi-target training set to train a student model as a new global model. The above steps are repeated until the new global model converges, which is used for supply chain business prediction of the next round of each supply chain participant. The scheme of the application can realize supply chain data collaboration of global model iterative optimization.

Description

Technical Field

[0001] This application relates to the field of data collaboration technology, and in particular to a supply chain data collaboration platform and method based on federated learning. Background Technology

[0002] In supply chain scenarios, various participants, such as manufacturers, distributors, and retailers, possess a large amount of local private data, such as order records, inventory levels, and logistics timeliness. This data cannot be directly shared due to the involvement of trade secrets, but joint modeling is of great value in improving the accuracy of supply chain demand forecasting, inventory optimization, and logistics analysis.

[0003] Existing federated learning typically employs federated averaging algorithms, where a central server aggregates parameter updates from each participant's local model to generate a global model. However, existing federated learning methods face the following technical challenges in practical applications: First, the data distribution among participants exhibits significant heterogeneity. For example, retailers in different regions may have data distribution shifts due to differences in consumption habits, and directly aggregating parameter updates can lead to performance degradation of the global model on some participants, i.e., the data heterogeneity problem. Second, simply transmitting model parameters still carries the risk of privacy leakage; attackers can reconstruct parts of the original data through gradient backpropagation or model inversion attacks. Third, existing solutions lack effective measurement and utilization of the similarity in the data distribution among participants, making it difficult to differentiate heterogeneous data. Therefore, how to achieve supply chain data collaboration for iterative optimization of the global model while protecting data privacy has become a challenge for the industry. Summary of the Invention

[0004] Based on this, this application provides a federated learning-based supply chain data collaboration platform and method for achieving global model iterative optimization of supply chain data collaboration.

[0005] Firstly, this application provides a supply chain data collaboration method based on federated learning, comprising the following steps:

[0006] The central server distributes the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data. The local teacher model is then used to predict unlabeled public datasets related to the supply chain and generate soft labels that represent knowledge of the local data distribution.

[0007] Each supply chain participant adds differential privacy noise to the generated soft tags before uploading them to the central server;

[0008] The central server calculates the similarity between soft tags of every two supply chain participants based on a preset distribution similarity metric, and obtains a soft tag difference matrix. Then, based on the soft tag difference matrix, all supply chain participants are divided into several knowledge clusters, wherein the supply chain participants in each knowledge cluster have similar data distributions.

[0009] Based on the amount of data from each participant, the soft labels within each knowledge cluster are weighted and fused to obtain the cluster-level fused soft labels for each knowledge cluster. All cluster-level fused soft labels are combined with each sample in the public dataset to form a multi-objective training set to train the student model as a new global model.

[0010] The central server distributes the new global model to each participant and repeats the above steps until the new global model converges, which can then be used by each supply chain participant for the next round of supply chain business forecasting.

[0011] In some embodiments, each supply chain participant trains a local teacher model based on the initial global model and local private data, specifically including:

[0012] After receiving the initial global model distributed by the central server, each supply chain participant obtains the corresponding local private data;

[0013] Each supply chain participant preprocesses its corresponding local private data and divides the preprocessed local private data into local training datasets and local validation datasets.

[0014] Each supply chain participant performs supervised iterative training on the initial global model based on the local training dataset. After each round of training, the performance of the initial global model is evaluated using the local validation dataset. Training stops when the loss function value of the initial global model on the local validation dataset no longer decreases. The model parameters obtained at this point are used as the final parameters of the local teacher model, thus obtaining the local teacher model.

[0015] In some embodiments, generating soft labels representing knowledge of local data distribution by predicting unlabeled public datasets related to the supply chain using the local teacher model specifically includes:

[0016] Each supply chain participant obtains an unlabeled public dataset related to the supply chain, which contains feature vectors of multiple unlabeled business samples;

[0017] Each supply chain participant inputs the feature vector of each sample in the corresponding unlabeled public dataset into the corresponding local teacher model, performs forward inference calculation, and obtains the probability output vector of each sample corresponding to each prediction category.

[0018] Each supply chain participant arranges the probability output vectors of all samples according to the original order of the samples in the public dataset, forming a soft label set that corresponds one-to-one with the samples in the public dataset. The soft label set is the soft label that represents the knowledge of local data distribution.

[0019] In some embodiments, each supply chain participant adds differential privacy noise to the generated soft tags before uploading them to the central server, specifically including:

[0020] Each supply chain participant adds differential privacy noise to the generated soft tag to obtain a noisy soft tag;

[0021] The noise-added soft tags are uploaded to the central server.

[0022] In some embodiments, the similarity between soft tags of every two supply chain participants is calculated based on a preset distribution similarity metric to obtain a soft tag difference matrix, specifically including:

[0023] The similarity value between the preprocessed soft tag features of each two supply chain participants is calculated based on a preset distribution similarity metric.

[0024] Combine all similarity values ​​into a soft-label difference matrix.

[0025] In some embodiments, classifying all supply chain participants into several knowledge clusters based on the soft-label difference matrix specifically includes:

[0026] A similarity graph is constructed based on the soft-label difference matrix. The nodes in the similarity graph are supply chain participants, and the weights of the edges are determined by the similarity values ​​in the soft-label difference matrix.

[0027] Calculate the Laplacian matrix of the similarity graph and perform eigenvalue decomposition on the Laplacian matrix to obtain eigenvectors;

[0028] Clustering is performed based on the feature vectors to obtain the cluster label of each supply chain participant, thereby dividing all participants into several knowledge clusters.

[0029] In some embodiments, the soft tags within each knowledge cluster are weighted and fused based on the amount of data from each participant to obtain the cluster-level fused soft tags for each knowledge cluster, specifically including:

[0030] For each knowledge cluster, obtain the noise-added soft tags corresponding to all supply chain participants within the knowledge cluster, as well as the local data volume of each supply chain participant;

[0031] For each sample in the unlabeled public dataset related to the supply chain, weights are assigned to the probability vector of the sample in the noisy soft labels of each supply chain participant based on the amount of local data, and a weighted average is performed to obtain the weighted average probability vector of the sample under the knowledge cluster.

[0032] The weighted average probability vectors of all samples are arranged in the original order of the samples in the unlabeled public dataset to obtain the cluster-level fused soft labels corresponding to the knowledge clusters.

[0033] In some embodiments, combining all cluster-level fused soft labels with each sample in the public dataset to form a multi-objective training set to train the student model as a new global model specifically includes:

[0034] The feature vector of each sample in the public dataset related to the supply chain is used as input, and the cluster-level fusion soft label of all knowledge clusters corresponding to each sample is used as multiple target outputs to construct a multi-objective training dataset.

[0035] Initialize a student model whose structure is consistent with the structure of the initial global model distributed by the central server;

[0036] The student model is trained in a supervised iterative manner using the multi-objective training dataset. In each round of training, the difference between the student model output and the soft label fused at each cluster level is calculated, and the weighted sum of the differences of all clusters is used as the total loss function. The student model parameters are updated through the backpropagation algorithm until the model converges.

[0037] The trained student model is used as the new global model.

[0038] In addition, before the central server calculates the similarity between any two soft tags of supply chain participants based on a preset distribution similarity metric, the central server also performs dimensionality reduction and normalization preprocessing on the soft tags, specifically including:

[0039] The central server retrieves the noise-added soft tags uploaded by all supply chain participants;

[0040] The dimensionality reduction of the noisy soft tags for each supply chain participant is performed to obtain the dimensionality-reduced soft tag features;

[0041] The reduced-dimensional soft label features are normalized to obtain preprocessed soft label features.

[0042] Secondly, this application provides a supply chain data collaboration platform based on federated learning, comprising:

[0043] The training module is used by the central server to distribute the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data, and uses the local teacher model to predict unlabeled public datasets related to the supply chain, generating soft labels that represent the knowledge of local data distribution.

[0044] The processing module is used by each supply chain participant to add differential privacy noise to the generated soft tags before uploading them to the central server;

[0045] The processing module is also used by the central server to calculate the similarity between the soft tags of every two supply chain participants based on a preset distribution similarity metric, to obtain a soft tag difference matrix, and then to divide all supply chain participants into several knowledge clusters based on the soft tag difference matrix, wherein the supply chain participants in each knowledge cluster have similar data distributions.

[0046] The processing module is also used to perform weighted fusion of soft labels within each knowledge cluster according to the amount of data of each participant, to obtain cluster-level fused soft labels for each knowledge cluster, and to form a multi-objective training set with all cluster-level fused soft labels and each sample in the public dataset to train the student model as a new global model.

[0047] The execution module is used by the central server to distribute the new global model to each participant and repeat the above steps until the new global model converges, so that it can be used by each supply chain participant for the next round of supply chain business forecasting.

[0048] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0049] In the federated learning-based supply chain data collaboration platform and method provided in this application, firstly, the central server distributes an initial global model to each supply chain participant. Each participant, based on the initial global model and their local private data, trains a local teacher model. This local teacher model then predicts on unlabeled public datasets related to the supply chain, generating soft labels representing the distribution knowledge of the local data. This step enables the extraction and representation of local data distribution knowledge, transforming private data knowledge into a shareable soft label format, providing a foundation for subsequent privacy protection and participant clustering, thereby improving the accuracy of knowledge transfer. Secondly, each supply chain participant... After differential privacy noise is added to the generated soft tags, they are uploaded to the central server. This step enhances the privacy of the soft tags during transmission. By adding differential privacy noise, attackers are prevented from inferring local private data from the soft tags, while preserving the statistical characteristics of the soft tags, thus improving privacy security during data collaboration. Subsequently, the central server calculates the similarity between the soft tags of every two supply chain participants based on a preset distribution similarity metric, obtaining a soft tag difference matrix. Then, based on this soft tag difference matrix, all supply chain participants are divided into several knowledge clusters. Within each knowledge cluster, the supply chain participants have similar data distributions. This step enables data segmentation among participants. This paper proposes an effective measure and clustering method for similarity among participants. By eliminating noise and dimensionality effects through dimensionality reduction and normalization, participants with similar distributions are grouped into the same knowledge cluster based on spectral clustering. This improves the targeting of subsequent fusion and the model's adaptability to data heterogeneity. Then, the soft labels within each knowledge cluster are weighted and fused according to the data volume of each participant, resulting in cluster-level fused soft labels for each knowledge cluster. All cluster-level fused soft labels are combined with each sample in the public dataset to form a multi-objective training set, which is used to train the student model as a new global model. This step enables weighted fusion of intra-cluster knowledge and multi-objective distillation training. The fused cluster-level soft labels represent the collective knowledge of each cluster. Through multi-objective training... This approach enables student models to adapt to different data distributions simultaneously, thereby improving the generalization ability of the global model under heterogeneous data. Finally, the central server distributes the new global model to each participant and repeats the above steps until the new global model converges, which is then used for the next round of supply chain business forecasting by each supply chain participant. This step enables iterative optimization and final deployment of the global model. By repeating the above steps, the model gradually converges to a balance point that takes into account the knowledge of each participant, and is ultimately used for actual supply chain business forecasting, thereby improving the model's practicality and prediction accuracy. In summary, the solution proposed in this application can achieve supply chain data collaboration through iterative optimization of the global model while protecting data privacy. Attached Figure Description

[0050] Figure 1 This is an exemplary flowchart of a federated learning-based supply chain data collaboration method according to some embodiments of this application;

[0051] Figure 2 This is a schematic diagram illustrating an application scenario of a supply chain data collaboration system according to some embodiments of this application;

[0052] Figure 3 This is a schematic diagram of the process for obtaining cluster-level fused soft tags according to some embodiments of this application;

[0053] Figure 4 This is a schematic diagram of the structure of a federated learning-based supply chain data collaboration platform according to some embodiments of this application;

[0054] Figure 5 This is a schematic diagram of the structure of a computer device that implements a federated learning-based supply chain data collaboration method according to some embodiments of this application. Detailed Implementation

[0055] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0056] refer to Figure 1 The figure is an exemplary flowchart of a federated learning-based supply chain data collaboration method according to some embodiments of this application. This federated learning-based supply chain data collaboration method mainly includes the following steps:

[0057] In step 101, the central server distributes the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data, and uses the local teacher model to predict unlabeled public datasets related to the supply chain, generating soft labels that represent knowledge of local data distribution.

[0058] In specific implementation, the initial global model can be constructed in the following manner: The central server first determines the model structure based on the target task type of supply chain business collaboration. The target task types include supply chain demand forecasting, inventory optimization, or logistics timeliness analysis. For different task types, the central server selects the corresponding deep neural network architecture as the base model. For example, for supply chain demand forecasting tasks, a recurrent neural network structure composed of long short-term memory networks or gated recurrent units is used; for inventory optimization tasks, a fully connected feedforward neural network structure is used; and for logistics timeliness analysis tasks, a structure combining convolutional neural networks and fully connected layers is used. Then, the central server determines the specific parameter configuration of the model. This includes the input layer dimension, the number of hidden layers, the number of neurons in each hidden layer, the number of neurons in the output layer, and the activation function type. The input layer dimension is consistent with the sample feature dimension of the unlabeled public dataset related to the supply chain described in subsequent steps. The number of neurons in the output layer is set according to the number of categories in the prediction task or the output dimension of the regression task. Next, the central server initializes all trainable weight parameters in the model using a random initialization method. The random initialization method includes randomly sampling initial values ​​from a uniform or normal distribution and initializing all bias term parameters to zero, thereby constructing an initial global model. Other methods can also be used in other embodiments, which are not limited here.

[0059] It should be noted that the initial global model in this application refers to the baseline model built by the central server and distributed to all supply chain participants when the federated learning process is started. Its role is to provide a unified starting point for parameters for each participant, enabling them to train on local private data and serve as the basis for subsequent iterative optimization. Finally, through multiple rounds of collaboration, it converges into a global model that can integrate the knowledge of each participant.

[0060] In specific implementation, the central server can distribute the initial global model to each supply chain participant in the following way: the central server packages the network structure definition file of the completed initial global model, the initialized weight parameter file, and the hyperparameter configuration file required for model training into a complete model distribution package, and distributes it to each supply chain participant through a secure transmission protocol, so that each participant can conduct subsequent teacher model training locally based on private data. Other methods can also be used in other embodiments, which are not limited here.

[0061] In some embodiments, each supply chain participant can train a local teacher model based on the initial global model and local private data using the following steps:

[0062] After receiving the initial global model distributed by the central server, each supply chain participant obtains the corresponding local private data;

[0063] Each supply chain participant preprocesses its corresponding local private data and divides the preprocessed local private data into local training datasets and local validation datasets.

[0064] Each supply chain participant performs supervised iterative training on the initial global model based on the local training dataset. After each round of training, the performance of the initial global model is evaluated using the local validation dataset. Training stops when the loss function value of the initial global model on the local validation dataset no longer decreases. The model parameters obtained at this point are used as the final parameters of the local teacher model, thus obtaining the local teacher model.

[0065] In specific implementation, after receiving the initial global model distributed by the central server, each supply chain participant can obtain the corresponding local private data in the following way: Each supply chain participant creates a dedicated data storage directory on the local server. This data storage directory is used to store historical data accumulated by its own business system. The historical data includes, but is not limited to, structured data related to the supply chain such as order records, inventory levels, logistics timeliness, and point-of-sale data. Each participant extracts data records within a preset time range from the local data warehouse through a database query interface or file reading interface, and stores the extracted data in a unified table format (such as a CSV file) in the data storage directory. At the same time, the total number of data samples and feature dimension information are recorded, thereby obtaining local private data for subsequent model training. Other methods can also be used in other embodiments, which are not limited here.

[0066] In specific implementation, each supply chain participant preprocesses its corresponding local private data, and the preprocessed local private data is divided into local training datasets and local validation datasets. This can be achieved in the following way: First, each supply chain participant cleans the acquired local private data, including deleting duplicate records, handling outliers, and filling missing values ​​using mean imputation or interpolation. Then, the cleaned data undergoes feature engineering, including one-hot encoding of categorical features, normalization or scaling of numerical features, and constructing new derived features according to business needs. Next, the feature-engineered dataset is randomly divided into local training datasets and local validation datasets according to a preset ratio. The preset ratio is typically 80% of the data used as the local training dataset for model parameter updates and 20% of the data used as the local validation dataset for model performance evaluation. Finally, the two datasets are saved as independent files or database tables, and their respective sample sizes are recorded to ensure accurate loading and use during subsequent training. Other methods can also be used in other embodiments, which are not limited here.

[0067] In practice, each supply chain participant performs supervised iterative training on the initial global model using the local training dataset. After each training round, the performance of the initial global model is evaluated using the local validation dataset. Training stops when the loss function value of the initial global model on the local validation dataset no longer decreases. The model parameters obtained at this point are used as the final parameters of the local teacher model. The local teacher model can be implemented as follows: each supply chain participant loads the network structure and initializes the weights of the initial global model, and inputs the local training dataset into the model in batches for forward propagation to calculate predicted values. Then, the loss is calculated based on the difference between the predicted values ​​and the true labels. The function value is then calculated, and the gradient is calculated using the backpropagation algorithm. The optimizer is then used to update the model weights to complete one training epoch. After each training epoch, the local validation dataset is input into the current model for forward propagation. The loss function value on the validation set is calculated and recorded. The above training and validation process is repeated. An early stopping mechanism is set, that is, when the loss function value on the validation set does not decrease in multiple consecutive training epochs, the model is considered to have converged and training is stopped. The weight parameters saved by the model when training stops are used as the final parameters, and together with the model structure definition, they form the local teacher model, which is stored locally for subsequent soft label generation. Other methods can also be used in other embodiments, which are not limited here.

[0068] In some embodiments, generating soft labels representing knowledge of local data distribution by predicting unlabeled public datasets related to the supply chain using the local teacher model can be achieved through the following steps:

[0069] Each supply chain participant obtains an unlabeled public dataset related to the supply chain, which contains feature vectors of multiple unlabeled business samples;

[0070] Each supply chain participant inputs the feature vector of each sample in the corresponding unlabeled public dataset into the corresponding local teacher model, performs forward inference calculation, and obtains the probability output vector of each sample corresponding to each prediction category.

[0071] Each supply chain participant arranges the probability output vectors of all samples according to the original order of the samples in the public dataset, forming a soft label set that corresponds one-to-one with the samples in the public dataset. The soft label set is the soft label that represents the knowledge of local data distribution.

[0072] It should be noted that the probability output vector in this application is the probability distribution vector output by the local teacher model after predicting each sample in the unlabeled public dataset related to the supply chain. Each element represents the confidence level of the sample belonging to the corresponding predicted category. Its function is to quantify the prediction result of the teacher model for the sample, thereby indirectly encoding the knowledge of the local data distribution.

[0073] In specific implementation, each supply chain participant obtains an unlabeled public dataset related to the supply chain. This unlabeled public dataset contains feature vectors of multiple unlabeled samples, which can be implemented in the following way: Each supply chain participant downloads a pre-collected public dataset file from a unified resource address provided by a central server. The content of the public dataset file comes from industry public databases or third-party data service providers, such as macroeconomic indicators, weather data, holiday information, logistics indices, and other multi-dimensional time series data potentially related to supply chain business. Each sample consists of a set of numerical or categorical features and does not contain any real labels for business forecasting tasks. The participants parse the downloaded file into a data structure in memory, extract the feature vectors of all samples, and number them according to the original order of the samples in the file, thereby obtaining an ordered unlabeled public dataset for subsequent soft label generation. Other methods can also be used in other embodiments, which are not limited here.

[0074] In practice, each supply chain participant inputs the feature vector of each sample in the corresponding unlabeled public dataset into the corresponding local teacher model, performs forward inference calculation, and obtains the probability output vector of each sample corresponding to each predicted category. This can be achieved as follows: each supply chain participant loads the trained local teacher model and sets the model to evaluation mode to disable dropout. The training layer is designed for training. Then, each sample in the unlabeled public dataset is sequentially traversed, and the feature vector of each sample is reshaped into a tensor format required by the model input. This tensor is then input into the local teacher model for a forward propagation calculation. The model output layer uses the softmax activation function to convert logits into a probability distribution, obtaining the probability values ​​of the sample corresponding to all predicted classes. These probability values ​​form a vector of the same length as the number of classes. The above operations are performed sequentially on all samples to generate a set of probability output vectors equal to the number of samples. Dropout is a regularization technique used during training. In each forward propagation, it randomly discards some neurons in the network with a certain probability, thereby preventing the model from over-relying on specific neurons and reducing overfitting. When switching to evaluation mode, the dropout layer is turned off, and all neurons participate in the calculation, ensuring the determinism of the inference result. Logits refer to the raw values ​​generated by the linear output of the last layer in the model. These values ​​have not yet been transformed by the softmax function, so they can be any real numbers and do not satisfy the constraint that the sum of probabilities is 1. The softmax function then compresses the logits into a probability distribution between 0 and 1 for classification prediction. Other methods can be used in other embodiments, which are not limited here.

[0075] In specific implementation, each supply chain participant arranges the probability output vectors of all samples according to the original order of the samples in the public dataset, forming a soft label set that corresponds one-to-one with the samples in the public dataset. This soft label set, representing the knowledge of local data distribution, can be implemented as follows: Each supply chain participant creates an empty list and appends the probability output vector of each sample to the empty list in the order of traversing the public dataset. After all samples have been processed, the entire list is converted into a two-dimensional array or tensor, where the first dimension index corresponds to the sample number and the second dimension index corresponds to the predicted category, thus obtaining a soft label matrix that is completely consistent with the order of the samples in the public dataset. This matrix is ​​the soft label set generated by the local teacher model on the public dataset, and its values ​​reflect the knowledge of local data distribution. For example, for the same public sample, the differences in soft labels from different participants reflect the differences in their local data distribution. Finally, the participants serialize the soft label matrix into a file, such as the .npy format of NumPy or the Parquet format, in preparation for the subsequent differential privacy noise addition and uploading steps. Other methods can also be used in other embodiments, which are not limited here.

[0076] It should be noted that the above steps can extract and represent knowledge from local data distribution, transforming private data knowledge into a shareable soft label form, providing a foundation for subsequent privacy protection and participant clustering, thereby improving the accuracy of knowledge transfer.

[0077] In some embodiments, reference Figure 2 As shown in the figure, this figure is a schematic diagram of the application scenario of the supply chain data collaboration system shown in some embodiments of this application. The figure includes three main components: a data acquisition device, a server, and a data storage device. The data acquisition device is responsible for collecting the local private data of each participant and sending the collected local private data to the server through a communication network. The supply chain data collaboration system runs on the server, and the server stores the processing results in the data storage device and visualizes them.

[0078] In step 102, each supply chain participant adds differential privacy noise to the generated soft tag and then uploads it to the central server.

[0079] In some embodiments, the process of adding differential privacy noise to the generated soft tags before uploading them to the central server by each supply chain participant can be achieved through the following steps:

[0080] Each supply chain participant adds differential privacy noise to the generated soft tag to obtain a noisy soft tag;

[0081] The noise-added soft tags are uploaded to the central server.

[0082] It should be noted that the differential privacy noise in this application is a random perturbation signal added by each supply chain participant after generating the soft tag. Its purpose is to perturb the original probability value before uploading the soft tag to the central server, making it difficult for attackers to reverse-engineer sensitive information of local private data from the uploaded data, while preserving the statistical characteristics of the soft tag.

[0083] In practice, each supply chain participant adds differential privacy noise to the generated soft tags. The noisy soft tags can be obtained as follows: Each supply chain participant first reads the generated soft tag matrix. The rows of this matrix correspond to samples in the unlabeled public dataset related to the supply chain, the columns correspond to the categories of the prediction task, and each element is a probability value between 0 and 1. Then, each supply chain participant presets a privacy budget parameter ε, which is set according to the privacy protection strength requirements, for example, a value of 1.0, and calculates the sensitivity of each element in the soft tag. Since the probability value ranges between 0 and 1 and a modification to a single sample may change a maximum probability value of 1, the sensitivity is usually set to 1. Then, the soft tag matrix is ​​processed... Each element in the matrix is ​​independently added with random noise following a Laplace distribution. This involves generating a Laplace random number with a mean of 0 and a scale equal to the sensitivity divided by the privacy budget, and adding it to the original probability value. The new value obtained after adding noise may exceed the range of 0 to 1, so truncation is required. Values ​​less than 0 are set to 0, and values ​​greater than 1 are set to 1. After truncation, the probability values ​​of each category for each sample are normalized again so that their sum is 1, thus ensuring that the soft label after adding noise still conforms to the probability distribution property. Finally, the processed soft label matrix is ​​saved as a new file named "Noisy Soft Label" and stored in the local directory to be uploaded. Other methods can also be used in other embodiments, which are not limited here.

[0084] In specific implementation, uploading the noisy soft tags to the central server can be achieved in the following way: Each supply chain participant reads the noisy soft tag file stored locally, and prepares a unique identifier for the supply chain participant, local data volume statistics, and the current federated learning round number; then, it establishes communication with the central server through a secure transmission protocol, such as HTTPS or TLS encrypted TCP connection, and packages the noisy soft tag file along with the aforementioned metadata into a request, for example, uploading it via a POST request in multipart / form-data format; after receiving the uploaded data, the central server first verifies the participant's identity and the integrity of the data, and then stores the noisy soft tag file in a designated storage area of ​​the server according to the participant ID and round number, and records the successful upload status for subsequent soft tag preprocessing and cluster analysis. Other methods can also be used in other embodiments, which are not limited here.

[0085] It should be noted that the above steps can achieve privacy-enhanced transmission of soft tags. By adding differential privacy noise, attackers are prevented from inferring local private data from the soft tags, while the statistical characteristics of the soft tags are preserved, thereby improving privacy and security in the data collaboration process.

[0086] In step 103, the central server calculates the similarity between the soft tags of every two supply chain participants based on a preset distribution similarity metric, and obtains a soft tag difference matrix. Then, based on the soft tag difference matrix, all supply chain participants are divided into several knowledge clusters, wherein the supply chain participants in each knowledge cluster have similar data distributions.

[0087] In some embodiments, the central server also performs dimensionality reduction and normalization preprocessing on the soft tags, which can be achieved through the following steps:

[0088] The central server retrieves the noise-added soft tags uploaded by all supply chain participants;

[0089] The dimensionality reduction of the noisy soft tags for each supply chain participant is performed to obtain the dimensionality-reduced soft tag features;

[0090] The reduced-dimensional soft label features are normalized to obtain preprocessed soft label features.

[0091] In specific implementation, the central server can obtain the noisy soft tags uploaded by all supply chain participants in the following way: The central server traverses all noisy soft tag files uploaded by participants in its storage directory. Each file is named with a unique identifier of the participant and the current federated learning round number. The server loads these files one by one through the file reading interface, reads the noisy soft tag data of each participant into memory, and builds an index according to the supply chain participant ID. At the same time, it records the dimension information of the soft tag matrix of each supply chain participant. Finally, a dictionary structure with the supply chain participant ID as the key and the noisy soft tag matrix as the value is obtained. Other methods can also be used in other embodiments, which are not limited here.

[0092] In practice, the dimensionality reduction of the noisy soft labels for each supply chain participant is achieved by the following method: The central server first concatenates the noisy soft label matrices of all supply chain participants into a large three-dimensional tensor. The dimension of this three-dimensional tensor is the number of supply chain participants multiplied by the number of samples in the public dataset multiplied by the number of predicted categories. Then, principal component analysis is used to reduce the dimensionality of this three-dimensional tensor. Specifically, the three-dimensional tensor is reshaped into a two-dimensional matrix, where rows correspond to each sample of each participant, and columns correspond to the original predicted category features. The covariance matrix of this two-dimensional matrix is ​​then calculated and further processed. The eigenvalue decomposition is performed, and the top principal components whose sum of contribution rates reaches a preset threshold, such as 95%, are selected as the target dimensions for dimensionality reduction. Then, the original two-dimensional matrix is ​​projected onto these principal components to obtain the dimensionality-reduced two-dimensional matrix. Finally, the two-dimensional matrix is ​​re-splittered according to the supply chain participants and samples, and the dimensionality-reduced soft label feature matrix is ​​the one corresponding to each supply chain participant. The number of rows in the feature matrix of each supply chain participant is still equal to the number of samples in the public dataset, and the number of columns is equal to the number of principal components selected. These dimensionality-reduced soft label feature matrices are the dimensionality-reduced soft label features. Other methods can also be used in other embodiments, which are not limited here.

[0093] In specific implementation, the dimensionality-reduced soft label features are normalized to obtain preprocessed soft label features. This can be achieved as follows: The central server performs L2 normalization on each row of the dimensionality-reduced soft label feature matrix of each supply chain participant. For the feature vector of each sample, the sum of squares of all its elements is calculated and the square root is taken to obtain the L2 norm. Then, each element in the feature vector is divided by the L2 norm, so that the normalized feature vector of each sample has a unit length. After performing the above operation on all samples of all supply chain participants, the preprocessed soft label feature matrix corresponding to each supply chain participant is obtained. This soft label feature matrix maintains a structure where the number of rows equals the number of samples and the number of columns equals the dimensionality after dimensionality reduction. The norm of each row vector is 1, providing dimensionless standardized input for subsequent similarity calculation. Other methods can also be used in other embodiments, which are not limited here.

[0094] In some embodiments, the central server calculates the similarity between soft tags of every two supply chain participants based on a preset distribution similarity metric, and the soft tag difference matrix can be obtained by the following steps:

[0095] The similarity value between the preprocessed soft tag features of each two supply chain participants is calculated based on a preset distribution similarity metric.

[0096] Combine all similarity values ​​into a soft-label difference matrix.

[0097] It should be noted that the distribution similarity metric preset in this application is a predefined mathematical measure used to quantitatively compare the similarity between soft tags of two supply chain participants. Its function is to transform the data distribution differences implied by the soft tags into calculable values, thereby constructing a soft tag difference matrix.

[0098] In specific implementation, the similarity value between the preprocessed soft-label features of each pair of supply chain participants can be calculated based on a preset distribution similarity metric in the following manner: The central server first obtains the preprocessed soft-label feature matrix of all supply chain participants. The rows of each soft-label feature matrix correspond to samples in the public dataset, and the columns correspond to the feature dimensions after dimensionality reduction. Then, for each pair of supply chain participants, the normalized feature vectors corresponding to each sample of the two supply chain participants are extracted in sequence, and the inner product of the two vectors is calculated. Since the vectors have been normalized, the inner product is the cosine similarity. The arithmetic mean of the cosine similarity values ​​calculated for all samples is taken to obtain the overall similarity value between the pair of supply chain participants. This value is between -1 and 1. The closer it is to 1, the more similar the data distributions of the two supply chain participants are. The above calculation is repeated for all possible pairs of supply chain participants to obtain all similarity values. The supply chain participant pairs also include supply chain participant pairs consisting of themselves. Other methods can also be used in other embodiments, which are not limited here.

[0099] In specific implementation, combining all similarity values ​​into a soft-label difference matrix can be achieved in the following way: The central server creates a two-dimensional matrix of size equal to the total number of supply chain participants multiplied by the total number of supply chain participants. The rows and columns of this two-dimensional matrix are arranged in a fixed order according to the supply chain participant IDs. The calculated similarity values ​​of each pair of supply chain participants are filled into the corresponding row and column intersection positions in the matrix. For diagonal elements with the same row and column index, the value 1 is filled in, indicating that the supply chain participant is completely similar to itself. The final matrix is ​​the soft-label difference matrix, which is a symmetric matrix, where each element represents the degree of similarity between the corresponding two participants in the knowledge distribution. Other methods can also be used in other embodiments, which are not limited here.

[0100] In some embodiments, dividing all supply chain participants into several knowledge clusters based on the soft-label difference matrix can be achieved by the following steps:

[0101] A similarity graph is constructed based on the soft-label difference matrix. The nodes in the similarity graph are supply chain participants, and the weights of the edges are determined by the similarity values ​​in the soft-label difference matrix.

[0102] Calculate the Laplacian matrix of the similarity graph and perform eigenvalue decomposition on the Laplacian matrix to obtain eigenvectors;

[0103] Clustering is performed based on the feature vectors to obtain the cluster label of each supply chain participant, thereby dividing all participants into several knowledge clusters.

[0104] In specific implementation, a similarity graph is constructed based on the soft-label difference matrix. The nodes in the similarity graph represent supply chain participants, and the edge weights are determined by the similarity values ​​in the soft-label difference matrix. This can be achieved as follows: the central server treats the soft-label difference matrix as a weighted adjacency matrix, where each element represents the weight of the edge between two corresponding supply chain participants. To construct the similarity graph, the similarity value first needs to be converted into a distance or weight suitable for graph construction. Typically, a Gaussian kernel function is used to convert the similarity value into non-negative edge weights. That is, for each pair of supply chain participants, their similarity value is calculated, and then... The numerical function mapping assigns larger edge weights to supply chain participants with high similarity, while the weights of low or negative similarity approaches zero after mapping. Alternatively, a fully connected graph can be constructed, where edges exist between all supply chain participants and their weights are determined by the mapped values. A k-nearest neighbor graph can also be used, where each node is connected only to its k most similar nodes, and the weights of the remaining edges are set to zero to reduce graph complexity. The final result is a similarity graph with supply chain participants as nodes and calculated non-negative weights as edges. Other methods can also be used in other embodiments, which are not limited here.

[0105] In specific implementation, the Laplacian matrix of the similar graph is calculated, and the Laplacian matrix is ​​decomposed into eigenvectors. This can be achieved in the following way: The central server first extracts the degree matrix from the similar graph. The degree matrix is ​​a diagonal matrix, and each element on the diagonal is equal to the sum of the weights of all edges connected to that node. Then, the Laplacian matrix is ​​constructed, usually using a normalized Laplacian matrix, which is obtained by multiplying the inverse square root of the degree matrix by (degree matrix minus adjacency matrix) and then by the inverse square root of the degree matrix. Next, the normalized Laplacian matrix is ​​decomposed into eigenvalues, and all eigenvalues ​​and their corresponding eigenvectors are calculated. The eigenvectors corresponding to the k smallest non-zero eigenvalues ​​are selected, where k is a preset number of clusters. These eigenvectors are arranged in columns to form a matrix. The number of rows in this matrix is ​​the number of participants, and the number of columns is k. Other methods can also be used in other embodiments, which are not limited here.

[0106] In specific implementation, clustering is performed based on the feature vectors to obtain the cluster label of each supply chain participant, thereby dividing all participants into several knowledge clusters. This can be achieved in the following way: The central server regards each row of the matrix composed of k feature vectors obtained in the previous step as a new representation of a supply chain participant in the k-dimensional feature space; then, the K-means clustering algorithm is performed on the row vectors of the matrix, setting the number of clusters to k, and allocating each row vector to k clusters through iterative optimization, so as to minimize the sum of squared distances from each sample to its cluster center; after clustering, each supply chain participant is assigned a cluster label, indicating the knowledge cluster to which it belongs; finally, all supply chain participants are divided into several knowledge clusters according to the cluster labels. Supply chain participants within each cluster are considered to have similar data distributions because they are close to each other in the feature vector space. Other methods can also be used in other embodiments, which are not limited here.

[0107] It should be noted that the above steps can effectively measure and cluster the similarity of the data distribution of the participants. By eliminating noise and the influence of dimensions through dimensionality reduction and normalization, participants with similar distributions are divided into the same knowledge cluster based on spectral clustering, thereby improving the targeting of subsequent fusion and the adaptability of the model to data heterogeneity.

[0108] In step 104, the soft labels within each knowledge cluster are weighted and fused according to the amount of data of each participant to obtain the cluster-level fused soft label of each knowledge cluster. All the cluster-level fused soft labels are combined with each sample in the public dataset to form a multi-objective training set to train the student model as a new global model.

[0109] In some embodiments, reference Figure 3 As shown in the figure, this is a schematic diagram of the process of obtaining cluster-level fused soft tags in some embodiments of this application. In this embodiment, the soft tags in each knowledge cluster are weighted and fused according to the amount of data of each participant. The cluster-level fused soft tags for each knowledge cluster can be obtained by the following steps:

[0110] In step 1031, for each knowledge cluster, the noise-added soft tags corresponding to all supply chain participants within the knowledge cluster and the local data volume of each supply chain participant are obtained.

[0111] In step 1032, for each sample in the unlabeled public dataset related to the supply chain, weights are assigned to the probability vector of the sample in the noisy soft labels of each supply chain participant based on the amount of local data, and a weighted average is performed to obtain the weighted average probability vector of the sample under the knowledge cluster.

[0112] In step 1033, the weighted average probability vectors of all samples are arranged in the original order of the samples in the unlabeled public dataset to obtain the cluster-level fusion soft label corresponding to the knowledge cluster.

[0113] It should be noted that the weighted average probability vector in this application is a comprehensive vector obtained by weighting the probability output vectors of all participants in the knowledge cluster for each sample in the unlabeled public dataset related to the supply chain according to their local data volume. Its function is to integrate the collective knowledge of each participant in the cluster to form a cluster-level fusion soft label representing the overall data distribution characteristics of the knowledge cluster.

[0114] In specific implementation, for each knowledge cluster, obtaining the noisy soft tags corresponding to all supply chain participants within the knowledge cluster and the local data volume of each supply chain participant can be achieved in the following way: The central server first groups the supply chain participants by cluster based on the cluster tags obtained from spectral clustering, obtaining a list of supply chain participant IDs contained in each knowledge cluster; then, iterates through each knowledge cluster, and for each supply chain participant ID within the cluster, reads the noisy soft tag file uploaded by the supply chain participant in this round from the server's storage area, and simultaneously extracts the local data volume information reported by the supply chain participant from the previously recorded supply chain participant metadata, that is, the total number of samples counted by the supply chain participant during local private data preprocessing; finally, loads the noisy soft tag data of all supply chain participants within each knowledge cluster into memory, and establishes a mapping with the supply chain participant ID as the key, while organizing the corresponding local data volume into a list or dictionary to ensure that the soft tag of each supply chain participant corresponds one-to-one with its data volume. Other methods can also be used in other embodiments, which are not limited here.

[0115] In specific implementation, for each sample in the unlabeled public dataset related to the supply chain, weights are assigned to the probability vector of the sample in the noisy soft labels of each supply chain participant based on the local data volume, and a weighted average is performed to obtain the weighted average probability vector of the sample under the knowledge cluster. This can be achieved in the following way: The central server first obtains the total number of samples in the unlabeled public dataset, denoted as M, and obtains the number of supply chain participants in the current knowledge cluster, denoted as N_c; then, a zero matrix of size M multiplied by the number of predicted categories is created to store the weighted average result of the cluster; then, the sample index is looped from 0 to M-1, and for each sample, the noisy soft label matrix of each supply chain participant in the knowledge cluster is retrieved. The probability vector corresponding to the sample has a length equal to the number of predicted categories. Simultaneously, the local data volume of each supply chain participant is used as a weight. The probability vectors of all supply chain participants for this sample are weighted and summed according to their respective data volumes. That is, each probability vector is multiplied by its corresponding data volume and then summed. Finally, this sum is divided by the total data volume of all supply chain participants within the knowledge cluster to obtain the weighted average probability vector of the sample. This weighted average probability vector is stored in the corresponding row of the previously created zero matrix. After performing the above operations on all samples, a matrix of size M multiplied by the number of predicted categories is obtained. Each row of this matrix is ​​the weighted average probability vector of the corresponding sample under the knowledge cluster. Other methods can also be used in other embodiments, which are not limited here.

[0116] In specific implementation, the weighted average probability vectors of all samples are arranged according to the original order of the samples in the unlabeled public dataset. The cluster-level fusion soft label corresponding to the knowledge cluster can be obtained in the following way: the central server directly saves the matrix obtained by the arrangement as a file, named it cluster-level fusion soft label, and marks the knowledge cluster number to which it belongs, while recording the dimension information of the matrix; finally, a cluster-level fusion soft label is obtained for each knowledge cluster. The cluster-level fusion soft label is the fusion representation of the knowledge of all participants in the knowledge cluster. Other methods can also be used in other embodiments, which are not limited here.

[0117] In some embodiments, training a student model as a new global model by combining all cluster-level fused soft labels with each sample in a public dataset to form a multi-objective training set can be achieved using the following steps:

[0118] The feature vector of each sample in the public dataset related to the supply chain is used as input, and the cluster-level fusion soft label of all knowledge clusters corresponding to each sample is used as multiple target outputs to construct a multi-objective training dataset.

[0119] Initialize a student model whose structure is consistent with the structure of the initial global model distributed by the central server;

[0120] The student model is trained in a supervised iterative manner using the multi-objective training dataset. In each round of training, the difference between the student model output and the soft label fused at each cluster level is calculated, and the weighted sum of the differences of all clusters is used as the total loss function. The student model parameters are updated through the backpropagation algorithm until the model converges.

[0121] The trained student model is used as the new global model.

[0122] In practical implementation, the feature vector of each sample in the public dataset related to the supply chain is used as input, and the cluster-level fused soft label of all knowledge clusters corresponding to each sample is used as multiple target outputs. The multi-target training dataset can be constructed in the following way: The central server first obtains the original sample feature matrix of the unlabeled public dataset related to the supply chain. The rows of the original sample feature matrix correspond to the samples, and the columns correspond to the feature dimensions. Then, it obtains the cluster-level fused soft label matrix of all knowledge clusters. Assuming there are K knowledge clusters, the size of the soft label matrix of each knowledge cluster is the number of samples multiplied by the number of predicted categories. The central server then uses the feature vector of each sample in the public dataset related to the supply chain as input, and the cluster-level fused soft label matrix of all knowledge clusters as outputs. The input matrix is ​​used as the input part. K cluster-level fused soft label matrices are stacked along a new dimension to form a multi-objective output tensor with a size of the number of samples multiplied by K multiplied by the number of predicted categories. Then, the input feature matrix is ​​paired with the multi-objective output tensor to construct a multi-objective training dataset, which is usually stored in memory as a data structure or in the form of a file. For example, the input feature matrix is ​​saved as one file and the multi-objective output tensor is saved as another file, and the correspondence between them is recorded to ensure that the input and all target outputs can be loaded simultaneously by sample index during training. Other methods can also be used in other embodiments, which are not limited here.

[0123] In specific implementation, a student model is initialized. The structure of the student model is consistent with the structure of the initial global model distributed by the central server. This can be achieved in the following way: The central server calls the model building module to re-instantiate a model object based on the network structure configuration file of the previously defined initial global model. This model object has the same input layer dimension, number of hidden layers, number of neurons in each layer, output layer dimension, and activation function settings. Then, all trainable weight parameters of this model are initialized using the same random initialization method as the initial global model, such as randomly sampling initial values ​​from a uniform distribution and initializing the bias term to zero. Finally, a student model with random initial weights but completely identical structure is obtained. This student model is used for subsequent multi-objective distillation training. Other methods can also be used in other embodiments, which are not limited here.

[0124] In specific implementation, the student model is trained using the multi-objective training dataset under supervised iterative conditions. In each training round, the difference between the student model output and the soft label fused at each cluster level is calculated, and the weighted sum of the differences of all clusters is used as the total loss function. The student model parameters are updated through the backpropagation algorithm until the model converges. This can be achieved as follows: The central server divides the multi-objective training dataset into multiple small batches according to batch size, such as 32 or 64. In each training round, all small batches are traversed. For each small batch, the input feature vector is input into the student model for forward propagation to obtain the student model's prediction output matrix for the samples in that batch, the size of which is equal to the batch size multiplied by the number of predicted categories. Then, for each sample in that batch, the cluster-level fusion soft label fused between the student model output and each knowledge cluster is calculated. The difference between labels is typically measured using cross-entropy or KL divergence. This involves calculating the loss element-wise between the probability vector output by the student model and the target probability vector. Then, the loss values ​​of all K clusters are summed using preset weights to obtain the total loss for that sample. These preset weights are, for example, a uniform weight of 1 / K or dynamically adjusted based on the number of participants within a cluster. The average loss of all samples within a batch is then averaged to obtain the average loss for that batch. The gradient of the loss function with respect to the trainable parameters of the student model is then calculated using the backpropagation algorithm, and the model parameters are updated using the optimizer. After one epoch, the average loss can be evaluated across the entire training set. If the loss value no longer decreases or reaches the preset maximum number of training epochs, training stops, indicating that the model has converged. Other methods can be used in other embodiments, which are not limited here.

[0125] In specific implementation, the trained student model can be used as a new global model in the following way: The central server extracts all weight parameters of the student model after training convergence, packages them together with its network structure definition file, and generates a new global model file; at the same time, the relevant information of the student model, such as the model version number, training round, and performance indicators on the validation set, is recorded in the model management database; finally, the student model is identified as the new global model for this round of federated learning and is ready to be distributed to each supply chain participant. Other methods can also be used in other embodiments, which are not limited here.

[0126] It should be noted that the above steps can achieve weighted fusion of knowledge within clusters and multi-objective distillation training. The fused cluster-level soft labels represent the collective knowledge of each cluster. Through multi-objective training, the student model can adapt to different data distributions at the same time, thereby improving the generalization ability of the global model under heterogeneous data.

[0127] In step 105, the central server distributes the new global model to each participant and repeats the above steps until the new global model converges, so that it can be used for the next round of supply chain business forecasting by each supply chain participant.

[0128] In practice, the central server distributes the new global model to each participant and repeats the above steps until the new global model converges, which can be used for the next round of supply chain business forecasting by each supply chain participant. This can be achieved as follows: The central server first uses the trained student model as the new global model. It packages and distributes the network structure definition file, weight parameter file, and hyperparameter configuration file of this new global model to all supply chain participants via a secure transmission protocol. Each supply chain participant receives the new model and replaces its local old model as the initial model for the next iteration. Then, the central server and each supply chain participant jointly initiate a new round of federated learning, repeating steps 101 to 105. Each supply chain participant retrains its local teacher model based on the newly received global model, generates noisy soft labels, and uploads them. The central server re-executes preprocessing, clustering, weighted fusion, and multi-objective distillation training to obtain another new global model. After each iteration, the central server utilizes reserved... The performance of the current global model is evaluated using a validation dataset (which can be a subset of samples from a public dataset with a small number of manually labeled real data). Evaluation metrics, such as prediction accuracy or loss value, are recorded and compared with the metrics of the previous model. If the improvement in metrics is less than a preset threshold (e.g., 0.1%) for multiple consecutive iterations, such as three consecutive iterations, or if the preset maximum number of iterations (e.g., 50 iterations) is reached, the global model is considered to have converged, and iteration is stopped. If it has not converged, the above distribution and training process is repeated. Finally, when the global model converges, the central server distributes the final global model to each supply chain participant. Each supply chain participant deploys the model to its local business system. In real-time supply chain business scenarios, the locally collected real-time feature data is input into the model to perform forward inference, and the prediction results are used for business decisions, such as demand forecasting, inventory optimization, or logistics timeliness analysis, thereby realizing supply chain data collaboration based on federated learning. Other methods can also be used in other embodiments, which are not limited here.

[0129] It should be noted that the above steps enable iterative optimization and final deployment of the global model. By repeating the above steps, the model gradually converges to a balance point that takes into account the knowledge of all participants, and is finally used for actual supply chain business forecasting, thereby improving the model's practicality and forecasting accuracy.

[0130] Furthermore, in another aspect of this application, in some embodiments, this application provides a supply chain data collaboration platform based on federated learning, referencing... Figure 4 The figure is a schematic diagram of the structure of a federated learning-based supply chain data collaboration platform according to some embodiments of this application. The federated learning-based supply chain data collaboration platform includes: a training module 401, a processing module 402, and an execution module 403, which are described below:

[0131] Training module 401, in this application, is mainly used by the central server to distribute the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data, and uses the local teacher model to predict the unlabeled public dataset related to the supply chain, generating soft labels that represent the knowledge of local data distribution.

[0132] Processing module 402, in this application, is mainly used by each supply chain participant to add differential privacy noise to the generated soft tag and then upload it to the central server;

[0133] The processing module 402 described in this application is also used by the central server to calculate the similarity between the soft tags of each two supply chain participants based on a preset distribution similarity metric, to obtain a soft tag difference matrix, and then to divide all supply chain participants into several knowledge clusters based on the soft tag difference matrix, wherein the supply chain participants in each knowledge cluster have similar data distributions.

[0134] The processing module 402 described in this application is further configured to perform weighted fusion of soft labels within each knowledge cluster based on the amount of data of each participant, to obtain cluster-level fused soft labels for each knowledge cluster, and to form a multi-objective training set with all cluster-level fused soft labels and each sample in the public dataset to train the student model as a new global model.

[0135] The execution module 403 in this application is mainly used by the central server to distribute the new global model to each participant and repeat the above steps until the new global model converges, so as to be used by each supply chain participant for the next round of supply chain business forecasting.

[0136] The modules in the aforementioned federated learning-based supply chain data collaboration platform can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can invoke and execute the corresponding operations of each module.

[0137] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 5As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores supply chain data collaboration data based on federated learning. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a supply chain data collaboration method based on federated learning.

[0138] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps described in the embodiment of the federated learning-based supply chain data collaboration method.

[0140] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the embodiment of the federated learning-based supply chain data collaboration method.

[0141] In one embodiment, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the embodiment of the federated learning-based supply chain data collaboration method.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for supply chain data collaboration based on federated learning, characterized in that, Includes the following steps: The central server distributes the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data. The local teacher model is then used to predict unlabeled public datasets related to the supply chain and generate soft labels that represent knowledge of the local data distribution. Each supply chain participant adds differential privacy noise to the generated soft tags before uploading them to the central server; The central server calculates the similarity between soft tags of every two supply chain participants based on a preset distribution similarity metric, and obtains a soft tag difference matrix. Then, based on the soft tag difference matrix, all supply chain participants are divided into several knowledge clusters, wherein the supply chain participants in each knowledge cluster have similar data distributions. Based on the amount of data from each participant, the soft labels within each knowledge cluster are weighted and fused to obtain the cluster-level fused soft labels for each knowledge cluster. All cluster-level fused soft labels are combined with each sample in the public dataset to form a multi-objective training set to train the student model as a new global model. The central server distributes the new global model to each participant and repeats the above steps until the new global model converges, which can then be used by each supply chain participant for the next round of supply chain business forecasting.

2. The method as described in claim 1, characterized in that, Each supply chain participant trains a local teacher model based on the initial global model and their local private data, specifically including: After receiving the initial global model distributed by the central server, each supply chain participant obtains the corresponding local private data; Each supply chain participant preprocesses its corresponding local private data and divides the preprocessed local private data into local training datasets and local validation datasets. Each supply chain participant performs supervised iterative training on the initial global model based on the local training dataset. After each round of training, the performance of the initial global model is evaluated using the local validation dataset. Training stops when the loss function value of the initial global model on the local validation dataset no longer decreases. The model parameters obtained at this point are used as the final parameters of the local teacher model, thus obtaining the local teacher model.

3. The method as described in claim 1, characterized in that, The local teacher model is used to predict unlabeled public datasets related to the supply chain, generating soft labels that represent knowledge of local data distribution. Specifically, this includes: Each supply chain participant obtains an unlabeled public dataset related to the supply chain, which contains feature vectors of multiple unlabeled business samples; Each supply chain participant inputs the feature vector of each sample in the corresponding unlabeled public dataset into the corresponding local teacher model, performs forward inference calculation, and obtains the probability output vector of each sample corresponding to each prediction category. Each supply chain participant arranges the probability output vectors of all samples according to the original order of the samples in the public dataset, forming a soft label set that corresponds one-to-one with the samples in the public dataset. The soft label set is the soft label that represents the knowledge of local data distribution.

4. The method as described in claim 1, characterized in that, Each supply chain participant adds differential privacy noise to the generated soft tags before uploading them to the central server. Specifically, this includes: Each supply chain participant adds differential privacy noise to the generated soft tag to obtain a noisy soft tag; The noise-added soft tags are uploaded to the central server.

5. The method as described in claim 1, characterized in that, The similarity between soft tags of every two supply chain participants is calculated based on a preset distribution similarity metric, resulting in a soft tag difference matrix that specifically includes: The similarity value between the preprocessed soft tag features of each two supply chain participants is calculated based on a preset distribution similarity metric. Combine all similarity values ​​into a soft-label difference matrix.

6. The method as described in claim 1, characterized in that, Based on the aforementioned soft-label difference matrix, all supply chain participants are divided into several knowledge clusters, specifically including: A similarity graph is constructed based on the soft-label difference matrix. The nodes in the similarity graph are supply chain participants, and the weights of the edges are determined by the similarity values ​​in the soft-label difference matrix. Calculate the Laplacian matrix of the similarity graph and perform eigenvalue decomposition on the Laplacian matrix to obtain eigenvectors; Clustering is performed based on the feature vectors to obtain the cluster label of each supply chain participant, thereby dividing all participants into several knowledge clusters.

7. The method as described in claim 1, characterized in that, Based on the data volume of each participant, the soft tags within each knowledge cluster are weighted and fused to obtain the cluster-level fused soft tags for each knowledge cluster, which specifically include: For each knowledge cluster, obtain the noise-added soft tags corresponding to all supply chain participants within the knowledge cluster, as well as the local data volume of each supply chain participant; For each sample in the unlabeled public dataset related to the supply chain, weights are assigned to the probability vector of the sample in the noisy soft labels of each supply chain participant based on the amount of local data, and a weighted average is performed to obtain the weighted average probability vector of the sample under the knowledge cluster. The weighted average probability vectors of all samples are arranged in the original order of the samples in the unlabeled public dataset to obtain the cluster-level fused soft labels corresponding to the knowledge clusters.

8. The method as described in claim 1, characterized in that, The process of combining all cluster-level fused soft labels with each sample from the public dataset to form a multi-objective training set for training the student model as a new global model specifically includes: The feature vector of each sample in the public dataset related to the supply chain is used as input, and the cluster-level fusion soft label of all knowledge clusters corresponding to each sample is used as multiple target outputs to construct a multi-objective training dataset. Initialize a student model whose structure is consistent with the structure of the initial global model distributed by the central server; The student model is trained in a supervised iterative manner using the multi-objective training dataset. In each round of training, the difference between the student model output and the soft label fused at each cluster level is calculated, and the weighted sum of the differences of all clusters is used as the total loss function. The student model parameters are updated through the backpropagation algorithm until the model converges. The trained student model is used as the new global model.

9. The method as described in claim 1, characterized in that, Before calculating the similarity between any two soft tags of supply chain participants based on a preset distribution similarity metric, the central server performs dimensionality reduction and normalization preprocessing on the soft tags, specifically including: The central server retrieves the noise-added soft tags uploaded by all supply chain participants; The dimensionality reduction of the noisy soft tags for each supply chain participant is performed to obtain the dimensionality-reduced soft tag features; The reduced-dimensional soft label features are normalized to obtain preprocessed soft label features.

10. A supply chain data collaboration platform based on federated learning, characterized in that, include: The training module is used by the central server to distribute the initial global model to each supply chain participant. Each supply chain participant trains a local teacher model based on the initial global model and local private data, and uses the local teacher model to predict unlabeled public datasets related to the supply chain, generating soft labels that represent the knowledge of local data distribution. The processing module is used by each supply chain participant to add differential privacy noise to the generated soft tags before uploading them to the central server; The processing module is also used by the central server to calculate the similarity between the soft tags of every two supply chain participants based on a preset distribution similarity metric, to obtain a soft tag difference matrix, and then to divide all supply chain participants into several knowledge clusters based on the soft tag difference matrix, wherein the supply chain participants in each knowledge cluster have similar data distributions. The processing module is also used to perform weighted fusion of soft labels within each knowledge cluster according to the amount of data of each participant, to obtain cluster-level fused soft labels for each knowledge cluster, and to form a multi-objective training set with all cluster-level fused soft labels and each sample in the public dataset to train the student model as a new global model. The execution module is used by the central server to distribute the new global model to each participant and repeat the above steps until the new global model converges, so that it can be used by each supply chain participant for the next round of supply chain business forecasting.

Images