Federated multitask learning in network automation

Abstract

A method, apparatus, and computer program product are provided for performing federated multitask learning using a federated learning aggregation environment and federated learning distributed nodes (FLDNs). In the context of a method, the method includes determining respective sets of FLDNs from a plurality of FLDNs for training corresponding personalized global machine learning (ML) models to perform different personalized tasks, where the different personalized tasks are determined based on one or more ML cases. The method further includes generating at least one first personalized global ML model for performing the first personalized task and at least one second personalized global ML model for performing the second personalized task. The method further includes performing a respective plurality of training iterations for training the first and second global personalized ML models using the corresponding FLDN set sums and by aggregating local model parameters of the local personalized ML models independently generated by the respective FLDNs.

Description

TECHNICAL FIELD

[0001] Example embodiments generally relate to federated multi-task learning (FMTL) techniques for managing network automation, covering both performance and trustworthiness aspects. BACKGROUND

[0002] Federated techniques for machine learning include training different versions of a machine learning model on different distributed nodes. However, due to the heterogeneity of different distributed nodes and local training data involved in training various model versions, conventional federated learning approaches can result in poor model performance trustworthiness. For example, a first model version trained by a first distributed node can exhibit sufficient performance, while a second model version trained by a second distributed node can exhibit poor performance. Typical federated machine learning approaches include averaging the first and second versions of the model, which can result in a third version of the model that underperforms compared to the first model version. As indicated above, current federated learning approaches exhibit potential deficiencies in which a distributed node can receive a final aggregated machine learning model that exhibits worse performance compared to an initial version of the machine learning model trained locally and independently by that distributed node. SUMMARY

[0003] A method, apparatus, and computer program product are provided in accordance with example embodiments to improve federated multi-task learning (FMTL) in network automation by generating personalized global machine learning models targeted to different subsets of multiple federated learning distributed nodes (FLDNs), where the FLDNs in each subset are associated with similar tasks, objectives, and / or characteristics.

[0004] Typical machine learning (ML) approaches can require centralizing all training data collected by distributed nodes onto a single central node for model training. To minimize data exchange between the distributed nodes and the central node where model training is performed, the present techniques perform federated learning (FL). For example, instead of training a model at a central node in previous approaches, various embodiments of the present techniques can iteratively train different versions of a model at different FLDNs, where each FLDN can include a different local training dataset. In some embodiments, during each training iteration, a central node (referred to herein as a federated learning aggregation (FLA) computing environment) aggregates the different locally versions of the model to generate a global version of the model, which can be distributed to the FLDNs for additional training iterations until model convergence is achieved.

[0005] In some embodiments, this technology performs FMTL, which includes training a shared global machine learning model using different versions of machine learning models trained collaboratively with different tasks in a set of FLDNs to generate a corresponding aggregated global machine learning model at the FLA for performing each task. For example, a first FLDN set may include one or more first training datasets associated with a first source domain and a first target domain, and a second FLDN set may include one or more second training datasets associated with a second source domain and a second target domain (e.g., the data in the first and second training datasets are heterogeneous). In some contexts, a first version of the machine learning model may be provided to the first and second FLDN sets and trained using the corresponding first or second training dataset. In some contexts, a global version of the machine learning model may be generated based on corresponding trained iterations of the first version of the machine learning model from the first and second FLDN sets. However, when applied to a first target domain, a second target domain, or both, the global version of the machine learning model may exhibit lower performance (e.g., lower statistical accuracy, etc.) and / or lower confidence. Various embodiments of this technology can overcome such challenges by providing personalized versions of machine learning models for training to different FLDN groups, wherein each personalized version of the machine learning model can be generated to be optimally suited to the local data distribution of the corresponding FLDN.

[0006] In various embodiments, the apparatus, method, and computer program product generate machine learning models that exhibit and / or enable various advantages for providing model credibility. For example, machine learning models generated by this technology can exhibit transparency, diversity, non-discrimination, and fairness, as well as technical robustness and security. As another example, machine learning models can enable or facilitate adequate privacy and data governance, accountability (e.g., auditability, minimization and reporting of negative impacts, trade-offs, and remedies), human factors and oversight, and social and environmental well-being.

[0007] In at least one embodiment, a method is provided, the method comprising: (i) for at least a first personalized task and a second personalized task among a plurality of personalized tasks, determining (503) the following: (1) a set of first FLDNs (105) associated with the first personalized task from a plurality of federated learning distributed nodes (FLDNs) (105), and (2) a set of second FLDNs (105) associated with the second personalized task from a plurality of FLDNs (105); (ii) generating (518) a first personalized machine learning (ML) model for performing the first personalized task, and a second personalized ML model for performing the second personalized task; and (iii) ) Perform (521) a plurality of model training iterations for training the first personalized ML model and the second personalized ML model, the plurality of model training iterations comprising: (1) in the respective initial model training iterations for the first personalized ML model and the second personalized ML model, (A) providing (521) the first personalized ML model to the respective FLDN (105) of the first set, and (B) providing (521) the second personalized ML model to the respective FLDN (105) of the second set; (2) in the respective subsequent model training iterations for the first personalized ML model and the second personalized ML model, (A) providing (521) the first personalized ML model to the respective FLDN (105) of the first set; 5) Provide (521) subsequent trained iterations of the first personalized ML model, and (B) provide (521) subsequent trained iterations of the first personalized ML model to the corresponding FLDN (105) of the second set, wherein: in the corresponding model training iteration, the corresponding FLDN (105) of the first set and the corresponding FLDN (105) of the second set independently train the corresponding personalized ML model to perform the corresponding personalized task; (3) in the corresponding model training iteration, receive (524) local model parameters from the corresponding FLDN (105), wherein the local model parameters are associated with the training of the corresponding personalized ML model, which is performed by the first set (4) (A) Based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the first set, generate (530) subsequent trained iterations of the first personalized ML model; and (B) Based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the second set, generate (530) subsequent trained iterations of the second personalized ML model, wherein: the output of the final iteration of each of the corresponding plurality of training iterations includes: a first personalized global ML model for performing the first personalized task and a second personalized global ML model for performing the second personalized task.

[0008] In some embodiments, the plurality of FLDNs (105) include one or more separate management domains; and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing the corresponding plurality of training iterations are performed using a cross-domain management domain (103).

[0009] In some embodiments, the plurality of FLDNs (105) include one or more network nodes; and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing corresponding multiple training iterations are performed using a Radio Access Network (RAN) management domain (103). In some embodiments, the plurality of FLDNs (105) include one or more Network Data Analysis Functions (NWDAFs); and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing corresponding multiple training iterations are performed using a core management domain (103). In some embodiments, the plurality of FLDNs (105) include one or more distributed NWDAFs; and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing corresponding multiple training iterations are performed using a central NWDAF (103).

[0010] In some embodiments, the plurality of FLDNs (105) include one or more user equipment (UE); and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing the corresponding plurality of training iterations are performed using network nodes (103).

[0011] In some embodiments, the plurality of FLDNs (105) include one or more real-time RAN intelligent controllers (RICs); and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing a plurality of training iterations accordingly are performed using a non-real-time RIC (103).

[0012] In some embodiments, the plurality of FLDNs (105) include one or more network nodes; and determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing a plurality of training iterations are performed using a real-time RIC (103). In some embodiments, determining a first set and a second set of the plurality of FLDNs (105), generating a first personalized ML model and a second personalized ML model, and performing a plurality of training iterations are performed using a first UE (103); and the plurality of FLDNs (105) include one or more additional UEs.

[0013] In some embodiments, the method further includes: in a corresponding model training iteration: (i) receiving (527) at least one training score associated with training of a corresponding personalized ML model performed by the FLDN (105) from a corresponding FLDN (105) of a first set and a corresponding FLDN (105) of a second set; and (ii) generating (530) training scores for corresponding aggregates of the first personalized ML model and the second personalized ML model based on the training scores received from the corresponding FLDN (105). In some embodiments, the method further includes: providing a computing device associated with an operator entity with the training scores of at least one aggregate associated with the first personalized ML model and the scores of at least one aggregate associated with the second personalized ML model.

[0014] In some embodiments, (i) a corresponding training score in at least one training score includes a performance score for performing the corresponding personalized task; and (ii) a corresponding training score in at least one aggregated training score includes a performance score for performing the aggregated personalized task. In some embodiments, (i) a corresponding training score also includes a credibility score for performing the corresponding personalized task; and (ii) a corresponding aggregated training score also includes a credibility score for performing the aggregated personalized task. In some embodiments, the method further includes: determining (506, 512) a plurality of personalized tasks based on user input received from a computing device associated with an operator entity, wherein the user input indicates at least one machine learning model use case.

[0015] In some embodiments, the method further includes: receiving (506) from a data storage source a plurality of historical personalization tasks based on at least one machine learning model use case, wherein the plurality of personalization tasks are determined based on the plurality of historical personalization tasks. In some embodiments, the method further includes: (i) generating (515) a mapping of at least one machine learning use case to a subset of the plurality of historical personalization tasks; and (ii) based on the mapping, generating (515) at least the following: (1) a first category comprising a first subset of a plurality of FLDNs (105), and (2) a second category comprising a second subset of a plurality of FLDNs (105), wherein the plurality of personalization tasks are further determined based on the first category and the second category.

[0016] In some embodiments, the method further includes: (i) providing (509) a plurality of historical personalization tasks to a computing device associated with an operator entity; (ii) receiving (512) a second user input from the computing device associated with the operator entity, wherein the second user input indicates at least one of the following: a filtered subset of the plurality of historical personalization tasks, or at least one additional personalization task; and (iii) further determining (506) a plurality of personalization tasks based on at least one of the following: a filtered subset of the plurality of historical personalization tasks, or at least one additional personalization task.

[0017] In some embodiments, the method further includes, between one or more model training iterations of a plurality of model training iterations: (i) determining (506) a third personalized task; (ii) determining (515) a third set of a plurality of FLDNs (105) associated with the third personalized task; (iii) generating (518) a third personalized machine learning model to perform the third personalized task; and (iv) providing (521) the third personalized machine learning model, or a subsequent trained iteration of the third personalized machine learning model, to the third set of FLDNs (105); wherein: (1) the corresponding FLDNs (105) of the third set train the third personalized machine learning model or a subsequent trained iteration of the third personalized machine learning model in the remaining subset of the plurality of model training iterations; and (2) in the final model training iteration, a personalized global machine learning model for the third personalized task is generated based on a plurality of corresponding local model parameters received from the corresponding FLDNs (105) of the third set during the remaining subset of the plurality of model training iterations. In some embodiments, one or more of the following items are determined based on one or more attributes or states of a plurality of FLDNs (105): a first set of a plurality of FLDNs (105), a second set of a plurality of FLDNs (105), or a third set of a plurality of FLDNs (105).

[0018] In some embodiments, the method further includes: (i) removing (506) at least one FLDN from a first set of a plurality of FLDNs (105) based on a change to one or more attributes or states of at least one FLDN; and (ii) adding at least one FLDN to a second set of a plurality of FLDNs (105) or a third set of a plurality of FLDNs (105) based on a change to one or more attributes or states of at least one FLDN. In some embodiments, one or more attributes or states of the plurality of FLDNs (105) include at least one of the following: computational resource availability, power availability, channel quality, data quality score, or data trust score. In some embodiments, one or more attributes or states of the plurality of FLDNs (105) include at least one of the following: location data, velocity data, trust score, or the number of available training data samples.

[0019] In some embodiments, the method further includes: (i) providing (515) a notification to a plurality of FLDNs (105) indicating a plurality of personalized tasks; (ii) receiving (512) a request from at least one of the plurality of FLDNs (105) to associate at least one FLDN with a first personalized task or a second personalized task; and (iii) determining (515) at least one of the following based on the request from at least one FLDN: a first set of the plurality of FLDNs (105) or a second set of the plurality of FLDNs (105).

[0020] In some embodiments, the method further includes: (i) receiving (512) a request from at least one FLDN to deassociate with a first set or a second set of a plurality of FLDNs (105); (ii) determining (506) that at least one training score associated with at least one FLDN and a personalized machine learning model trained by at least one FLDN meets a predetermined threshold; and (iii) in response to at least one training score meeting the predetermined threshold, providing (506) a request to at least one FLDN to retain at least one FLDN in the corresponding set of the plurality of FLDNs (105) for a plurality of training iterations, wherein the request to retain at least one FLDN indicates a fund retention incentive. In some embodiments, (i) the request for deassociation includes at least one deassociation factor; and (ii) the at least one deassociation factor is at least one of the following: computational overload, bandwidth limitation, or power limitation.

[0021] In various embodiments, as further described below, one or more operations of the above method are performed by a device having components for performing the one or more operations.For example, an apparatus may include: (i) components for determining (503) the following for at least one first personalization task and one second personalization task among a plurality of personalization tasks: (1) a first set of FLDNs (105) associated with the first personalization task from a plurality of federated learning distributed nodes (FLDNs (105)), and (2) a second set of FLDNs (105) associated with the second personalization task from a plurality of FLDNs (105); (ii) components for generating (518) the following: a first personalization machine learning (ML) model for performing the first personalization task, and a second personalization ML model for performing the second personalization task; and (iii) components for performing (521) A component for training a plurality of model training iterations for a first personalized ML model and a second personalized ML model, the plurality of model training iterations comprising: (1) providing (521) the component of the first personalized ML model to the corresponding FLDN (105) of the first set and (B) providing (521) the component of the second personalized ML model to the corresponding FLDN (105) of the second set in a corresponding initial model training iteration for the first personalized ML model and the second personalized ML model; (2) providing (521) the component of the second personalized ML model to the corresponding FLDN (105) of the first set in a corresponding subsequent model training iteration for the first personalized ML model and the second personalized ML model. 5) A component for providing (521) subsequent trained iterations of the first personalized ML model, and (B) a component for providing (521) subsequent trained iterations of the first personalized ML model to the corresponding FLDN (105) of the second set, wherein: in the corresponding model training iteration, the corresponding FLDN (105) of the first set and the corresponding FLDN (105) of the second set independently train the corresponding personalized ML model to perform the corresponding personalized task; (3) a component for receiving (524) local model parameters from the corresponding FLDN (105) in the corresponding model training iteration, wherein the local model parameters are associated with the training of the corresponding personalized ML model, which is performed by the first (3) The corresponding FLDN (105) of the first set or the second set is used; and (4) for (A) generating (530) components of the subsequent trained iterations of the first personalized ML model based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the first set; and (B) generating (530) components of the subsequent trained iterations of the second personalized ML model based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the second set, wherein: the output of the final iteration of each of the corresponding plurality of training iterations includes: a first personalized global ML model for performing the first personalized task and a second personalized global ML model for performing the second personalized task.In the same example, the device may embody additional aspects and / or include additional components for performing additional operations of the methods described above.

[0022] As further described below, in some embodiments, one or more operations of the above methods are performed by a device including at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the device to perform the one or more operations.For example, an apparatus may include at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to perform a function comprising: (i) determining (503) the following for at least one first personalized task and one second personalized task among a plurality of personalized tasks: (1) a set of first FLDNs (105) associated with the first personalized task from a plurality of federated learning distributed nodes (FLDNs) (105), and (2) a set of second FLDNs (105) associated with the second personalized task from a plurality of FLDNs (105); (ii) generating (518) a first personalized machine learning (MLM) for performing the first personalized task. (iii) performing (521) a series of model training iterations for training the first personalized ML model and the second personalized ML model, the series of model training iterations comprising: (1) in the corresponding initial model training iterations for the first personalized ML model and the second personalized ML model, (A) providing (521) the first personalized ML model to the corresponding FLDN (105) of the first set, and (B) providing (521) the second personalized ML model to the corresponding FLDN (105) of the second set; (2) in the corresponding subsequent model training iterations for the first personalized ML model and the second personalized ML model. In the training iterations, (A) (521) subsequent trained iterations of the first personalized ML model are provided to the corresponding FLDN (105) of the first set, and (B) (521) subsequent trained iterations of the first personalized ML model are provided to the corresponding FLDN (105) of the second set, wherein: in the corresponding model training iterations, the corresponding FLDN (105) of the first set and the corresponding FLDN (105) of the second set independently train the corresponding personalized ML model to perform the corresponding personalized task; (3) in the corresponding model training iterations, (524) local model parameters are received from the corresponding FLDN (105), wherein the local model parameters are related to the training of the corresponding personalized ML model. The training is associated with the corresponding FLDN (105) of the first set or the second set; and (4) (A) generating (530) subsequent trained iterations of the first personalized ML model based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the first set; and (B) generating (530) subsequent trained iterations of the second personalized ML model based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the second set, wherein: the output of the final iteration of each of the corresponding plurality of training iterations includes: a first personalized global ML model for performing the first personalized task and a second personalized global ML model for performing the second personalized task.In the same example, the device may also perform other operations and / or embody additional aspects of the methods described above.

[0023] In various embodiments, as further described below, a computer program product is provided herein, including at least one non-transitory computer-readable storage medium having computer-executable program code instructions stored thereon, the computer-executable program code instructions including program code instructions configured to perform one or more operations and / or embody additional aspects of the methods described above.For example, a non-transitory computer-readable medium includes program instructions stored thereon for performing at least the following: (i) for at least one first personalized task and a second personalized task among a plurality of personalized tasks, determining (503) the following: (1) a set of first FLDNs (105) associated with the first personalized task from a plurality of federated learning distributed nodes (FLDNs) (105), and (2) a set of second FLDNs (105) associated with the second personalized task from a plurality of FLDNs (105); (ii) generating (518) a first personalized machine learning (ML) model for performing the first personalized task, and a second personalized machine learning (ML) model for performing the second personalized task. Personalized ML models; and (iii) performing (521) corresponding multiple model training iterations for training the first personalized ML model and the second personalized ML model, the corresponding multiple model training iterations comprising: (1) in the corresponding initial model training iterations for the first personalized ML model and the second personalized ML model, (A) providing (521) the first personalized ML model to the corresponding FLDN (105) of the first set, and (B) providing (521) the second personalized ML model to the corresponding FLDN (105) of the second set; (2) in the corresponding subsequent model training iterations for the first personalized ML model and the second personalized ML model, (A) providing (521) the first personalized ML model to the corresponding FLDN (105) of the first set. (A) The FLDN (105) provides (521) subsequent trained iterations of the first personalized ML model to the corresponding FLDN (105) of the second set, and (B) provides (521) subsequent trained iterations of the first personalized ML model to the corresponding FLDN (105) of the second set, wherein: in the corresponding model training iteration, the corresponding FLDN (105) of the first set and the corresponding FLDN (105) of the second set independently train the corresponding personalized ML model to perform the corresponding personalized task; (3) in the corresponding model training iteration, receives (524) local model parameters from the corresponding FLDN (105), wherein the local model parameters are associated with the training of the corresponding personalized ML model, the training (4) (A) Based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the first set or the second set, (530) a subsequent trained iteration of the first personalized ML model is generated; and (B) Based on the aggregation of local model parameters associated with the corresponding FLDN (105) of the second set, (530) a subsequent trained iteration of the second personalized ML model is generated, wherein: the output of the final iteration of each of the corresponding plurality of training iterations includes: a first personalized global ML model for performing the first personalized task and a second personalized global ML model for performing the second personalized task.

[0024] In at least one embodiment, a method is provided comprising: (i) storing (603) training data in a memory (204) accessible to the device (105, 200); (ii) receiving (606) a request from a Federated Learning Aggregation (FLA) computing environment (101) for local environment data associated with the device; (iii) providing (609) the local environment data to the FLA computing environment (101), the local environment data including at least one of the following: computing resource availability, power availability, channel quality, data quality, or data credibility, wherein based on the local environment data, the FLA computing environment (101) associates the device with at least one of a plurality of personalized tasks; and (iv) performing a plurality of model training iterations, wherein a corresponding model training iteration of performing the plurality of model training iterations includes: (1) receiving (621) a personalized machine learning model for performing at least one personalized task from the FLA computing environment (101). (v) training (624) the personalized machine learning model or its subsequent trained iterations using training data to generate a plurality of first local model parameters; and (vi) providing (627) local model parameters to the FLA computing environment (101), wherein the FLA computing environment (101) generates a plurality of first local model parameters based on an aggregation of a plurality of first local model parameters and a plurality of second local model parameters received by the FLA computing environment (101) from a second device, the second device having trained the personalized machine learning model or its subsequent trained iterations; (v) generating (630) at least one training score for the personalized machine learning model or its subsequent iterations after at least a subset of the plurality of model training iterations; and (vi) providing at least one training score to the FLA computing environment (101).

[0025] In some embodiments, the FLA computing environment (101) includes a cross-domain management domain, and the means (105) for performing the method includes one or more separate management domains. In some embodiments, the FLA computing environment (101) includes a radio access network (RAN) management domain, and the means for performing the method includes a network node. In some embodiments, the FLA computing environment (101) includes a core management domain, and the means (105) for performing the method includes a network data analysis function (NWDAF). In some embodiments, the FLA computing environment (101) includes a central NWDAF, and the means (105) for performing the method includes a distributed NWDAF. In some embodiments, the FLA computing environment (101) includes a network node, and the means (105) for performing the method includes a user equipment (UE). In some embodiments, the FLA computing environment (101) includes a non-real-time RAN intelligent controller (RIC), and the means (105) for performing the method includes a real-time RIC. In some embodiments, the FLA computing environment (101) includes a real-time RIC, and the means for performing the steps of the method includes a network node. In some embodiments, the FLA computing environment (101) includes a first UE; and the apparatus (105) for performing the method includes a second UE.

[0026] In some embodiments, at least one training score includes a performance score for performing at least one personalized task. In some embodiments, at least one training score also includes a credibility score for performing at least one personalized task.

[0027] In some embodiments, the method further includes: (i) receiving (606) at least one local environment data threshold; (ii) determining (609) that a change to the local environment data satisfies at least one local environment data threshold; and (iii) providing (609) a notification to the FLA computing environment (101) indicating a change to the local environment data, wherein the notification causes the FLA computing environment (101) to remove the association between the device and at least one personalized task. In some embodiments, the method further includes: (i) providing (609) at least one attribute or state of the device to the FLA computing environment (101), wherein: (1) at least one attribute or state of the device includes at least one of the following: location data indicating the location of the device, the speed of the device, the trust score of the device, or the number of training data samples of the device; and (2) the FLA computing environment (101) associating the device with at least one personalized task based on at least one attribute or state of the device; (ii) receiving (606) at least one attribute or state threshold from the FLA computing environment (101); (iii) determining (609) that a change in at least one attribute or state of the device satisfies at least one attribute or state threshold; and (iv) providing (609) a notification to the FLA computing environment (101) indicating a change in at least one attribute or state of the device, wherein the notification causes the FLA computing environment (101) to remove the association between the device and at least one personalized task. In some embodiments, at least one attribute or state threshold includes a predetermined proximity to a geographic area, and a change in the at least one attribute or state indicates that the device's location exceeds the predetermined proximity to the geographic area. In some embodiments, at least one attribute or state threshold is a speed range associated with at least one task, and a change in the at least one attribute or state indicates that the device's speed is outside that speed range.

[0028] In some embodiments, at least one of the following further causes the FLA computing environment (101) to associate the device with an additional personalized task among a plurality of tasks: (i) a notification indicating a change to local environment data, or (ii) a notification indicating a change to at least one attribute or state. In some embodiments, the method further includes: (i) receiving (621) a second personalized machine learning model for performing the additional personalized task, or a subsequent trained iteration of the second personalized machine learning model, from the FLA computing environment (101); (ii) training (624) the second personalized machine learning model, or the subsequent iteration of the second personalized machine learning model, using training data to generate an additional plurality of local model parameters; and (iii) locally providing (627) the additional plurality of local model parameters to the FLA computing environment (101).

[0029] In some embodiments, the training data includes a plurality of local training samples. In some embodiments, the method further includes: (i) generating (624) synthetic data; and (ii) generating (624) additional local training samples based on the synthetic data, wherein the training data also includes the additional local training samples. In some embodiments, the method further includes: instructing (609) the number of local training samples in the training data to an FLA computing environment (101), wherein the FLA computing environment (101) also associates the device (105) with at least one of a plurality of personalized tasks based on the number of local training samples in the training data.

[0030] In some embodiments, the method further includes: (i) receiving (612) a notification from the FLA computing environment (101) indicating multiple personalized tasks; and (ii) providing (615) a request to the FLA computing environment (101) to associate the device with at least one personalized task, wherein the FLA computing environment (101) associates the device (105) with at least one personalized task at least in part based on the request. In some embodiments, the method further includes: (i) providing (618) a request to the FLA computing environment (101) to unassociate with at least one personalized task; and (ii) receiving (618) a request from the FLA computing environment (101) to maintain association with at least one personalized task and continue performing multiple training iterations, wherein: (1) the request to maintain at least one FLDN indicates at least one retention incentive; and (2) the FLA computing environment (101) provides the request to the device in response to at least one training score satisfying a predetermined threshold. In some embodiments, the request for unassociation includes at least one unassociation factor. In some embodiments, at least one decoupling factor is at least one of the following: computational overload, bandwidth limitation, or power limitation. In some embodiments, the decoupling request also includes the duration of device unavailability.

[0031] In various embodiments, as further described below, one or more operations of the above method are performed by a device having components for performing the one or more operations. For example, an apparatus may include: (i) a component for storing (603) training data in a memory (204) accessible to the apparatus (105, 200); (ii) a component for receiving (606) a request from a Federated Learning Aggregation (FLA) computing environment (101) for local environment data associated with the apparatus; (iii) a component for providing (609) local environment data to the FLA computing environment (101), the local environment data including at least one of the following: computing resource availability, power availability, channel quality, data quality, or data credibility, wherein based on the local environment data, the FLA computing environment (101) associates the apparatus with at least one of a plurality of personalized tasks; (iv) a component for performing a plurality of model training iterations, wherein the component for performing a corresponding model training iteration of the plurality of model training iterations includes: (1) receiving (621) a personalized machine learning model for performing at least one personalized task from the FLA computing environment (101), or the personalized machine learning model for performing at least one personalized task. (v) A component for training (624) a personalized machine learning model, or a subsequent iteration of the personalized machine learning model, using training data to generate a plurality of first local model parameters; and (vi) a component for locally providing (627) local model parameters to an FLA computing environment (101), wherein the FLA computing environment (101) generates a subsequent trained iteration of the personalized machine learning model based on an aggregation of a plurality of first local model parameters and a plurality of second local model parameters received by the FLA computing environment (101) from a second device, the second device having trained the personalized machine learning model, or a subsequent trained iteration of the personalized machine learning model; (v) a component for generating (630) at least one training score for the personalized machine learning model, or a subsequent iteration of the personalized machine learning model, after at least a subset of the plurality of model training iterations; and (vi) a component for providing (633) at least one training score to the FLA computing environment (101). In the same example, the device may embody additional aspects and / or include additional components for performing additional operations of the methods described above.

[0032] As further described below, in some embodiments, one or more operations of the above methods are performed by an apparatus including at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to perform the one or more operations. For example, an apparatus may include at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to perform a function including: (i) storing (603) training data in a memory (204) accessible to the apparatus (105, 200); (ii) receiving (606) a request from a Federated Learning Aggregation (FLA) computing environment (101) for local environment data associated with the apparatus; (iii) providing (609) local environment data to the FLA computing environment (101), the local environment data including at least one of the following: computing resource availability, power availability, channel quality, data quality, or data credibility, wherein based on the local environment data, the FLA computing environment (101) associates the apparatus with at least one of a plurality of personalized tasks; (iv) performing a plurality of model training iterations, wherein a corresponding model training iteration of performing the plurality of model training iterations includes: (1) receiving (621) instructions from the FLA computing environment (101) for performing (2) Performing at least one personalized machine learning model for a personalized task, or a subsequent trained iteration of the personalized machine learning model; (3) Training (624) the personalized machine learning model, or a subsequent iteration of the personalized machine learning model, using training data to generate a plurality of first local model parameters; and (4) Locally providing (627) local model parameters to an FLA computing environment (101), wherein the FLA computing environment (101) generates a subsequent trained iteration of the personalized machine learning model based on an aggregation of the plurality of first local model parameters and a plurality of second local model parameters received by the FLA computing environment (101) from a second device, the second device having trained the personalized machine learning model, or a subsequent trained iteration of the personalized machine learning model; (5) After at least a subset of the plurality of model training iterations, generating (630) at least one training score for the personalized machine learning model, or a subsequent iteration of the personalized machine learning model; and (66) Providing (633) at least one training score to the FLA computing environment (101). In the same example, the device may also perform other operations and / or embody additional aspects of the methods described above.

[0033] In various embodiments, as further described below, a computer program product is provided herein, including at least one non-transitory computer-readable storage medium having computer-executable program code instructions stored thereon, the computer-executable program code instructions including program code instructions configured to perform one or more operations and / or embody additional aspects of the methods described above. For example, a non-transitory computer-readable medium having program instructions stored thereon for performing at least the following: (i) storing (603) training data in a memory (204) accessible to the device (105, 200); (ii) receiving (606) a request from a Federated Learning Aggregation (FLA) computing environment (101) for local environment data associated with the device; (iii) providing (609) local environment data to the FLA computing environment (101), the local environment data including at least one of the following: computing resource availability, power availability, channel quality, data quality, or data credibility, wherein based on the local environment data, the FLA computing environment (101) associates the device with at least one of a plurality of personalization tasks; (iv) performing a plurality of model training iterations, wherein the corresponding model training iteration for performing the plurality of model training iterations includes: (1) receiving (621) personalization instructions from the FLA computing environment (101) for performing at least one personalization task. (2) Using training data to train (624) the personalized machine learning model, or subsequent trained iterations of the personalized machine learning model, to generate a plurality of first local model parameters; and (3) locally providing (627) local model parameters to an FLA computing environment (101), wherein the FLA computing environment (101) generates subsequent trained iterations of the personalized machine learning model based on an aggregation of the plurality of first local model parameters and a plurality of second local model parameters received by the FLA computing environment (101) from a second device, the second device having trained the personalized machine learning model, or subsequent trained iterations of the personalized machine learning model; (v) after at least a subset of the plurality of model training iterations, generating (630) at least one training score for the personalized machine learning model, or subsequent iterations of the personalized machine learning model; and (vi) providing (633) at least one training score to the FLA computing environment (101). In the same example, the program instructions may also be configured to perform additional operations and / or embody additional aspects of the methods described above. Attached Figure Description

[0034] Therefore, some example implementations of this disclosure have been described in general terms, and reference will now be made to the accompanying drawings, which are not necessarily drawn to scale, and in the drawings:

[0035] Figure 1The illustration shows an example of a communication system that can be implemented according to the exemplary embodiments of the present disclosure;

[0036] Figure 2 The illustration shows a block diagram of an apparatus that can be configured according to exemplary embodiments of the present disclosure;

[0037] Figures 3A-3B A cross-functional diagram is illustrated for Federated Multitask Learning (FMTL) according to an example embodiment of the present disclosure;

[0038] Figure 4 This is an example scheme for the management of FMTL according to an example embodiment of this disclosure;

[0039] Figure 5 This is an example flowchart of the FMTL training process from the perspective of a Federated Learning Aggregation (FLA) computing environment, according to an example embodiment of this disclosure; and

[0040] Figure 6 This is an example flowchart of the FMTL training process from the perspective of Federated Learning Distributed Nodes (FLDN) according to an example embodiment of this disclosure. Detailed Implementation

[0041] Some embodiments will now be described more fully below with reference to the accompanying drawings, in which some, but not all, of the embodiments are illustrated. In fact, various embodiments may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will satisfy applicable legal requirements. The same reference numerals refer to the same elements throughout the specification. As used herein, the terms “data,” “content,” “information,” and similar terms are used interchangeably to refer to data capable of being transmitted, received, and / or stored according to the described embodiments. Therefore, the use of any such terms should not be construed as limiting the spirit and scope of the embodiments.

[0042] Additionally, as used herein, the term "circuit system" means: (a) a hardware circuit implementation (e.g., an implementation in an analog circuit system and / or a digital circuit system); (b) a combination of circuitry and (multiple) computer program products, including software and / or firmware instructions stored on one or more computer-readable storage media, which work together to cause a device to perform one or more functions described herein; (c) a circuitry, such as, for example, (multiple) microprocessors or portions thereof, which requires software or firmware to operate, even if the software or firmware is not physically present. This definition of "circuit system" applies to all uses of the term herein, including in any claim. As another example, as used herein, the term "circuit system" also includes an implementation comprising one or more processors and / or portions thereof, along with accompanying software and / or firmware. As yet another example, as used herein, the term "circuit system" also includes, for example, a baseband integrated circuit or application processor integrated circuit for a mobile phone, or a similar integrated circuit in a server, cellular network device, other network device (such as a core network device), or a field-programmable gate array, and / or other computing device.

[0043] The term “comprising” means “including, but not limited to” and should be interpreted in the manner in which it is typically used in the patent content. The use of broader terms such as “comprising,” “including,” and “having” should be understood to support narrower terms such as “consisting of,” “mainly consisting of,” and “substantially comprising.” Furthermore, the terms “comprising” and “including,” and variations thereof, as used in the Detailed Description or Claims, are intended to be inclusive in a manner similar to “comprising.”

[0044] The phrases “in one embodiment,” “according to one embodiment,” “in some embodiments,” “in various embodiments,” etc., generally refer to the fact that a particular feature, structure, or characteristic following the phrase may be included in at least one embodiment of this disclosure, but not necessarily in all embodiments of this disclosure. Therefore, the particular feature, structure, or characteristic may be included in more than one embodiment of this disclosure, such that these phrases do not necessarily refer to the same embodiment.

[0045] As used herein, the terms “example,” “exemplary,” etc., are used to mean “serving as an example, instance, or illustration.” Any implementation, aspect, or design described herein as an “example” or “exemplary” should not necessarily be construed as being preferred or superior to other implementations, aspects, or designs. Rather, the use of the terms “example,” “exemplary,” etc., is intended to present concepts in a concrete manner.

[0046] If this specification indicates that a component or feature is "may", "can", "should", "will", "preferably", "may", "usually", "optionally", "for example", "often", or "maybe" (or other such expressions) included or has the characteristic, then that particular component or feature is not required to include or have that characteristic. In some embodiments, the component or feature may be optionally included or may be excluded.

[0047] As used herein, the term "computer-readable medium" refers to signals, non-transitory computer-readable media, etc. The term "non-transitory computer-readable medium" refers to non-transitory storage hardware, non-transitory storage devices, or non-transitory computer system memory that can be accessed by a controller, microcontroller, computing system, or module of a computing system to encode computer-executable instructions or software programs thereon. A non-transitory "computer-readable medium" can be accessed by a computing system or module of a computing system to retrieve and / or execute computer-executable instructions or software programs encoded on that medium. Examples of non-transitory computer-readable media include, but are not limited to, one or more types of hardware memory, non-transitory tangible media (e.g., one or more magnetic storage disks, one or more optical disks, one or more USB flash drives), computer system memory, or random access memory (such as DRAM, SRAM, EDO RAM), etc.

[0048] As used herein, the term "personalization task" refers to the action of generating or implementing a machine learning model architecture that satisfies a use case for network automation relative to different local datasets or groups of local datasets. For example, a personalization task may include generating a machine learning model architecture for performing network automation operations at one or more FLDNs such that the machine learning model architecture limits the number of radio link failures (RLFs) experienced by network services and devices to a threshold limit (e.g., different personalization tasks for this use case may be associated with different values ​​for that threshold limit).

[0049] As another example, personalization tasks could include generating machine learning model architectures for FLDNs that are associated with varying levels of data heterogeneity. Use cases could include scenarios where different FLDN groups comprise different dataset sizes (e.g., due to differences in data sampling rates). Personalization tasks could be defined to perform network automation services for each FLDN group. Existing techniques for FMTL training generate corresponding personalized global machine learning models for each FLDN group to perform network automation.

[0050] As another example, personalization tasks may include generating machine learning model architectures for FLDNs associated with varying levels of system heterogeneity. Use cases may include scenarios where different FLDN groups are associated with different levels of compute and / or communication capabilities. Existing techniques for FMTL training can be implemented to generate a corresponding personalized global machine learning model for each FLDN group, which is used to perform network automation services. The operational entity may provide the gNodeB (e.g., embodying an FLA aggregation environment as described herein) with an analysis indicating the possible compute and communication capabilities for the UE within the cell coverage area. The gNodeB can execute this FMTL training technique to determine different categories of UEs based on compute and / or communication capabilities. The gNodeB can generate a corresponding personalized global machine learning model for each category of UE, which is used to perform personalized network automation operations.

[0051] As another example, the personalization task may include generating machine learning model architectures for FLDNs with varying levels of objective heterogeneity. Use cases may include scenarios where different FLDN groups leverage machine learning model architectures to support different objectives. For example, a first FLDN group may utilize a federated learning machine learning model for mobility use cases, and a second FLDN group may utilize an FL machine learning model for supporting energy efficiency. This technique for FMTL training can be executed to generate a first personalized global machine learning model to satisfy the objectives of the first FLDN group, and a second personalized global machine learning model to satisfy the objectives of the second FLDN group.

[0052] As used herein, the term "personalized machine learning model" (including personalized global machine learning models and personalized local machine learning models) refers to a machine learning model configured to perform one or more tasks within a network to support network automation. In some embodiments, the satisfaction of the personalized task is based on (i) model performance (e.g., statistical accuracy for performing the personalized task) and / or (ii) model reliability (e.g., the degree to which the machine learning model architecture can be generalized to distinguish network elements associated with different attributes, states, interpretability requirements, etc.).

[0053] As used herein, the term "subsequent trained iterations" refers to a version of the personalized machine learning model generated based on local model parameters acquired by a set of Federated Learning Distributed Nodes (FLDNs) during at least one iteration of FMTL training. For example, a subsequent iteration of a personalized global machine learning model could be a personalized global machine learning model generated based on multiple personalized local machine learning models (e.g., defined by local model parameters) generated by corresponding FLDNs, where each corresponding FLDN uses its local training data to train previous iterations of the personalized global machine learning model. SUMMARY

[0054] In Federated Learning (FL), the goal is to generate a single global machine learning model that achieves “average” performance and credibility across all FL Distributed Nodes (FLDNs) participating in the FL training process. Different FLDNs, or groups of FLDNs, can utilize different non-IID (non-independent and identically distributed) local training datasets. Furthermore, different local training datasets can exhibit further heterogeneity due to missing, biased, or noisy features or labels in the local training data. Additionally, different FLDNs can exhibit different computational, storage, bandwidth, and / or power capabilities for performing model training. Due to the heterogeneity of FLDNs and local training data, the performance and / or credibility of the trained global machine learning model may be sufficient or ideal in the first set of participating FLDNs, but insufficient or poor in the second set of participating FLDNs. For example, a scenario may arise where, for a subset of participating FLDNs, the performance and credibility of the trained global machine learning model are worse than those of a model trained locally and independently by that subset of participating FLDNs. In various embodiments, the apparatus, method, and computer program product overcome such challenges using Federated Multi-Task Learning (FMTL) technology, which includes: determining personalized tasks for machine learning models and target use cases; determining different subsets of FLDNs associated with different personalized tasks; and generating personalized global machine learning models for each corresponding subset of FLDNs intended to participate in (e.g., where the FLDNs in the subsets may exhibit similar tasks, objectives, and / or characteristics).

[0055] In some embodiments, the apparatus for performing FMTL includes a Federated Learning Aggregation (FLA) computing environment and a plurality of FLDNs. In some embodiments, the FLA computing environment includes components for determining possible or necessary personalized tasks for the machine learning model and one or more use cases during one or more iterations of FMTL training. In some embodiments, the FLA computing environment includes components for determining the association between personalized tasks and candidate FLDNs during one or more iterations of FMTL training. For example, the FLA computing environment may include components for generating mappings between one or more FLDNs from a plurality of candidate FLDNs and one or more personalized tasks determined for the machine learning model.

[0056] In some embodiments, the FLA computing environment includes components for generating a personalized machine learning model for performing a personalized task (e.g., which may also be trained by an FLDN to perform the personalized task), and components for providing the personalized machine learning model to the corresponding FLDN for training. In some embodiments, the components for generating the personalized machine learning model include components for generating subsequent iterations of the personalized machine learning model based on local model parameters received from the FLDN and generated during the training of the machine learning model at each FLDN. For example, the FLA computing environment may aggregate the received local model parameters to generate subsequent iterations of the personalized machine learning model (also referred to herein as a personalized global machine learning model).

[0057] In some embodiments, the FLA computing environment includes components for determining the performance and / or confidence of a personalized machine learning model trained by a corresponding FLDN, or a personalized global machine learning model generated based on one or more model training iterations performed by a group of FLDNs. For example, the FLA computing environment may include components for generating one or more training scores for the personalized machine learning model based on local model parameters acquired during one or more training iterations. These training scores may include performance scores, confidence scores, etc. In some embodiments, the FLA computing environment includes components for receiving instructions from FLDNs intending to leave a group of FLDNs participating in the FMTL training process, or instructions from new FLDNs requesting to join that group. In some embodiments, the FLA computing environment includes components for predicting the impact of the departure of a participating FLDN, or the joining of another FLDN, on the performance and / or confidence of the personalized global machine learning model being trained during the ongoing FMTL training process.

[0058] In some embodiments, the apparatus includes one or more computing devices associated with an operator entity, wherein the FLA computing environment can receive data, commands, requests, user input, etc., from the operator entity via the computing devices. In some embodiments, the FLA computing environment includes components for receiving information from the operator entity's computing devices regarding intent and / or one or more use cases for a machine learning model. In some embodiments, the FLA computing environment includes components for generating or determining one or more personalized tasks for the machine learning model based on information from the operator entity. In some embodiments, the FLA computing environment includes components for providing instructions for personalized tasks to the computing devices for presentation to and viewing by the operator entity. In some embodiments, the FLA computing environment includes components for receiving a filtered subset of personalized tasks, one or more additional personalized tasks, etc., from the operator entity's computing devices. In some embodiments, the FLA computing environment includes components for generating new personalized machine learning models based on personalized tasks indicated by input from the computing devices, wherein the FLA computing environment can provide new personalized machine learning models for training on current or newly determined multiple FLDNs. In some embodiments, the FLA computing environment includes a component for pausing the training of a personalized machine learning model based on input from a computing device (e.g., input from an operator entity may enable the FLA computing environment to activate or deactivate a personalized task and initiate or pause the training of the corresponding personalized machine learning model).

[0059] In some embodiments, the FLA computing environment includes components for generating or retrieving one or more personalized machine learning models from a data store based on determined personalized tasks, wherein the respective personalized machine learning models are configured to perform specific personalized tasks. In some embodiments, the FLA computing environment includes components for generating reports indicating the following: multiple personalized tasks determined for the machine learning models; one or more use cases indicated for the machine learning models; one or more personalized machine learning models generated or retrieved based on multiple personalized tasks; one or more training scores for the respective personalized machine learning models; a personalized global machine learning model; or one or more participating FLDNs, etc. In some embodiments, the FLA computing environment includes components for providing reports to computing devices associated with an operator entity. Example Embodiments

[0060] like Figure 1As shown, a network environment 100 for performing Federated Multitask Learning (FMTL) according to various embodiments of the present disclosure is provided. In some embodiments, the network environment 100 includes a Federated Learning Aggregation (FLA) computing environment 103 communicating with a plurality of Federated Learning Distributed Nodes (FLDNs), which are depicted as FLDNs 105A, 105B, and 105C for illustrative and descriptive purposes. By way of example, the network environment 100 can be deployed in a radio access architecture based on LTE-Advanced (LTE-A) and / or New Radio (NR, 5G). However, this network environment can be deployed in other network architectures, including other communication networks, such as those developed in the future, such as sixth-generation (6G) networks, and any of a plurality of existing networks, including Universal Mobile Telecommunications System (UMTS) radio access networks (UTRAN, E-UTRAN, or NG-RAN), wireless local area networks (WLAN or WiFi). TM ), global microwave access interoperability (WiMAX) TM Bluetooth®, Personal Communication Services (PCS), ZigBee®, Wideband Code Division Multiple Access (WCDMA), systems using Ultra Wideband (UWB) technology, sensor networks, Mobile Ad Hoc Networks (MANET), and Internet Protocol Multimedia Subsystem (IMS), or any combination thereof.

[0061] In some embodiments, network environment 100 includes one or more computing devices 101 that communicate with FLA computing environment 103 to provide or receive data associated with performing an FMTL training process. In some embodiments, computing device 101 is associated with an operator entity. In some embodiments, computing device 101 includes components for receiving input (such as user input) from the operator entity. In some embodiments, this input includes or indicates one or more intents, use cases, personalization tasks, etc., for a machine learning model. In some embodiments, computing device 101 includes components for providing input to FLA computing environment 103. For example, computing device 101 may include components for indicating intents or use cases of a machine learning model to FLA computing environment 103 (e.g., these intents or use cases may then translate into personalization tasks for FMTL training of one or more personalized global machine learning models). In another example, computing device 101 may include components for providing FLA computing environment 103 with information for determining one or more personalization tasks. Information may include, for example, performance intent, credibility intent, one or more use cases, service level agreement, thresholds for model performance, thresholds for model credibility (e.g., including model interpretability), etc. In some embodiments, computing device 101 includes components for configuring a personalization task for which FMTL training will be performed. For example, computing device 101 may include components for filtering, activating, or deactivating a subset of proposed personalization tasks, which are determined by FLA computing environment 103 and provided to computing device 101 for presentation to the operational entity.

[0062] In some embodiments, such as Figure 3A As further described in section B, network environment 100 includes or embodies one or more data analysis functions 107 (including network data analysis function (NWDAF)). Figure 3A (as shown in B, NWDAF107A), and one or more Management Data Analysis Functions (MDAFS) (e.g., such as...) Figure 3AExamples include MDAF 107B shown in Figure B. In some embodiments, the data analytics function 107 embodies any number of computing devices and / or other systems embodied in hardware, software, firmware, etc., which (i) collect data from the FLDN, other network functions, Operations, Administration and Maintenance (OAM) systems, and other network elements, and / or (ii) generate analytics outputs based on the collected data. For example, the NWDAF may receive, analyze, and report collected data from other analytics functions, and generate data insights based on this data, such as compliance with service level parameters, the status of network device configurations and attributes, and changes thereto. In some embodiments, the FLA computing environment 103 may embody the NWDAF, such as by executing instructions via at least one processor for performing the functions of the FLA computing environment 103 as described herein. In another example, the MDAF may report network-related parameters, such as load levels and resource utilization, and generate predictions and recommendations for network management based on historical network behavior.

[0063] FLA computing environment 103 may include any number of computing devices and / or other systems embodied in hardware, software, firmware, etc., which communicate with computing device 101 and FLDNs (e.g., FLDN105A, 105B, 105C, etc.) via one or more networks, determine personalized tasks for machine learning models, determine the association between personalized tasks and FLDNs, generate personalized global machine learning models, determine training scores, and other metrics associated with the FMTL training process, etc. In some embodiments, FLA computing environment 103 includes or embodies a cross-domain management domain. For example, in some contexts, FLA computing environment 103 is a cross-domain management domain, including components for performing FMTL training processes for multiple FLDNs, which include or embody multiple separate management domains. In some embodiments, a management domain refers to a computing environment that includes any number of computing devices and / or other systems embodied in hardware, software, firmware, etc., which govern and manage network services, functions, and topologies for a specific domain (such as a Radio Area Network (RAN) domain, a transport domain, a fifth-generation (5G) core domain, etc.). In some embodiments, a cross-domain management domain refers to multiple computing devices and / or other systems embodied in hardware, software, firmware, etc., which aggregate data, analytics, and / or resource requirements across individual management domains and coordinate data sharing for network performance across all individual management domains or subsets thereof.

[0064] In some embodiments, the FLA computing environment 103 includes or embodies a Radio Access Network (RAN) management domain. For example, in some contexts, the FLA computing environment 103 is a RAN management domain, including components for performing an FMTL training process for multiple FLDNs, which include or embody multiple network nodes (e.g., gNodeBs, etc.). In some embodiments, the FLA computing environment 103 includes or embodies a core management domain. For example, in some contexts, the FLA computing environment 103 is a core management domain, including components for performing an FMTL training process for multiple FLDNs, which include or embody multiple Network Data Analysis Functions (NWDAFs). In some embodiments, the FLA computing environment 103 includes or embodies a core NWDAF. For example, in some contexts, the FLA computing environment 103 is a core NWDAF, including components for performing an FMTL training process for multiple FLDNs, which include or embody multiple distributed NWDAFs.

[0065] In some embodiments, the FLA computing environment 103 includes or embodies a network node. For example, in some contexts, the FLA computing environment 103 is a gNodeB or other suitable network node that includes components for performing an FMTL training process for multiple FLDNs that include or embody multiple User Equipments (UEs). In some embodiments, the FLA computing environment 103 includes or embodies a non-real-time RAN Intelligent Controller (RIC). For example, in some contexts, the FLA computing environment 103 is a non-real-time RIC, such as an rApp, that includes components for performing an FMTL training process for multiple FLDNs that include or embody multiple real-time RICs (e.g., multiple xApps, etc.). In some embodiments, the FLA computing environment 103 includes or embodies a real-time RIC. For example, in some contexts, the FLA computing environment 103 is a real-time RIC (such as an xApp) that includes components for performing an FMTL training process for multiple FLDNs that include or embody multiple network nodes (such as a gNodeB).

[0066] In some embodiments, the FLA computing environment 103 includes components for determining associations between various personalized tasks and FLDNs. For example, the FLA computing environment 103 may include components for determining, for at least a first personalized task and a second personalized task among a plurality of personalized tasks: (i) a first subset of a plurality of Federated Learning Distributed Nodes (FLDNs) associated with the first personalized task, and (ii) a second subset of a plurality of FLDNs associated with the second personalized task. In some embodiments, the FLA computing environment 103 includes components for generating an initial iteration or subsequent iteration of a personalized global machine learning model, wherein the corresponding personalized global machine learning model is configured to perform a specific personalized task. For example, the FLA computing environment 103 may include components for generating a first personalized machine learning model for performing the first personalized task and a second personalized machine learning model for performing the second personalized task, wherein the corresponding personalized task is associated with one or more use cases for the machine learning model.

[0067] In some embodiments, the FLA computing environment 103 includes components for performing FMTL training, which includes coordinating multiple model training iterations between the FLA computing environment 103 and various participating FLDN groups, wherein each FLDN group may be associated with a different personalized task and trains a corresponding personalized machine learning model locally. For example, the FLA computing environment 103 may include: (i) components for providing a first personalized machine learning model, or subsequent trained iterations of the first personalized machine learning model, to a first subset of the plurality of FLDNs; and (ii) components for providing a second personalized machine learning model, or subsequent trained iterations of the second personalized machine learning model, to a second subset of the plurality of FLDNs (e.g., wherein the respective FLDNs train their corresponding personalized machine learning models independently to perform the first or second personalized task). In some embodiments, the FLA computing environment includes components for receiving personalized local machine learning models independently and locally trained by the respective FLDNs from the FLDNs. In some embodiments, the personalized local machine learning model may include or embody multiple local model parameters. In one example, the FLA computing environment 103 may include a component for receiving, from a first subset and a second subset of multiple FLDNs, corresponding local model parameters associated with the training of a corresponding personalized global machine learning model performed by the FLDNs of each subset (e.g., where the corresponding subset of FLDNs is associated with FMTL training of a machine learning model for performing one of multiple personalized tasks).

[0068] In some embodiments, the FLA computing environment 103 includes a component for generating subsequent iterations of a personalized global machine learning model based on the aggregation of local model parameters from a plurality of trained personalized local machine learning models (e.g., where the corresponding personalized local machine learning models are generated by FLDNs of a version of the personalized global machine learning model provided during training). For example, the FLA computing environment 103 may include: (i) a component for generating subsequent trained iterations of a first personalized global machine learning model based on the aggregation of local model parameters received from a first subset of a plurality of FLDNs trained by an FMTL executing a first personalized global machine learning model; and (ii) a component for generating subsequent trained iterations of a second personalized machine learning model based on the aggregation of local model parameters received from a second subset of a plurality of FLDNs trained by an FMTL executing a second personalized global machine learning model.

[0069] In some embodiments, the FLA computing environment 103 includes components for receiving, in a corresponding model training iteration, one or more training scores from each participating FLDN for each corresponding personalized global machine learning model being trained. For example, the FLA computing environment 103 may include components for receiving at least one training score associated with training of the corresponding personalized machine learning model performed by the FLDN from corresponding FLDNs of a first subset and corresponding FLDNs of a second subset. In some embodiments, the training scores include performance scores, confidence scores, etc. In some embodiments, the FLA computing environment 103 includes components for generating training scores for one or more aggregates of personalized global machine learning models based on the corresponding training scores. For example, the FLA computing environment 103 may include components for generating aggregate performance scores based on performance scores from multiple local implementations, and generating aggregate confidence scores based on confidence scores from multiple local implementations. In some embodiments, the FLA computing environment 103 includes components for providing the training scores for one or more aggregates of one or more personalized global machine learning models to an operational entity. For example, the FLA computing environment 103 may include components for providing computing device 101 and / or one or more agent entities associated with the operator entity with performance scores and credibility scores of corresponding aggregates for the first personalized global machine learning model and the second personalized global machine learning model.

[0070] In some embodiments, the FLA computing environment 103 includes components for determining one or more personalized tasks based on machine learning model use cases received from a computing device 101 associated with an operator entity, or one or more agent entities associated with an operator entity. In some embodiments, the FLA computing environment 103 receives machine learning model use cases in the form of user input. In one example, machine learning model use cases include network performance improvements for mobility, network improvements for stability, model reliability improvements, etc. In some embodiments, the FLA computing environment 103 includes components for retrieving data from data storage (e.g., such as...). Figure 2 The memory 204 shown receives components for one or more historical personalization tasks based on one or more machine learning model use cases. In some embodiments, the FLA computing environment 103 includes components for determining one or more personalization tasks based at least in part on one or more historical personalization tasks.

[0071] In some embodiments, the FLA computing environment 103 includes components for generating mappings of one or more machine learning use cases to subsets of a plurality of historical personalization tasks. In some embodiments, the FLA computing environment 103 includes components for generating categories of candidate FLDNs based on the mapping. For example, the FLA computing environment 103 may include components for generating, based on the mapping,: (i) a first category comprising a first subset of the plurality of FLDNs, and (ii) a second category comprising a second subset of the plurality of FLDNs. In some embodiments, the FLA computing environment 103 includes components for determining one or more personalization tasks based on the categories of one or more generated FLDNs.

[0072] In some embodiments, the FLA computing environment 103 includes components for providing a plurality of historical personalization tasks to a computing device 101 associated with an operator entity. In some embodiments, the FLA computing environment 103 includes components for receiving a filtered subset of the plurality of historical personalization tasks, or one or more additional personalization tasks, from the operator entity (e.g., via computing device 101 or one or more agent entities). For example, the FLA computing environment 103 may include components for receiving user input indicating a subset of the plurality of historical personalization tasks or one or more additional personalization tasks. In some embodiments, the FLA computing environment 103 includes components for determining personalization tasks for FMTL training based on the received filtered subset of the plurality of historical personalization tasks and / or one or more additional personalization tasks.

[0073] In some embodiments, the FLA computing environment 103 includes components for determining one or more additional personalized tasks during or between iterations of FMTL training for other personalized tasks and personalized global machine learning models. In some embodiments, the FLA computing environment 103 includes components for generating one or more personalized global machine learning models for performing the additional determined personalized tasks. For example, the FLA computing environment 103 may include, between model iterations of FMTL training for a first personalized global machine learning model and a second personalized global machine learning model (which are used for the first and second personalized tasks), the following: (i) components for determining a third personalized task; (ii) components for determining a third subset of a plurality of FLDNs associated with the third personalized task; (iii) components for generating a third personalized machine learning model to perform the third personalized task; and (iv) components for providing the third personalized machine learning model or subsequent trained iterations of the third personalized machine learning model to the third subset of FLDNs. The corresponding FLDNs of the third subset may be used to train the third personalized machine learning model or subsequent trained iterations of the third personalized machine learning model in the remaining subsets of the plurality of model iterations. In the final model training iteration, the personalized global machine learning model for the third personalized task can be generated by the FLA computing environment 103 based on the corresponding local model parameters received from the corresponding FLDN of the third subset during the remaining subset of multiple model training iterations.

[0074] In some embodiments, the FLA computing environment 103 includes components for determining FLDNs for FMTL training based on one or more attributes or states of one or more FLDNs. In some embodiments, these attributes or states include computing resource availability, power availability, channel quality, data quality score, data trust score, location data, velocity data, trust score, number of available training data samples, etc. In some embodiments, the FLA computing environment 103 includes components for receiving indications of attributes or states, or changes thereof, from FLDNs or analysis functions associated with those FLDNs (e.g., NWDAF, MDAF, etc.). In some embodiments, the FLA computing environment 103 includes components for removing one or more participating FLDNs from additional FMTL training of the personalized global machine learning model based on changes to one or more attributes or states of the FLDNs. In some embodiments, the FLA computing environment 103 includes components for reassigning removed FLDNs to a second personalized task (e.g., for participation in FMTL training of a corresponding personalized global machine learning model) based on changes to one or more attributes or states of the FLDNs.

[0075] In some embodiments, the FLA computing environment 103 includes components for providing notifications to one or more FLDNs indicating multiple personalized tasks. In some embodiments, the FLA computing environment 103 includes components for receiving requests from one or more FLDNs to associate an FLDN with an indicated personalized task. In some embodiments, the FLA computing environment 103 includes components for determining participant FLDNs for FMTL training based on such requests from FLDNs that associate FLDNs with personalized tasks.

[0076] In some embodiments, the FLA computing environment 103 includes components for receiving a request from a participant FLDN to unassociate a group of participant FLDNs associated with ongoing FMTL training of a personalized global machine learning model. In some embodiments, the FLA computing environment 103 includes components for determining that the current or historical training score associated with the requesting FLDN and the personalized machine learning model trained by the requesting FLDN meets one or more predetermined thresholds. In some embodiments, the FLA computing environment 103 includes components for providing a request to the requesting FLDN to retain the FLDN in a corresponding group of FLDNs for multiple training iterations in response to at least one training score meeting a predetermined threshold. In some embodiments, the request to retain the FLDN indicates a fund retention incentive. This fund retention incentive may include an amount or rate of monetary compensation for continued participation in the FTTL training process. For example, an operator entity may provide a certain amount of money to one or more UEs to participate in (or continue participating in) FMTL training. In another example, an operator entity may allocate resources or services to a UE in exchange for UE participation in FMTL training. In some embodiments, the request to unassociate a participant FLDN includes one or more unassociation factors, including computational overload, bandwidth limitation, power limitation, etc. In some embodiments, a request to unassociate includes the duration for which the participant's FLDN is unavailable.

[0077] In some embodiments, the FLA computing environment 103 includes components for receiving, during FMTL training, a request to add an additional FLDN to a group of participant FLDNs associated with the training. In some embodiments, the FLA computing environment 103 includes components for determining the impact on the performance and / or confidence estimate of the corresponding personalized global machine learning model if the requested FLDN is added to the group of participant FLDNs used to train the personalized global machine learning model. In some embodiments, the FLA computing environment 103 includes components for adding the requested FLDN to the group of participant FLDNs in response to determining that the impact on the estimate of model performance and / or model confidence meets a predetermined threshold.

[0078] FLDN 105A, 105B, or 105C may include any number of computing devices(s) and / or other systems embodied in hardware, software, firmware, etc., that communicate with FLA computing environment 103 via one or more networks, provide data to FLA computing environment 103 (e.g., local environment information, machine learning model data, etc.), store local training data, receive and train personalized global machine learning models (e.g., including generating local model parameters), generate training scores for machine learning models, request association with or unassociation with personalized tasks, etc. In some embodiments, FLDN includes or embodies a separate management domain, network node, NWDAF, user equipment, real-time RIC, etc. For example, in some contexts, FLDN is a network node, such as gNodeB, including components for receiving and independently and locally training personalized global machine learning models using local training data. In another example, in some contexts, FLDN is a distributed NWDAF.

[0079] In some embodiments, the FLDN includes components for storing training data in a memory accessible to the FLDN. For example, the FLDN includes components for storing local training data that can be used by the FLDN to train one or more iterations of a personalized global machine learning model. In some embodiments, the FLDN includes components for receiving a request from the FLA computing environment 103 to provide local environment data associated with the FLDN. In some embodiments, the FLDN includes components for providing local environment data to the FLA computing environment 103. In some embodiments, the local environment data includes resource-related information, training data-related information, etc. In some embodiments, the local environment data includes one or more indications of: computing resource availability, power availability, channel quality, data quality, data reliability, etc. In some embodiments, the FLDN includes components for providing the FLA computing environment 103 with indications of one or more attributes or states of the FLDN, and / or indications of changes to one or more attributes or states of the FLDN. In some embodiments, the FLDN includes components for receiving one or more predetermined thresholds for detecting specific changes or magnitudes to one or more attributes or states of the FLDN. In some embodiments, the FLDN includes components for determining whether a change to an attribute or state of the FLDN (including a change to local environment information) satisfies a predetermined threshold, wherein a positive determination would cause the FLDN to provide an indication of the change to the FLA computing environment 103. In some embodiments, the attributes or state of the FLDN include FLDN location, FLDN velocity, trust score, number of available training data samples, etc. In some embodiments, the FLDN location is indicated by or determined based on location data associated with the FLDN, such as Global Positioning System (GPS) data. In some embodiments, the FLDN velocity is indicated by or determined based on location data, sensor data (such as data from an accelerometer), etc.

[0080] In some embodiments, the FLDN includes components for performing any number of iterations of FMTL training to train a personalized global machine learning model (e.g., including its subsequent trained iterations), thereby generating a personalized local machine learning model that includes local model parameters. For example, the FLDN may include components for training the personalized global machine learning model locally and independently using local training data to generate multiple local model parameters that may define the personalized local machine learning model. In some embodiments, the FLDN includes components for providing the personalized local machine learning model and / or corresponding local model parameters to the FLA computing environment 103. In some embodiments, the FLDN includes components for determining one or more training scores or personalized global (or local) machine learning models. In some embodiments, the one or more training scores include performance scores, confidence scores, etc. In some embodiments, the training scores include metadata indicating the associated personalized machine learning model and / or the local training data used to train the personalized machine learning model. For example, the metadata may include or indicate a machine learning model identifier, the number of local training data samples, the label or features of the local training data, the distribution of the local training data, noise in the local training data, any missing information or samples in the local training data, etc. In some embodiments, the FLDN includes components for providing one or more training scores to the FLA computing environment 103.

[0081] In some embodiments, the FLDN includes components for receiving notifications from the FLA computing environment 103 indicating a personalized task. In some embodiments, the FLDN includes components for providing the FLA computing environment 103 with a request associated with a specific personalized task, wherein the association of the FLDN with the specific personalized task results in the FLDN being included in the FMTL training of a personalized global machine learning model used to perform that specific task. In some embodiments, the FLDN determines the personalized task based on local environment information associated with the FLDN and / or other attributes or states associated with the FLDN, and requests the association for that personalized task.

[0082] In some embodiments, the FLDN includes components for providing a request to the FLA computing environment 103 for deassociation from the personalized task, the FLDN currently associated with the personalized task at the FLA computing environment 103. In some embodiments, the request for deassociation causes the FLA computing environment 103 to remove the FLDN from an additional FMTL training iteration for a personalized global machine learning model used to perform the corresponding personalized task. In some embodiments, the request for deassociation includes or indicates one or more deassociation factors, including computational overload, bandwidth limitations, etc. In some embodiments, the request for deassociation includes or indicates the duration for which the FLDN cannot participate in FMTL training for the currently associated personalized task and / or other personalized tasks. In some embodiments, the request for deassociation causes the FLA computing environment 103 to request the FLDN to remain associated with the personalized task and continue participating in FMTL training of the corresponding personalized global machine learning model. In some embodiments, the request to retain the FLDN indicates or includes one or more retention incentives. For example, the request to retain the FLDN may indicate an amount or other resources that will be provided to the FLDN in response to the FLDN remaining associated with the personalized task and continuing to participate in FMTL training of the personalized global machine learning model. In some embodiments, the FLDN includes a component for accepting or rejecting a request to the FLA computing environment 103 to retain the FLDN, wherein accepting or rejecting may cause the FLA computing environment 103 to retain or remove the FLDN as a participant in FMTL training for a corresponding personalized task.

[0083] Figure 2 An example apparatus 200 according to one embodiment is shown. In some embodiments, Figure 1 The Federated Learning Aggregation (FLA) computing environment shown and described herein is embodied in one or more devices 200. In some embodiments, Figure 1 The FLDN shown and described herein is embodied in device 200 (e.g., such embodiment of device 200 includes components for performing the operations and functions of the FLDN described herein). In one example, FLA computing environment 103 may be a first user equipment (UE) embodied in first device 200, which is configured to perform process 500 ( Figure 5 ); and FLDN 105 may be a second UE embodied as a second device 200, which is configured to perform process 600 ( Figure 6 ).

[0084] The device 200 may be an embodiment of a network device, or may be embodied in or otherwise associated with a network device. For example, the device may be embodied in or otherwise associated with an application function (AF), a multicast and broadcast session management function (MB-SMF), and / or a multicast and broadcast user plane function (MB-UPF).

[0085] Regardless of the device embodying apparatus 200, the apparatus may include processor 202, memory 204, and network interface 206. Apparatus 200 may be configured to perform the operations described herein. For example, processor 202, memory 204, and / or network interface 206A may perform the actions and processes described herein for implementing and managing personalized global machine learning models based on federated multi-task learning (FMTL) training. Although these components are described with respect to the execution of various functions, it should be understood that a particular implementation necessarily involves the use of specific hardware. It should also be understood that some of these components may include similar or common hardware. For example, two circuit sets may both utilize the same processor, network interface, storage medium, etc., to perform their associated functions, such that no duplicate hardware is required for each circuit set.

[0086] In some embodiments, processor 202 (and / or coprocessor or auxiliary processor or any other processing circuitry system otherwise associated with the processor) may communicate with memory 204 via a bus for transferring information between components of the device. Memory 204 is non-transitory and may include, for example, one or more volatile and / or non-volatile memories. In other words, memory 204 may be, for example, an electronic storage device (e.g., a non-transitory computer-readable storage medium). Memory 204 may be configured to store information, data, content, applications, instructions, etc., to enable the device to perform various functions according to the example embodiments disclosed herein.

[0087] Processor 202 can be embodied in a variety of different ways, for example, it can include one or more processing devices configured to execute independently. In some non-limiting embodiments, processor 202 can include one or more processors configured in series via a bus to enable independent execution, pipelining, and / or multithreading of instructions. The term "processor" can be understood to include single-core processors, multi-core processors, multiple processors within a device, and / or remote or "cloud" processors.

[0088] In some embodiments, processor 202 may be configured to execute instructions stored in memory 204 and / or otherwise accessible to processor 202 in a circuit system. In some embodiments, processor 202 may be configured to execute hard-coded functions. Thus, whether configured by hardware or software methods, or a combination thereof, processor 202 may represent an entity (e.g., physically embodied in a circuit system) capable of performing operations according to the embodiments disclosed herein while being configured accordingly. Alternatively, as another example, when processor 202 is embodied as an executor of software instructions, these instructions may be specifically configured to perform the algorithms and / or operations described herein when executed.

[0089] In some embodiments, device 200 may optionally include an input / output circuitry system that can communicate with processor 202 to provide output to a user and / or other entities, and in some embodiments, to receive input indications. The input / output circuitry system may include a user interface and may include a display, and may include a web user interface, mobile application, query-initiating computing device, self-service terminal, etc. In some embodiments, the input / output circuitry system may also include a keyboard, mouse, joystick, touchscreen, touch area, softkeys, microphone, speaker, or other input / output mechanisms. The processor and / or the user interface circuitry system including the processor may be configured to control one or more functions of one or more user interface elements via computer program instructions (e.g., software and / or firmware) stored in processor-accessible memory (e.g., memory 204, etc.).

[0090] Network interface 206 can be any component, such as a device or circuitry embodied in hardware or a combination of hardware and software, configured to receive and / or transmit data from / to the network and / or to / from any other device, circuitry, or module communicating with device 200. In this regard, network interface 206 may, for example, include network interfaces for enabling communication with wired or wireless communication networks, such as Application Functions (AF), Multicast and Broadcast Service Functions (MBSF), Multicast and Broadcast User Plane Functions (MB-UPF), and / or Multicast and Broadcast Session Management Functions (MB-SMF). For example, network interface 206 may include one or more network interface cards, antennas, buses, switches, routers, modems, and any other hardware and / or software supporting or suitable for enabling communication via a network. Additionally or alternatively, network interface 206 may include circuitry for interacting with one or more antennas to induce the transmission of signals via one or more antennas, or to process the reception of signals via one or more antennas.

[0091] Figures 3A-3B A cross-functional diagram for Federated Multitask Learning (FMTL) according to an example embodiment of the present disclosure is illustrated.

[0092] like Figure 3A As shown, the FLA computing environment 103 (i.e., the computing device 101 of the operating entity) can execute a preparation sequence 300A for FMTL. In some embodiments, the preparation sequence 300A is executed before the initiation of multiple training iterations in the FMTL training process. In some embodiments, the operating entity can interact directly with the FLA computing environment 103 via the computing device 101. In some embodiments, the operating entity can interact indirectly with the FLA computing environment 103 via one or more proxy entities, including intent and / or policy managers, network and service coordinators, machine learning coordinators, machine learning trust engines, etc.

[0093] In some embodiments, the entity operator, via computing device 101, indicates intents for use cases for a machine learning model to the FLA computing environment 103 in the form of one or more inputs provided to the FLA computing environment 103 (Box 1). These inputs may be user inputs, inputs defined by an automation entity (such as an agent entity), etc. In some embodiments, intents include performance intents, trust intents, etc. In some embodiments, intents are defined by a Service Level Agreement (SLA) between the operator entity and a vendor entity, which may be associated with the FLA computing environment 103, one or more FLDNs 105, etc. In some embodiments, the use case includes handover optimization, where FMTL training can be performed to train personalized global machine learning models for UEs associated with different intents (e.g., and the resulting different personalized tasks).

[0094] As an example, in a handover optimization use case, intents may include or embody network performance-related intents for different user equipments (UEs) associated with different rates, and thus generate different network performance thresholds, such as threshold limits for radio link failures (RLF). For example, a high-speed UE may be associated with a higher threshold limit for RLF, a medium-speed UE with a medium threshold limit for RLF, and a low-speed UE with a lower threshold limit for RLF. As another example, in a handover optimization use case, intents may include or embody network performance-related intents for UEs associated with different cell coverages, which can be identified using tracking area identifiers (TAIs). For example, a UE within cell coverage in a first TAI (TAI 1) may be associated with a higher threshold limit for RLF, a UE within cell coverage in a second TAI (TAI 2) may be associated with a medium threshold limit, and a UE within cell coverage in a third TAI (TAI 3) may be associated with a lower threshold limit for RLF.

[0095] As another example, in handover optimization use cases, intents can include or reflect intents related to machine learning trustworthiness. For instance, a UE within the cell coverage of TAI 1 can be associated with a higher threshold for machine learning model interpretability, a UE within the cell coverage of TAI 2 can be associated with a medium threshold for machine learning model interpretability, and a UE within the cell coverage of TAI 3 can be associated with a lower threshold for machine learning model interpretability.

[0096] In some embodiments, the FLA computing environment 103 retrieves one or more previously stored personalization tasks and associated metadata from memory 204 to leverage past experience to enhance the personalization task generation process and improve the reusability of identified personalization tasks (box 2). As described herein, in some embodiments, memory 204 includes or embodies one or more data stores or repositories that may be locally or remotely accessible to the FLA computing environment 103. In some embodiments, the FLA computing environment 103 determines one or more personalization tasks for FMTL based on information provided by the operator entity in box 1, for a given use case. In some embodiments, the FLA computing environment 103 generates one or more mappings of intents to FMTL personalization tasks, where the corresponding intents may be associated with a category of FLDN (such as a category of UE). For example, the FLA computing environment 103 may classify multiple UEs (e.g., FLDNs) based on their respective speeds (e.g., low speed, medium speed, high speed), where the FLA computing environment 103 may map each category of UE to a network performance-focused personalization task (e.g., RLF). As another example, the FLA computing environment 103 can classify multiple UEs based on their respective geographic locations (e.g., TAI 1, 2, 3, etc.), where the FLA computing environment 103 can map each category of UE to a personalized task focused on network performance. In another example, the FLA computing environment 103 can classify multiple UEs based on their respective geographic locations, where each category is mapped to a personalized task focused on machine learning credibility (e.g., machine learning model interpretability).

[0097] In some embodiments, the FLA computing environment 103 may report to the computing device 101 the determined personalized tasks for FMTL from blocks 2 and / or 3 for verification by the operating entity (e.g., filtering, activation, deactivation, and addition of personalized tasks) (block 4). In some embodiments, the FLA computing environment 103 stores the determined personalized tasks in memory 204. For example, the FLA computing environment 103 may store the determined personalized tasks in one or more data stores that are accessible to other FLA computing environments.

[0098] like Figure 3BAs shown, the FLA computing environment 103, the computing device 101 of the operating entity, multiple FLDNs 105, and optional NWDAF 107A and / or MDAF 107B can execute training sequences 300B for FMTL. In one embodiment, training sequences 300B, or subsets thereof, are repeatedly executed for multiple training iterations. In some embodiments, the operator entity filters a subset of personalization tasks from the overall personalization tasks reported at box 4 via computing device 101 and provides the filtered subset of personalization tasks to the FLA computing environment 103 (box 5A). In some embodiments, at box 5A, the operating entity provides the FLA computing environment 103 with recommendations for one or more new personalization tasks, in addition to or replacing the overall personalization tasks reported at box 4. In some embodiments, the operating entity configures (e.g., creates, updates, deletes, disables, enables, etc.) the filtered and / or recommended personalization tasks in the FLA computing environment 103 via one or more agent entities (box 5B).

[0099] In some embodiments, the FLA computing environment 103 provides requests to candidate FLDNs 105, one or more NWDAFs 107A, and / or one or more MDAFs 107B for reporting the local environment of the respective FLDN (box 6A). In some embodiments, the FLA computing environment 103 receives reports from FLDNs 105, NWDAFs 107A, and / or MDAFs 107B (box 6B) (e.g., these reports may be used as input in further determinations of personalized tasks, participant FLDNs, etc.). In some embodiments, requests for information provided to FLDNs 105 or analytics functions include a reporting deadline set by the FLA computing environment 103.

[0100] The FLA computing environment 103 may provide one or more requests between each iteration of FMTL training or within a subset of model training iterations. In some embodiments, the FLA computing environment 103 indicates reporting parameters for the requested report to the candidate FLDN 105, NWDAF 107A, and / or MDAF 107B. In some embodiments, the reporting parameters include computing resource availability, power availability, channel quality, data quality, data trustworthiness, etc. In some embodiments, the FLA computing environment 103 provides requests for additional information to the candidate FLDN 105 and / or one or more analysis functions (such as Network Data Analysis Function (NWDAF) or Management Data Analysis Function (MDAF)). In some embodiments, the additional information includes location information (e.g., the current location of the FLDN 105, its association with a tracking area identifier, etc.), velocity information (e.g., current velocity, acceleration, etc.), FLDN trust score, historical data, the number of available training samples accessible to the FLDN 105, etc.

[0101] In some embodiments, the FLA computing environment 103 subscribes to FLDN 105 and / or analytics functions for notification of changes to one or more attributes or states of the FLDN local environment and / or the corresponding FLDN 105. In some embodiments, the FLA computing environment 103 provides one or more thresholds or conditions to the FLDN 105 and / or analytics functions (such as NWDAF 107A or MDAF 107B). In some embodiments, the FLDN or analytics function determines whether a change to an aspect of the FLDN local environment, or a change to one or more states or attributes of the FLDN 105, satisfies the provided threshold. In some embodiments, in response to determining that a threshold is satisfied, the FLDN 105 or analytics function indicates to the FLA computing environment 103 the change in the FLDN local environment and / or the FLDN attribute or state. In one example, in response to FLDN 105, NWDAF 107A, or MDAF 107B determining that the availability of computing resources at FLDN 105 is less than two central processing units (CPUs) (e.g., or another suitable threshold), FLDN 105, NWDAF 107A, or MDAF 107B provides an indication of the change in computing resource availability to FLA environment 103. In another example, in response to FLDN 105, NWDAF 107A, or MDAF 107B determining that the location of FLDN 105 has changed from a first TAI to a second TAI, FLDN 105, NWDAF 107A, or MDAF 107B provides an indication of the change in TAI association to FLA environment 103. In another example, in response to FLDN 105, NWDAF 107A, or MDAF 107B determining that the availability of computing resources at FLDN 105 exceeds or falls below a speed threshold (e.g., 50 km / hr, 60 km / hr, or another suitable threshold), FLDN 105, NWDAF 107A, or MDAF 107B provides an indication of a change in the FLDN speed to FLA environment 103. In some embodiments, where a reporting deadline is specified and FLDN 105 or the analysis function determines that no threshold is met, FLDN or the analysis function may indicate a lack of change (“no change”) to FLA computing environment 103 (e.g., instead of sending a full report of the current FLDN local environment and / or attributes or status).

[0102] In some embodiments, based on the information collected in boxes 5 and / or 6A to B, the FLA computing environment 103 also determines new personalized tasks (box 7) in addition to the tasks identified in box 2. The FLA computing environment 103 can determine the new personalized tasks by classifying the FLDNs 105 based on the received information. In some embodiments, the FLA computing environment 103 can classify the FLDNs 105 based on the total number of their respective local training samples. For example, the FLA computing environment 103 can classify FLDNs with fewer than 10,000 samples into a first personalized task and FLDNs with more than 10,000 samples into a second personalized task. In some embodiments, the FLA computing environment 103 classifies the FLDNs based on their trust scores. For example, the FLA computing environment 103 can classify FLDNs with trust scores within a high threshold range into a first personalized task and FLDNs with trust scores within a low threshold range into a second personalized task.

[0103] In some embodiments, the FLA computing environment 103 stores the identified personalized tasks, the association between the personalized tasks and FLDNs, and the corresponding metadata in memory 204. For example, the FLA computing environment 103 may store the identified personalized tasks and corresponding metadata in a data repository (e.g., this data repository may be associated with analytics data repository functionality), allowing other FLA computing environments to utilize this information by directly querying the data repository.

[0104] In some embodiments, based on the information collected in boxes 4 through 6, the FLA computing environment 103 determines the association between candidate FLDNs and identified personalized tasks (e.g., also considering the intersection between personalized tasks identified in boxes 2 and 6) (box 8). In some embodiments, the FLA computing environment 103 provides FLDNs 105 with indications of personalized tasks and corresponding metadata information, enabling the FLA computing environment 103 to receive self-selections from FLDNs 105 to participate in the FMTL training of the corresponding personalized global machine learning model. For example, the FLA computing environment 103 may broadcast all identified personalized tasks and metadata information for each task to candidate FLDNs 105. FLDNs 105 may self-select the personalized tasks they wish to participate in and report their selections back to the FLA computing environment 103.

[0105] In some embodiments, the FLA computing environment 103 generates and / or retrieves from memory 204 a personalized global machine learning model mapped to the corresponding personalized task (box 9). Also at box 9, in some embodiments, the FLA computing environment 103 provides the personalized global machine learning model to the corresponding FLDN 105 as determined in step 7. In some embodiments, the corresponding FLDN 105 trains the personalized global machine learning model locally and independently to perform the corresponding personalized task, thereby generating a personalized local machine learning model. In some embodiments, in addition to the personalized global machine learning model, the FLA computing environment 103 may also include metadata information related to the personalized task selected for the FLDN 105. In some embodiments, the FLDN 105 uses such information to customize its corresponding local training to improve the performance and reliability of the trained personalized model. For example, based on the metadata information, the FLDN 105 may generate synthetic data to increase the number of training samples for local training, such as when the total number of available local training samples for the personalized task in the FLDN is determined.

[0106] In some embodiments, after an iteration of FMTL training at FLDN 105, FLA compute environment 103 receives a corresponding personalized local machine learning model from FLDN 105 (box 10), wherein the personalized local machine learning model may be provided in the form of local model parameters, one or more training scores (e.g., performance scores, confidence scores, etc.), metadata (e.g., the number of training samples, data quality indicators, etc.). For example, after training its corresponding local model, FLDN 105 may send its local model parameters back to FLA compute environment 103 within a reporting deadline set by FLA compute environment 103.

[0107] In some embodiments, the FLA computing environment 103 aggregates local model parameters received from the FLDN 105 to obtain subsequent iterations of the personalized global machine learning model corresponding to each personalized task (box 11). In some embodiments, the FLA computing environment 103 stores one or more versions of the personalized global machine learning model in memory 204. For example, the FLA computing environment 103 may store time-series records of versions of the personalized global machine learning model acquired throughout the iterations of the entire workflow 300B.

[0108] In some embodiments, once the personalized global machine learning model converges and / or after a predetermined number of model training iterations (e.g., 3, 5, or any suitable number), the FLA computing environment 103 receives the implemented personalized model performance and personalized model confidence from each participant FLDN 105 (box 12). In some embodiments, the implemented personalized model performance and personalized model confidence are provided in the form of one or more training scores, such as personalized local model performance scores and personalized local model confidence scores. In some embodiments, the FLA computing environment 103 generates averaged model performance and averaged model confidence for each personalized task by averaging the individual model performance and individual model confidence measurements reported by each corresponding FLDN 105.

[0109] In some embodiments, based on the average model performance and average model confidence generated for each personalized task, the FLA computing environment 103 can also calculate the “actual contribution” (box 12) of each participant FLDN 105 to the corresponding personalized global machine learning model. In some embodiments, according to Equation 1, the FLA computing environment 103 averages the personalized model performance / confidence of all participants FLDN 105 within the personalized task and generates a contribution score for each participant FLDN 105. Score for each FLDN = (Equation 1)

[0110] In some embodiments, in response to determining that a particular FLDN 105 has a low score (e.g., below a predetermined threshold), the FLA computing environment 103 initiates a troubleshooting process by requesting additional information from the corresponding FLDN 105 and analytics functions (e.g., NWDAF 107A, MDAF 107B, etc.). In some embodiments, based on the score and / or the new information, the FLA computing environment 103 (i) reassigns the participant FLDN 105 to a new or alternative existing personalized task for model training iterations, or (ii) removes the participant FLDN 105 from the candidate FLDN list, thereby excluding the FLDN 105 from subsequent model training iterations. In some embodiments, without requiring the participant FLDN to send training scores for performance and credibility, the FLA computing environment 103 determines the individual implemented personalized model performance and implemented personalized model credibility of the participant FLDN. For example, in some embodiments, where the FLA computing environment 103 is able to access a general test dataset for use cases, or receive a general test dataset from the computing device 101 of the operator entity, to perform such evaluations on the received personalized global machine learning model, the FLA computing environment 103 can independently determine the individual contribution (e.g., performance and reliability) of the FLDN 105.

[0111] In some embodiments, after model convergence or a predetermined number of model training iterations, the FLA computing environment 103 generates and provides to the operating entity the corresponding average model performance and average model confidence for the final iteration of each personalized global machine learning model (box 13). In some embodiments, the FLA computing environment 103 provides the averaged model information to the operating entity via transmission to computing device 101, one or more agent entities, etc.

[0112] Figure 4 An example scheme 400 for managing Federated Multitask Learning (FMTL) according to an example embodiment of the present disclosure is illustrated. In some embodiments, scheme 400 includes an FLA consumer 401, an FLA computing environment 403, and a plurality of FLDN producers 405. The FLA consumer 401 represents the operator entity described herein, including computing devices and / or agent entities associated with the operator entity. In some embodiments, the plurality of FLDN producers 405 represent Federated Learning Distributed Nodes (FLDNs) described herein. In some embodiments, as... Figure 4 As shown, the FLA computing environment 403 communicates with the FLA consumer 401 and multiple FLDN producers to coordinate and manage the FMTL training of personalized global machine learning models for performing corresponding personalized tasks (e.g., corresponding to one or more machine learning use cases).

[0113] In some embodiments of FMTL, FLDN producer 405 reports information about the availability of computational and radio resources at each model training iteration. In this approach, participant FLDNs can be selected first for each training iteration. For example, candidate FLDNs can report their computational resource availability for the training task to FLA computing environment 403 (such as a federated learning (FL) server). FLA computing environment 403 can then select participant FLDNs based on reports from candidate FLDNs and other conditions, such as radio channel conditions. Previous approaches for federated learning may not be able to collect and consider information on FLDNs available for training data. In some embodiments, approach 400 includes enabling FLA computing environment 403 to collect information about the distribution of training data, data quality, and local datasets of candidate FLDNs in order to detect potential biases and / or anomalous behavior and react promptly. Such embodiments can address the potential challenges of generating FLDN selections for federated learning; however, such approaches may not be able to achieve high average statistical accuracy of the output machine learning model across heterogeneous FLDNs and corresponding local environments. Such challenges in federated learning can be overcome by implementing federated multitasking learning (FMTL) technology, including a novel management mechanism for FMTL training developed by the inventors.

[0114] In some embodiments, scheme 400 includes configuring and / or enabling FLA consumer 401 to provide information about the intent of one or more machine learning models for use cases to assist FLA computing environment 403 in determining personalization tasks for FMTL. In some embodiments, FLA consumer 401 further includes components for discovering supported / available personalization tasks from FLA computing environment 403. In some embodiments, FLA consumer 401 includes components for configuring the activation and deactivation of personalization tasks at FLA computing environment 403, thereby controlling the personalization tasks for which FMTL training is performed. In some embodiments, FLA consumer 401 includes components for receiving reports from FLA computing environment 403 related to the personalization global model.

[0115] In some embodiments, scheme 400 includes configuring and / or enabling the FLA computing environment 403 to determine possible or necessary personalization tasks for use cases during one or more iterations of the FMTL training process. In some embodiments, the FLA computing environment 403 includes components for discovering mappings between FLDN producers 405 (e.g., candidate FLDNs) and their associated personalization tasks during one or more iterations of the FMTL training process. In some embodiments, the FLA computing environment 403 includes components for determining the performance and / or credibility of (multiple) personalized global models from each participant FLDN producer 405 or group of FLDN producers (e.g., for which average performance and / or credibility can be determined by the FLA computing environment 403). In some embodiments, the FLA computing environment 403 includes components for ensuring that the performance and credibility of (multiple) personalized global models are not adversely affected beyond a threshold level if one or more participant FLDN producers 405 leave the FMTL training process, or if one or more candidate FLDN producers 405 determine to join the FMTL training process.

[0116] For reference Figure 5 The diagram shows an example flowchart of a Federated Multi-Task Learning (FMTL) training process 500, which can be performed by, for example... Figure 2 The device 200 performs the device operation. The device 200 can embody Figure 1 The FLA computing environment 103 shown includes components for performing various functions and operations as described herein.

[0117] In some embodiments, at block 503, the means for performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for receiving intents for one or more use cases for a machine learning model from a computing device of an operator entity. In one example, the intent may include improving network performance in device mobility. In another example, the intent may include improving network stability. In yet another example, the intent may include improving the reliability of models used for one or more global machine learning networks associated with providing automated network management. In some embodiments, the use case refers to improvements in corresponding network aspects for a specific category of network devices. For example, a use case for improving network performance may include generating a corresponding global machine learning model to minimize radio link failures (RLFs) for different categories of network devices, under varying threshold constraints, based on device location (e.g., the location of a device within cell coverage of a specific tracking area identifier) ​​or device speed. In another example, a use case for improving model reliability may include generating corresponding global machine learning models with low, medium, and high levels of interpretability for devices within different cell coverage areas.

[0118] In some embodiments, at block 506, the apparatus for performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for determining multiple personalized tasks based on the intent and / or use cases of block 503, wherein the multiple personalized tasks include at least a first personalized task and a second personalized task. In some embodiments, apparatus 200 determines multiple personalized tasks based on one or more historical personalized tasks. In some embodiments, apparatus 200 determines multiple personalized tasks based on local environment data and / or other attributes or states of the corresponding candidate federated learning distributed nodes (FLDNs) (e.g., where such information may be provided by the FLDN or one or more network analysis functions). In some embodiments, apparatus 200 determines one or more personalized tasks based on one or more modifications to the personalized tasks acquired at block 512. For example, apparatus 200 may remove one or more personalized tasks based on a filtered subset of multiple tasks received from an operating entity. In another example, apparatus 200 may add one or more personalized tasks received from or indicated by an operating entity. In some embodiments, apparatus 200 determines one or more additional personalized tasks based on data acquired during one or more model training iterations. For example, device 200 can determine one or more additional personalized tasks based on detected changes to local environment data and / or other attributes or states of the FLDN (e.g., indications received from the FLDN or one or more network analysis functions).

[0119] In some embodiments, at block 509, the means for performing process 500 may optionally include components such as processor 202, memory 204, network interface 206, etc., for providing a notification to a computing device of an operating entity indicating a plurality of personalized tasks. In some embodiments, the notification includes a list of personalized tasks. In some embodiments, the notification includes metadata associated with the determination of personalized tasks, including use cases, categories of one or more FLDNs, local environment data associated with one or more FLDNs, one or more attributes or states of the corresponding FLDN, etc.

[0120] In some embodiments, at block 512, the means of executing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for receiving one or more modifications to a plurality of personalized tasks from a computing device of an operating entity. For example, the means 200 may receive from computing device 101 a filtered subset of the plurality of personalized tasks, activation or deactivation of one or more personalized tasks (e.g., initiating, pausing, or resuming an associated FMTL training process), or instructions for one or more additional personalized tasks to be included in FMTL training.

[0121] In some embodiments, at block 515, the apparatus for performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for determining the association between candidate FLDNs and each personalization task. In some embodiments, apparatus 200 determines the association based on local environment data and / or other attributes or states of the corresponding candidate FLDNs. In some embodiments, apparatus 200 adjusts the association of participant FLDNs based on changes to local environment data and / or other attributes or states of FLDNs. For example, apparatus 200 may reclassify an FLDN from a first personalization task to a second personalization task based on changes in FLDN speed or TAI association. In some embodiments, apparatus 200 adjusts the association of participant FLDNs based on the average performance score determined at block 530. For example, apparatus 200 may remove the association between a participant FLDN and a personalization task based on determining that the average performance score of a participant FLDN fails to meet a predetermined threshold. In some embodiments, apparatus 200 determines participant FLDNs for personalization tasks based on historical training scores associated with FLDNs. In some embodiments, apparatus 200 determines participant FLDNs for personalization tasks in response to receiving a request from candidate FLDNs to participate in FMTL training.

[0122] In some embodiments, at block 518, the apparatus for performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for generating a corresponding personalized global machine learning model for each of a plurality of personalized tasks. For example, apparatus 200 may generate or retrieve from memory 204 a first personalized global machine learning model for performing a first personalized task and a second personalized global machine learning model for performing a second personalized task.

[0123] In some embodiments, at block 521, the means of performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for providing a corresponding personalized global machine learning model (e.g., including its initial or subsequent trained iterations) to the corresponding FLDN set for training. At or after block 521, the corresponding FLDN independently and locally trains its corresponding personalized machine learning model to perform the corresponding personalized task, wherein the personalized machine learning model may be an initially received version of the personalized machine learning model or a subsequent trained iteration of the personalized machine learning model.

[0124] In some embodiments, at block 524, the means of executing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for receiving a corresponding trained personalized local machine learning model and / or associated data from the corresponding FLDN set. In some embodiments, the components for receiving the trained personalized local machine learning model include components for receiving local model parameters that define the corresponding personalized local machine learning model trained by the participant FLDN. In some embodiments, the means of executing process 500 includes components such as memory 204 for storing one or more versions of a personalized global machine learning model (including personalized local machine learning models associated with the training of the personalized global machine learning model). For example, during one or more iterations of FMTL training, the means 200 may store the received local model parameters in memory 204, which may include data storage accessible to other means of executing the FMTL training process (e.g., other FLA computing environments).

[0125] In some embodiments, at block 527, the means of executing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for receiving one or more training scores for a corresponding personalized local machine learning model. In one example, a first set of FLDNs trains a first personalized global machine learning model for performing a first personalized task, and a second set of FLDNs trains a second personalized global machine learning model for performing a second personalized task. The means 200 can receive confidence scores and performance scores associated with locally trained versions of the personalized global machine learning model (e.g., also referred to as personalized local machine learning models) from the corresponding FLDNs of the first and second sets. In some embodiments, the means of executing process 500 includes components such as memory 204 for storing training scores. For example, the means 200 can store training scores associated with a stored version of the corresponding personalized global machine learning model in memory 204, thereby providing a time-series record of personalized global machine learning model versions and corresponding training scores. In some embodiments, block 527 occurs in each iteration of FMTL training or after a predetermined number of model training iterations (e.g., 3, 5, or any suitable number of model training iterations). In some embodiments, in response to the participant FLDN and / or the apparatus of execution process 500 determining that the personalized global machine learning model has achieved convergence (e.g., the model error has stabilized within the error range and / or the model performance is close to a threshold or limit), box 527 occurs.

[0126] In some embodiments, at block 530, the apparatus for performing process 500 further includes components such as processor 202, memory 204, network interface 206, etc., for generating an average training score for each personalized task by averaging the individual training scores reported by each corresponding FLDN. For example, apparatus 200 can generate an average model performance score and an average model confidence score for each personalized task by averaging the individual model performance and individual model confidence measurements reported by each corresponding FLDN.

[0127] In some embodiments, at block 530, the apparatus for performing process 500 further includes components such as processor 202, memory 204, network interface 206, etc., for generating corresponding average performance scores for multiple participating FLDNs according to Equation 1. In some embodiments, the average performance scores for FLDNs are used at block 515 to modify the participating FLDNs. For example, at block 515, the apparatus for performing process 500 may remove or reclassify FLDNs based on a comparison of the FLDN's performance score with a minimum performance threshold.

[0128] In some embodiments, the apparatus of execution process 500 uses the average performance score of the participant's FLDN as input to determine the association between the personalization task and the FLDN at block 515. For example, at block 515, apparatus 200 may compare the average performance score of the participant's FLDN for the first personalization task with a predetermined minimum performance threshold. In response to determining that the average performance score fails to meet the minimum performance threshold, apparatus 200 may exclude the FLDN from further iterations of FMTL training for the first personalization task and, possibly, associate the FLDN with FMTL training for the second personalization task.

[0129] In some embodiments, at block 533, the means of executing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for generating subsequent iterations of the personalized global machine learning model. In some embodiments, the means of executing process 500 includes components for generating subsequent iterations of the personalized global machine learning model by aggregating local model parameters (e.g., as reported to the means by the corresponding participant FLDN) of the personalized local machine learning model for the corresponding personalized task. Subsequent iterations of the personalized global machine learning model can be provided to the corresponding participant FLDN for additional FMTL training. For example, after block 533, process 500 can continue to execute subsequent iterations of block 521, including providing the corresponding set of participant FLDNs with subsequent iterations of the personalized global machine learning model.

[0130] In some embodiments, at block 536, the means for executing flow 500 includes components such as processor 202, memory 204, network interface 206, etc., for providing reports to an operational entity (e.g., by providing reports to a computing device associated with the operational entity). In some embodiments, the report includes one or more average training scores for the final iteration of a corresponding personalized global machine learning model for each personalized task. For example, the report may include an average performance score and an average confidence score for the final iteration of the personalized global machine learning model. In some embodiments, the report includes data associated with participant FLDNs, including average performance scores, the number of participant FLDNs, the category of participant FLDNs, etc.

[0131] For reference Figure 6 The diagram shows an example flowchart of a Federated Multi-Task Learning (FMTL) training process 600, which can be performed by a device, such as... Figure 2 Device 200. Device 200 can embody Figure 1 The FLDN 105 shown includes components for performing various functions and operations as described herein. For example, the means for performing process 600 may be a user equipment configured to perform the various functions and operations of the corresponding FLDN described herein.

[0132] In some embodiments, at block 603, the means of executing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for storing training data for performing local training of a machine learning model. For example, the FLDN of executing process 600 may acquire and store training data for training a machine learning model to generate an optimal machine learning model architecture for various use cases, such as improving network stability, energy saving, mobility handover, and / or network performance for various categories of network devices (e.g., speed-based categories, cell coverage area-based categories, etc.). In some embodiments, the training data includes measurements for network device location, channel state information (CSI) data, allocated power, etc.

[0133] In some embodiments, at block 606, the means for performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for receiving a request to provide local environment data to the Federated Learning Aggregation (FLA) computing environment. In one example, the FLDN may receive a request from the FLA computing environment 103 to report resource-related information and training data-related information. In some embodiments, the request indicates one or more reporting parameters that indicate the requested information. For example, the reporting parameters may instruct the FLDN to report computing resource availability, power availability, channel quality, data quality, data trustworthiness, etc. In some embodiments, the request includes reporting parameters for other attributes or states of the FLDN. For example, the request may include reporting parameters instructing the FLDN to report location information, velocity information, trust score, historical information, the number of available training samples, etc.

[0134] In some embodiments, at block 609, the means for performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for providing local environment data (and possibly other requested information) to the FLA computing environment 103. For example, the FLDN may provide the FLA computing environment 103 with indications of computing resource availability, power availability, channel quality, data quality, data reliability, etc. In another example, the FLDN may provide current speed, location (e.g., association with TAI), number of local training data samples, etc. In some embodiments, based on the received information, the FLA computing environment 103 determines one or more personalized tasks, and the association between the FLDN and the personalized tasks, for the purpose of FMTL training.

[0135] In some embodiments, at block 612, the apparatus for executing process 600 may optionally include components such as processor 202, memory 204, network interface 206, etc., for receiving notifications of one or more personalized tasks, for which the apparatus may participate in FMTL training. For example, FLDN can receive notifications instructing multiple personalized tasks from the FLA computing environment.

[0136] In some embodiments, at block 615, the means of executing process 500 may optionally include components such as processor 202, memory 204, network interface 206, etc., for providing the FLA computing environment with a request for a personalized task. For example, the FLDN may provide the FLA computing environment with a request to associate the FLDN with one of a plurality of personalized tasks reported to the FLDN, wherein the association with the personalized task includes the FLDN's participation in FMTL training of a personalized global machine learning model for performing the task.

[0137] In some embodiments, at block 618, the apparatus of execution process 600 optionally includes components such as processor 202, memory 204, network interface 206, etc., for providing a request to the FLA computing environment for deassociating the device from the personalized task, the device currently associated with the personalized task. For example, an FLDN may provide a request to the FLA computing environment for deassociating from the currently associated personalized task, where deassociation may result in the FLDN being excluded as a participant from additional FMTL training for the personalized task. In some embodiments, the request for deassociation includes one or more deassociation factors, including computational overload, bandwidth limitation, power limitation, etc. In some embodiments, the apparatus of execution process 500 optionally includes components such as processor 202, memory 204, network interface 206, etc., for generating a request for deassociation and one or more deassociation factors based on monitoring or predicting changes to one or more attributes, states, or other local environment data associated with the device. For example, the FLDN can determine that the computing load, bandwidth, and / or power levels meet predetermined thresholds, where such determination can cause the FLDN to generate and provide a request to the FLA computing environment for unassociation with one or more personalized tasks. In some embodiments, the request for unassociation indicates the duration for which the device is unavailable for execution process 600. The duration of unavailability can be generated by the device executing process 600 based on local environment data or other attributes or states of the device.

[0138] In various embodiments, after one or more of blocks 609 to 615, the FLA computing environment associates the device performing process 600 with the personalized task, causing the device to participate in one or more model training iterations of the FMTL training process to train a personalized global machine learning model to perform the personalized task.

[0139] In some embodiments, at block 621, the means for performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for receiving an initial version of a personalized global machine learning model for performing a personalized task, or for receiving subsequent iterations of a personalized global machine learning model after one or more model training iterations. For example, FLDN receives a personalized global machine learning model from an FLA computing environment.

[0140] In some embodiments, at block 624, the means of performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for training an initial version or subsequent iteration of a personalized global machine learning model to perform a corresponding personalized task. For example, FLDN can use local training data including multiple local training samples to train the personalized global machine learning model locally and independently to perform the personalized task. In some embodiments, at block 624, the means generates a personalized local machine learning model based on the personalized global machine learning model. In some embodiments, the means generates local model parameters that define the personalized local machine learning model. In some embodiments, the means of performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for generating synthetic data and generating additional local training samples based on the synthetic data to improve the training of the local model by using the additional synthetic training samples.

[0141] In some embodiments, at block 627, the means of performing process 600 includes components such as processor 202, memory 204, network interface 206, etc., for providing a personalized local machine learning model to an FLA computing environment, wherein the FLA computing environment can generate subsequent iterations of a personalized global machine learning model based on the personalized local machine learning model received from the corresponding participant FLDN. For example, the FLDN provides the FLA computing environment with local model parameters acquired via local training. In some embodiments, the means reports metadata associated with local training, such as the number of training samples.

[0142] In some embodiments, at block 630, the apparatus for performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for generating one or more training scores for iterations of a personalized global machine learning model trained by the apparatus. In some embodiments, the training scores include performance scores, confidence scores, etc. For example, during each iteration or subset of iterations of the FMTL training process, FLDN generates performance scores and confidence scores for the trained model.

[0143] In some embodiments, at block 633, the means of performing process 500 includes components such as processor 202, memory 204, network interface 206, etc., for providing one or more training scores to the FLA computing environment. For example, during each iteration or subset of iterations of the FMTL training process, the FLDN provides the FLA computing environment with performance scores and confidence scores for the current trained iteration of the personalized global machine learning model (e.g., which may be a personalized local machine learning model at the FLDN).

[0144] In various embodiments, the methods, apparatus, and computer program products of this disclosure provide management mechanisms for performing Federated Multi-Task Learning (FMTL) training for machine learning models. These methods, apparatus, and computer program products provide improved solutions for integrating FMTL into the training of machine learning models for network automation. In various embodiments, these methods, apparatus, and computer program products improve network automation by overcoming the technical challenges associated with achieving high statistical accuracy and model reliability in machine learning models performing network management functions.

[0145] It should be understood that each block of the flowcharts shown in the figures and described herein, as well as combinations of blocks in the flowcharts, can be implemented by various components, such as hardware, firmware, processors, circuit systems, and / or communication devices associated with the execution of software including one or more program instructions. For example, one or more of the processes or operations described above can be implemented by computer program instructions. In this regard, computer program instructions embodying the processes or operations described above can be stored in memory 204 of an apparatus employing the disclosed embodiments (e.g., user equipment (UE)) and executed by processor 202. It should be understood that any such computer program instructions can be loaded onto a computer or other programmable device (e.g., hardware) to produce a machine, such that the resulting computer or other programmable device performs the functions specified in the flowchart blocks. These computer program instructions can also be stored in a computer-readable storage medium that can instruct a computer or other programmable device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture whose execution performs the functions specified in the flowchart blocks. Computer program instructions may also be loaded onto a computer or other programmable device such that a series of operations will be performed on the computer or other programmable device to produce a computer-implemented process, such that the instructions executed on the computer or other programmable device provide operations for implementing the function specified in the flow block.

[0146] Benefiting from the teachings presented in the foregoing description and associated drawings, those skilled in the art will conceive of numerous modifications and other embodiments of the present disclosure set forth herein. Therefore, it should be understood that the present disclosure is not, and is not limited to, the specific embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Furthermore, although the foregoing description and associated drawings are exemplary embodiments described in the context of certain example combinations of elements and / or functions, it should be understood that different combinations of elements and / or functions may be provided by alternative embodiments without departing from the scope of the appended claims. In this regard, for example, different combinations of elements and / or functions, in addition to those explicitly described above, are also contemplated as being set forth in some of the claims in the appended claims. Although specific terminology is used herein, it is used only in a general and descriptive sense and not for limiting purposes.

Claims

1. A method (500) for a network device in a federated learning aggregation (FLA) computing environment of a communication network, the method comprising: For at least a first personalized task and a second personalized task among a plurality of personalized tasks for performing network automation operations for the communication network, determine (503) the following: (i) a set of first FLDNs (105) associated with the first personalized task from a plurality of federated learning distributed nodes (FLDNs) (105), and (ii) a set of second FLDNs (105) associated with the second personalized task from the plurality of FLDNs (105); Generate (518) a first personalized machine learning (ML) model for performing the first personalized task and a second personalized ML model for performing the second personalized task; and Execute (521) a plurality of corresponding model training iterations for training the first personalized ML model and the second personalized ML model, the plurality of corresponding model training iterations including: In the corresponding initial model training iterations for the first personalized ML model and the second personalized ML model, (i) the first personalized ML model is provided (521) to the corresponding FLDN (105) of the first set, and (ii) the second personalized ML model is provided (521) to the corresponding FLDN (105) of the second set. In the subsequent model training iterations for the first personalized ML model and the second personalized ML model, (i) the subsequent trained iterations of the first personalized ML model are provided (521) to the corresponding FLDN (105) of the first set, and (ii) the subsequent trained iterations of the first personalized ML model are provided (521) to the corresponding FLDN (105) of the second set, wherein: In the corresponding model training iteration, the corresponding FLDN(105) of the first set and the corresponding FLDN(105) of the second set independently train the corresponding personalized ML model to perform the corresponding personalized task; In the corresponding model training iteration, local model parameters are received (524) from the corresponding FLDN (105), wherein the local model parameters are associated with the training of the corresponding personalized ML model, the training being performed by the corresponding FLDN (105) of the first set or the second set; and (i) Based on the aggregation of the local model parameters associated with the corresponding FLDN (105) of the first set, generate (530) the subsequent trained iterations of the first personalized ML model; and (ii) Based on the aggregation of the local model parameters associated with the corresponding FLDN (105) of the second set, generate (530) the subsequent trained iterations of the second personalized ML model, wherein: The final iteration output of each of the corresponding plurality of training iterations includes: a first personalized global ML model for performing the first personalized task and a second personalized global ML model for performing the second personalized task.

2. The method according to claim 1, wherein: The plurality of FLDNs (105) include one or more separate management domains; and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using a cross-domain management domain (103).

3. The method according to claim 1, wherein: The plurality of FLDNs (105) includes one or more network nodes; and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using the Radio Access Network (RAN) management domain (103).

4. The method according to claim 1, wherein: The plurality of FLDNs (105) include one or more Network Data Analysis Functions (NWDAF); and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using the core management domain (103).

5. The method according to claim 1, wherein: The plurality of FLDNs (105) include one or more distributed NWDAFs; and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using a central NWDAF (103).

6. The method according to claim 1, wherein: The plurality of FLDNs (105) include one or more user equipment (UEs); and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using network nodes (103).

7. The method according to claim 1, wherein: The plurality of FLDNs (105) include one or more real-time RAN intelligent controllers (RICs); and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using non-real-time RIC (103).

8. The method according to claim 1, wherein: The plurality of FLDNs (105) includes one or more network nodes; and The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using real-time RIC (103).

9. The method according to claim 1, wherein: The determination of the first set and the second set of the plurality of FLDNs (105), the generation of the first personalized ML model and the second personalized ML model, and the execution of the corresponding plurality of training iterations are performed using the first UE (103); and The plurality of FLDNs (105) include one or more additional UEs.

10. The method according to any one of claims 1 to 9, further comprising, in the corresponding model training iteration: From the corresponding FLDN (105) of the first set and the corresponding FLDN (105) of the second set, receive (527) at least one training score associated with the training of the corresponding personalized ML model performed by the FLDN (105); and Based on the training scores received from the corresponding FLDN (105), (530) training scores of the corresponding aggregates for the first personalized ML model and the second personalized ML model are generated.

11. The method of claim 10, further comprising: (536) Provide the computing device associated with the operator entity with the training score of the at least one aggregate associated with the first personalized ML model and the score of the at least one aggregate associated with the second personalized ML model.

12. The method of claim 11, wherein: The corresponding training score in the at least one training score includes a performance score for performing the corresponding personalized task; as well as The corresponding training score in the training score of the at least one aggregate includes the performance score of the aggregate used to perform the corresponding personalized task.

13. The method of claim 12, wherein: The corresponding training score also includes a credibility score for performing the corresponding personalized task; and The training score of the corresponding aggregation also includes the credibility score of the aggregation used to perform the corresponding personalized task.

14. The method according to any one of claims 11 to 13, further comprising: Based on user input received from the computing device associated with the operator entity, the plurality of personalized tasks are determined (506, 512), wherein the user input indicates at least one machine learning model use case.

15. The method of claim 14, further comprising: Receive (506) from data storage a plurality of historical personalization tasks based on the at least one machine learning model use case, wherein the plurality of personalization tasks are determined based on the plurality of historical personalization tasks.

16. The method of claim 15, further comprising: Generate (515) a mapping of the at least one machine learning use case to a subset of the plurality of historical personalization tasks; as well as Based on the mapping, generate (515) at least the following: (i) a first category comprising the first subset of the plurality of FLDNs (105), and (ii) a second category comprising the second subset of the plurality of FLDNs (105), wherein the plurality of personalized tasks are also determined based on the first category and the second category.

17. The method of claim 15, further comprising: Provide (509) the plurality of historical personalized tasks to the computing device associated with the operator entity; Receive (512) a second user input from the computing device associated with the operator entity, wherein the second user input indicates at least one of the following: a filtered subset of the plurality of historical personalization tasks, or at least one additional personalization task; as well as The plurality of personalized tasks are also determined based on at least one of the following: a subset of the filtered plurality of historical personalized tasks, or the at least one additional personalized task.

18. The method according to any one of claims 1 to 16, further comprising, between one or more model training iterations of the plurality of model training iterations: Determine (506) the third personalized task; Determine (515) a third set of the plurality of FLDNs (105) associated with the third personalized task; Generate (518) a third personalized machine learning model to perform the third personalized task; as well as Provide the third personalized machine learning model (521), or a subsequent trained iteration of the third personalized machine learning model, to the third FLDN (105) set; wherein: The corresponding FLDN (105) of the third set trains the third personalized machine learning model or the subsequent trained iterations of the third personalized machine learning model in the remaining subset of the multiple model iterations. as well as In the final model training iteration, the personalized global machine learning model for the third personalized task is generated based on the corresponding local model parameters received from the corresponding FLDN (105) of the third set during the remaining subset of the multiple model training iterations.

19. The method of claim 18, wherein: One or more of the following items are determined based on one or more attributes or states of the plurality of FLDNs (105): the first set of the plurality of FLDNs (105), the second set of the plurality of FLDNs (105), or the third set of the plurality of FLDNs (105).

20. The method of claim 17 or claim 18, further comprising: Based on changes to one or more attributes or states of the at least one FLDN, remove (506) at least one FLDN from the first set of the plurality of FLDNs (105); and Based on the change to one or more attributes or states of the at least one FLDN, the at least one FLDN is added to the second set of the plurality of FLDNs (105) or the third set of the plurality of FLDNs (105).

21. The method according to claim 19, wherein: The one or more attributes or states of the plurality of FLDNs (105) include at least one of the following: computational resource availability, power availability, channel quality, data quality score, or data credibility score.

22. The method according to claim 19 or claim 20, wherein: The one or more attributes or states of the plurality of FLDNs (105) include at least one of the following: location data, velocity data, trust score, or the number of available training data samples.

23. The method according to any one of claims 1 to 21, further comprising: Provide (515) notifications to the plurality of FLDNs (105) indicating the plurality of personalized tasks; Receive (512) a request from at least one of the plurality of FLDNs (105) to associate the at least one FLDN with the first personalization task or the second personalization task; as well as Based on the request from the at least one FLDN, determine (515) at least one of the following: the first set of the plurality of FLDNs (105), or the second set of the plurality of FLDNs (105).

24. The method according to any one of claims 1 to 22, further comprising: Receive (512) a request to unassociate with the first set of the plurality of FLDNs (105) or the second set of the plurality of FLDNs (105) from at least one FLDN; Determine (506) that at least one training score associated with the at least one FLDN and the personalized machine learning model trained by the at least one FLDN satisfies a predetermined threshold; as well as In response to the at least one training score satisfying the predetermined threshold, a request (506) is provided to the at least one FLDN regarding the following: for the plurality of training iterations, the at least one FLDN is retained in the corresponding set of the plurality of FLDNs (105), wherein the request to retain the at least one FLDN indicates a fund retention incentive.

25. The method of claim 24, wherein: The request to unassociate includes at least one unassociation factor; and The at least one decoupling factor is at least one of the following: computational overload, bandwidth limitation, or power limitation.

26. An apparatus for a federated learning aggregation (FLA) computing environment for a communication network, the apparatus comprising components for performing the method according to any one of claims 1 to 25.

27. A computer program comprising instructions that, when executed by means of a federated learning aggregation (FLA) computing environment for a communication network, cause the means to perform the method according to any one of claims 1 to 25.

Images