The first node of the first machine learning model and the method executed by it.

Abstract

A method performed by a first node (111). The first node (111) obtains (402) one or more criteria to be satisfied by first data, which is used as input to obtain a first ML model for prediction in a first domain. A second ML model has been trained with second data from a second domain to make predictions in the second domain. The first node (111) obtains (403) information characterizing the second data and determines (406) a first data sample in the second data whose satisfaction level of the one or more criteria does not exceed a first threshold. The first data sample has an influence on the second ML model. The first node (111) obtains (409) the first ML model by performing a forgetting of the second ML model, which is performed by reducing the influence of the first data sample. The first node (111) then provides (410) an instruction to the first ML model.

Description

Technical Field

[0001] This disclosure generally relates to a method for processing a first node of a first machine learning model and performing such processing thereon. This disclosure also generally relates to a computer program and a computer-readable storage medium having a computer program for performing the method stored thereon. Background Technology

[0002] A computer system within a communication network or system may include one or more nodes. A node may include one or more processors (which, along with computer program code, can perform various functions and actions), as well as memory, receive ports, and transmit ports. A node may, for example, be a server. Nodes may perform their functions entirely in the cloud.

[0003] Computer systems can be included in telecommunications networks. Telecommunications networks can cover geographical areas that can be divided into cell areas, each served by a type of node, namely a network node, radio network node, or transport point (TP) in a radio access network (RAN), such as an access node, a base station (BS), or a radio base station (RBS), which may sometimes be referred to as a gNB, evolved Node B (“eNB”), “eNodeB”, “NodeB”, “B node”, or base transceiver station (BTS), depending on the technology and terminology used. Based on transmission power and thus cell size, base stations can have different categories, such as, for example, wide-area base stations, medium-range base stations, local base stations, and home base stations. A cell can be understood as a geographical area that can be radio-covered by a base station at a base station site. A base station located at a base station site can serve one or more cells. Furthermore, each base station can support one or more communication technologies. Telecommunications networks may also include network nodes that can serve receiving nodes, such as user equipment, through a service beam.

[0004] The standards organization 3GPP is currently defining a new air interface known as next-generation radio or new air interface (NR), and a fifth-generation (5G) packet core network, which can be referred to as the 5G core network (5GC). The advantages of 5G NR can include higher bandwidth, more resources, lower latency, and network slicing. 5G can provide services for a variety of applications, such as enhanced mobile broadband (eMBB), machine-to-machine type communication (mMTC), and ultra-reliable low-latency communication (URLLC).

[0005] 5G can be understood as bringing considerable flexibility with technological advancements and innovations in cloud computing and artificial intelligence (AI). This can be interpreted as bringing a whole new set of opportunities to the enterprise sector.

[0006] For many businesses, mobile cellular technology has proven to bring tremendous value to their digitalization processes, which can encompass numerous use cases such as autonomous robots, enhanced video services, connected vehicles, remote operation, and hazard and maintenance sensors. This can be understood as not only increasing productivity in connected factories but also making workplaces safer.

[0007] During the operation of a telecommunications network, data can be collected through the telecommunications network, which enables the monitoring and management of various functions.

[0008] For example, the emergence of the Internet of Things (IoT) has led to an exponential increase in the amount of data to be monitored. The availability of massive amounts of data (such as data collected from IoT devices) can be understood as enabling the analysis of such data to predict events with high predictive power. Predicting events can be understood as building mathematical models that can fit those data, and these models can then be used to predict such events. In this context, machine learning models can be used to analyze the collected data and enable improved management of different types of operations via telecommunications networks.

[0009] Machine Learning Machine learning (ML) can be understood as the study of computer algorithms that automatically improve through experience. It is considered a part of AI. ML algorithms can build models based on sample data (called "training data") to make predictions or decisions in situations where they are not explicitly programmed to do so. ML algorithms can be used in a wide variety of applications, such as email filtering and computer vision, where developing conventional algorithms to perform the desired tasks can be difficult or infeasible.

[0010] There are basically three types of ML algorithms: supervised learning, unsupervised learning, and reinforcement learning (RL).

[0011] Supervised learning algorithms can include target / outcome variables or dependent variables, which may need to be predicted based on a given set of predictors (i.e., independent variables). Using this set of variables, a function can be generated that maps the input to the desired output. The training process can continue until the model can reach the desired level of accuracy with respect to the training data. Once the ML model has likely been trained, the inference process can begin, allowing the ML model to be used to compute the output on new data. Examples of supervised learning include regression, decision trees, random forests, KNN, logistic regression, etc.

[0012] In unsupervised learning algorithms, there may be no target or outcome variable to predict / estimate. It can be used to cluster populations into different groups, which can be widely used to segment customers into different groups for specific interventions. Examples of unsupervised learning include K-means, mean-shift clustering, density-based spatial clustering with noise (DBSCAN), expectation-maximization (EM) clustering using Gaussian mixture models (GMM), agglomerative hierarchical clustering, and more.

[0013] Cluster analysis, or clustering, can be understood as a machine learning (ML) technique that involves grouping a set of objects in such a way that objects within the same group (which may be called a cluster) are, in a sense, more similar to each other than objects in other groups (i.e., other clusters). It can be understood as a primary task of exploratory data mining and a common technique for statistical data analysis, used in many fields including pattern recognition, image analysis, information retrieval, bioinformatics, data compression, computer graphics, and machine learning.

[0014] Using Reinforcement (RL) algorithms, machines can be trained to make specific decisions. It can be understood as follows: the machine is exposed to an environment where it can continuously train itself using trial and error. The machine can learn from past experiences and attempt to acquire the best possible knowledge to make accurate decisions. An example of RL could be a Markov Decision Process (MDP). Training using RL can include generating ML models. To train such an ML model, given an environment state, an agent can take an action in that environment and receive a reward. This action can lead to a new state of the environment. This process can be repeated in loops. Over time, the agent can learn to take actions that may lead to greater immediate and future rewards, meaning it can be understood that what is in the agent's best interest is not taking the action that might only lead to the highest reward in the next state, but rather the action that might cumulatively lead to the next state and many future states.

[0015] Agents can include neural networks that can take state as input and generate actions. Several ML algorithms can be used to train the network, such as policy-based learning algorithms like the actor-critic approach, or value-based learning algorithms like deep q-networks.

[0016] Network Data Analytics Function (NWDAF) can be understood as a Network Function (NF) within the 5G core network (5GC). NWDAF is designed to collect data from various data sources, such as User Equipment (UE), other NFs within the 5GC, as well as Operations, Management and Maintenance (OAM) systems, cloud, and edge networks. NWDAF can utilize and process the collected data to train machine learning (ML) models, which can then be used to generate analytical reports based on predictions and information from the past, and deliver them to the NFs. Accordingly, different NFs (called service consumers) can subscribe to NWDAF. Different NFs can benefit from the capabilities of NWDAF and request information about network status, such as network data related to slice load levels, user equipment (UE) related data, and user data congestion analysis.

[0017] Figure 1 This is a diagram illustrating one of the potential problems. Consider an ML model, which we might call source model 1. Machine learning using existing methods may face some challenges, and it may have already been trained on data used to solve the task. The data may have been obtained from source domain 2, where many data samples 3 are accessible. Consider target domain 4, where there may be a limited number of data samples available, making it insufficient to train the ML model using only these. In this context, it might be desirable to transfer source model 1 to target domain 4 and use source model 1 to enrich the ML model for target domain 4 (referred to as the target model in this paper). The first complexity in this scenario is that source model 1 may have been trained using data samples that may not conform to the target domain's specifications. Specifications can be understood as one or more criteria that data in a particular domain may need to meet. Specifically, the source domain may contain data samples that the target domain may not be able to access, or may contain data samples that may not support the use cases in the target domain, such as data samples corresponding to different versions, or being unlabeled, etc. Therefore, source model 1 cannot be used directly because it may infringe on data privacy rights or may not be suitable for the use cases in the target domain. Figure 1 The following data samples are illustrated to illustrate how the source model 1 was trained: a) data sample 5, which may not conform to the target domain specification; b) data sample 6, which may not conform to the source specification; and c) data sample 7, which may conform to both the source specification and the data target specification.

[0018] Figure 2This is a schematic diagram depicting existing techniques (sample selection and source model selection) for handling this problem. According to the first method (which may be referred to as sample selection) schematically depicted in panel a), samples that do not meet the requirements of the target domain can be removed, and the source model can be trained to be personalized for the target domain. According to the second method (which may be referred to as source model selection) schematically depicted in panel b), a source model can be selected from a dictionary of source models that best correspond to the target specification.

[0019] Machine learning based on the existing methods described above can lead to a waste of time and computing resources, and is expensive. Summary of the Invention

[0020] As part of the development of the embodiments herein, one or more problems of the prior art will first be described and discussed.

[0021] exist Figure 2 The first method depicted in panel a) has several problems. First, retraining the new source model will take time and requires sufficient computational resources to be available at the compute nodes, since the source model may typically be trained on many samples. Second, it requires all data samples to be available at the compute nodes (or alternatively in a cloud implementation), which requires back-and-forth communication between the compute nodes and the data nodes from which data can be obtained.

[0022] exist Figure 2 The second method depicted in panel b) has two main problems. First, it requires maintaining different source models trained on data samples with different specifications. Managing the lifecycle of many models can become expensive. Second, the source models selected from the dictionary of source models may still only be partially relevant to the target domain. Furthermore, it may be found that source models already trained on data samples that do not conform to the target specification are unsuitable for use. According to the third method (which can be called a target-agnostic model, and which is not included in...),... Figure 2 (Illustrated schematically) This allows for the construction of a target-agnostic model. That is, a highly generalized model that can be applied to a large class of use cases in the target domain. However, in practice, such a model may only be suitable for use cases in the target domain that are somewhat related to the source domain. Finally, this approach is unsuitable when some of the data used in training the target-agnostic model contains privacy-sensitive information or erroneous data samples.

[0023] A second complication might be that the source model was trained on a dataset containing errors or biased samples that were unknown at training time. Existing methods remove these samples and retrain the source model from scratch. However, retraining the source model can be expensive, especially with large datasets.

[0024] Based on the foregoing, one objective of the embodiments described herein is to improve the handling of machine learning models in computer systems.

[0025] According to a first aspect of the embodiments herein, this objective is achieved by a computer-implemented method executed by a first node. The method is used to dispose of a first machine learning model. The first node operates in a computer system. The first node obtains one or more criteria to be satisfied by first data, which is to be used as input to obtain a first machine learning model for making predictions in a first domain. A second machine learning model has been trained with second data from a second domain to make predictions in the second domain. The first node obtains information characterizing the second data. The first node then determines a first data sample in the second data whose satisfaction level with the one or more criteria does not exceed a first threshold. The determined first data sample has an impact on the second machine learning model. The first node obtains the first machine learning model by performing a forgetting process on the second machine learning model. The first node performs the forgetting process by reducing the impact of the determined first data sample on the second machine learning model to exceed a second threshold. The first node then provides an indication of the obtained first machine learning model.

[0026] According to a second aspect of the embodiments herein, this objective is achieved by the first node. The first node can be understood as being used to dispose of the first machine learning model. The first node is configured to operate in the computer system. The first node is configured to obtain one or more criteria to be satisfied by the first data, which is to be used as input to obtain the first machine learning model for making predictions in the first domain. The second machine learning model has been trained with the second data from the second domain to make predictions in the second domain. The first node is also configured to obtain the information configured to characterize the second data. The first node is further configured to determine first data samples in the second data whose satisfaction level for the one or more criteria does not exceed a first threshold. The first data samples are configured to be determined to have the effect on the second machine learning model. The first node is further configured to obtain the first machine learning model by performing the forgetting process of the second machine learning model, the forgetting process being performed by reducing the effect of the determined first data samples on the second machine learning model to exceed the second threshold. The first node is further configured to provide the indication configured to instruct the first machine learning model obtained by multiplying the configuration.

[0027] According to a third aspect of the embodiments herein, this objective is achieved by a computer-readable storage medium having a computer program stored thereon, the computer program including instructions that, when executed on at least one processing circuit, cause the at least one processing circuit to perform a method executed by a first node.

[0028] According to a fourth aspect of the embodiments herein, this objective is achieved by a computer program comprising instructions that, when executed on at least one processing circuit, cause the at least one processing circuit to perform the method executed by the first node.

[0029] By obtaining one or more criteria to be satisfied by the first data (which will be used as input to obtain a first machine learning model for making predictions in a first domain), the first node can then be enabled to check whether a suitable machine learning model exists, which may have already been trained with data that can adequately meet the one or more criteria, and is therefore suitable for obtaining the first machine learning model, rather than having to train a new model from scratch.

[0030] By obtaining information representing the second data, the first node can then be enabled to determine whether the second machine learning model is suitable as a starting point for adaptation, so as to obtain a first machine learning model for making predictions in the first domain.

[0031] By identifying first data samples in the second data whose satisfaction level with one or more criteria does not exceed a first threshold, a first node can be enabled to identify, for example, unwanted, noisy, biased, erroneous, and privacy-sensitive samples from a fully trained second machine learning model. The first node can then be enabled to remove the influence of such samples from the second machine learning model without retraining it through a forgetting process, thereby minimizing the potential impact of the first data samples on the second machine learning model. In other words, this is done to "erase" their potential influence on the second machine learning model.

[0032] The first machine learning model is obtained by performing a forgetting process on the second machine learning model. The first node can then be enabled to customize the second machine learning model according to one or more criteria of the first domain via the model forgetting process. The first node can then be enabled to provide the first machine learning model as an adaptive, on-demand customized model. This allows the first node to obtain the on-demand customized model in a shorter amount of time than would otherwise take to retrain the second machine learning model or train the first machine learning model from scratch, and therefore does not require high computational resources. Furthermore, the first node can obtain the first machine learning model even if not all data samples are available. Therefore, back-and-forth communication between the first node and the data source may not be necessary.

[0033] Custom models can be understood as models that do not require maintenance. Therefore, a minimum number of source models may need to be maintained and continuously evolved. Consequently, the potentially expensive storage requirements and maintenance work that would otherwise exist due to a large number of models can be advantageously reduced.

[0034] Since the second machine learning model can be understood as not needing to be retrained, but rather potentially needing to be maintained, this can be interpreted as providing potential energy savings.

[0035] One particular advantage may be that the first node can enable the resolution of requirements such as EU directives on AI, which may require compliance with the private rights of individuals to have their data removed from databases and systems (e.g., containing ML models).

[0036] It can also further enable the first node to send a custom model, now without the source node's unwanted representation.

[0037] By providing this instruction, the first node can enable the use of the first machine learning model, for example, to make predictions in the first domain. Attached Figure Description

[0038] Examples of embodiments described herein are described in more detail below with reference to the accompanying drawings.

[0039] Figure 1 This is a schematic diagram illustrating an example of a source model trained using samples that may not conform to the target domain specification according to existing methods.

[0040] Figure 2 This is a conceptual visualization of existing solutions for sample selection and source model selection.

[0041] Figure 3 These are schematic diagrams illustrating two non-limiting examples of computer systems according to embodiments of the present article (in panels a) and b).

[0042] Figure 4 This is a flowchart depicting the method in the first node according to the embodiments herein.

[0043] Figure 5 This is a schematic diagram depicting a component of a first node according to a non-limiting example of an embodiment of this document.

[0044] Figure 6 This is a schematic diagram depicting a particular aspect of another non-limiting example of a method performed by a first node according to embodiments herein.

[0045] Figure 7This is a schematic diagram illustrating another specific aspect of a non-limiting example of a method performed by a first node according to embodiments herein.

[0046] Figure 8 This is a signaling diagram depicting a non-limiting example of a method performed by a first node according to embodiments of this document.

[0047] Figure 9 is a schematic diagram illustrating another specific aspect of a method performed by a first node according to embodiments herein.

[0048] Figure 10 This is a signaling diagram depicting a non-limiting example of a method performed by a first node according to embodiments of this document.

[0049] Figure 11 is a signaling diagram depicting other aspects of a non-limiting example of a method performed by a first node according to embodiments of this document.

[0050] Figure 12 This is a schematic block diagram illustrating an embodiment of the first node according to the embodiments described herein. Detailed Implementation

[0051] Certain aspects of this disclosure and its embodiments address the challenges of existing methods identified in the Background and Summary of the Invention sections, and provide solutions to the challenges discussed.

[0052] The embodiments described herein can be understood as relating to a method and apparatus for creating an adaptive, on-demand customized model from a source model. In general terms, the method may include an apparatus (first node) receiving a request for a customized ML model for a specific target domain. The first node may receive data specifications or guidelines from both the source and target domains. Next, the first node may identify out-of-policy and / or irrelevant data samples, which can be understood as data samples in the source domain that may not meet the specifications of the target domain. The first node may then execute a machine forgetting method that allows the removal of the effects of the out-of-policy and / or irrelevant data samples from the source model, thereby constructing an adaptive version of the source model that can be customized to the needs of the target, with the aim of ensuring that the customized model follows the data guidelines and specifications from both the source and target domains. The new model for the target domain can then be passed to the requesting entity.

[0053] The embodiments described herein can be understood as relating to how to maintain a minimum number of source models and how to customize them as needed for the target domain.

[0054] Some embodiments under consideration will now be described more fully below with reference to the accompanying drawings, in which examples are illustrated. In this section, embodiments described herein will be illustrated in more detail by way of a number of exemplary embodiments. However, other embodiments are included within the scope of the subject matter disclosed herein. The disclosed subject matter should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided as examples to convey the scope of the subject matter to those skilled in the art. It should be noted that the exemplary embodiments described herein are not mutually exclusive. Components from one embodiment may be assumed by default to exist in another embodiment, and how these components can be used in other exemplary embodiments will be apparent to those skilled in the art.

[0055] This document includes several embodiments and examples. It should be noted that the embodiments and / or examples herein are not mutually exclusive. Components from one embodiment or example may be assumed to exist in another embodiment or example by default, and how these components can be used in other exemplary embodiments and / or examples will be apparent to those skilled in the art.

[0056] Figure 3 Two non-limiting examples of a computer system 100 in which embodiments of this document can be implemented are depicted (in panels “a” and “b”, respectively). In some example implementations (such as in Figure 3 (As depicted in the non-limiting example of a), computer system 100 can be a computer network. In other example implementations (such as in...) Figure 3(As depicted in the non-limiting example of b), computer system 100 can be implemented in a telecommunications system (sometimes also referred to as a telecommunications network, cellular radio system, cellular network, or wireless communication system). In some examples, the telecommunications system may include network nodes that can serve receiving nodes, such as wireless devices. Computer system 100 may be, for example, a network such as a 5G system or an updated system supporting similar functionality. The telecommunications system may additionally support other technologies, such as, for example, Long Term Evolution (LTE), such as LTE Frequency Division Duplex (FDD), LTE Time Division Duplex (TDD), LTE Half-Duplex Frequency Division Duplex (HD-FDD), or LTE operating in unlicensed frequency bands. Telecommunication systems can also support other technologies, such as Wideband Code Division Multiple Access (WCDMA), Universal Mobile Telecommunications System Terrestrial Radio Access (UTRA) TDD, Global System for Mobile Communications (GSM) networks, GSM / GSM Evolution Enhanced Data Rate (EDGE) Radio Access Network (GERAN) networks, Ultra Mobile Broadband (UMB), EDGE, any combination of Radio Access Technologies (RATs) (such as Multi-Standard Radio (MSR) base stations, multi-RAT base stations, etc.), any 3GPP cellular network, Wireless Local Area Network (WLAN) or WiFi network, Global Microwave Access Interoperability (WiMax), Low Power Short Range (LPWAN) networks based on IEEE 802.15.4 (such as IPv6 (6LowPAN) based on Low Power Wireless Personal Area Networks), Zigbee, Z-Wave, Bluetooth Low Energy (BLE), or any cellular network or system. Telecommunication systems can, for example, support Low Power Wide Area Networks (LPWANs). LPWAN technologies can include long-range physical layer protocols (LoRa), Haystack, SigFox, LTE-M, and narrowband IoT (NB-IoT).

[0057] Computer system 100 includes a first node 111. Computer system 100 may include additional nodes. In some embodiments, computer system 100 may include a second node 112. Figure 3 In other embodiments depicted in the non-limiting examples, the computer system 100 may alternatively or additionally include a plurality of third nodes 113. The computer system 100 may include additional nodes.

[0058] Any one of the first node 111, the second node 112, and the plurality of third nodes 113 can be understood as a first computer system or server, a second computer system or server, and the plurality of third computer systems or servers, respectively. Any one of the third nodes 111, 112, and 113 can be implemented as an independent server, such as a host computer, in the cloud 115. Figure 3(b) The non-limiting example depicts the first node 111, the second node 112, and the plurality of third nodes 113. In other examples, any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 may be a distributed node or a distributed server (such as a virtual node in cloud 115), and may perform some of its respective functions locally, for example, by a client manager, and some of its functions in cloud 115, for example, by a server manager. In other examples, any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 may perform its functions entirely on cloud 115, or partially in cooperation with or co-located with radio network nodes. In still other examples, any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 may also be implemented as a processing resource in a server farm.

[0059] In other examples, any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 can also be implemented as a virtual network function (e.g., according to a network function virtualization (NFV) architecture).

[0060] Any one of the first node 111, the second node 112, and the plurality of third nodes 113 may be under the ownership or control of the service provider, or may be operated by the service provider or on behalf of the service provider.

[0061] exist Figure 3 In some examples not depicted, any one of the first node 111, the second node 112, and any one of the plurality of third nodes 113 may be co-located or be the same node. However, in typical embodiments, the first node 111, the second node 112, and the plurality of third nodes 113 may be different nodes.

[0062] Any of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 may have the ability to perform a computer-implemented learning process (which may also be referred to as “machine learning” (ML)).

[0063] In some non-limiting examples, any one of the third nodes 111, 112, and 113 may have the ability to manage an artificial neural network. An artificial neural network can be understood as a machine learning framework that may include a set of connected nodes, where each node or perceptron may have a basic decision unit. Each such node may have one or more inputs and one output. Inputs to a node may come from the output of another node or from a data source. Each node and connection may have specific weights or parameters associated with it. To solve a decision task, the weights may be learned or optimized with respect to a dataset representative of the decision task. Most commonly, nodes may allow each input to be weighted individually, and the sum may be obtained through a non-linear function, which may be called an activation function. The properties of the connections and nodes can determine the type of neural network, such as a feedforward network, a recurrent neural network, etc. Any one of the third nodes 111, 112, and 113 may have the ability to manage an artificial neural network, which can be understood herein as having the ability to: store training datasets and models that may be generated from machine learning, train new models, and use the model to make predictions once the model has been trained.

[0064] Any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 can, for example, support running Python / Java using Tensorflow, PyTorch, Theano, etc. Any one of the third nodes among the first node 111, the second node 112, and the plurality of third nodes 113 can also have GPU capabilities.

[0065] In some examples, any one of the third nodes 111, 112, and 113 can be a network analysis logic function managed by the operator. That is, it can be a node that has the ability to handle the collection and analysis of data from different sources in computer system 100. Any of the third nodes 111, 112, and 113 can interact with different entities for different purposes, such as: collecting data provided by, for example, Access and Mobility Functions (AMF), Session Management Functions (SMF), Policy Control Functions (PCF), Unified Data Management Functions (UDM), Application Functions (AF) (directly or via Network Exposure Functions (NEF) based on event subscriptions), and Operations and Management (OAM); retrieving information from data repositories, such as retrieving subscriber-related information from the Unified Data Repository (UDR) via the UDM, retrieving information about NFs, such as retrieving NF-related information from the NRF, and retrieving slice-related information from the Network Slice Selection Function (NSSF); providing analytics to consumers on demand; and storing two types of data in analytics data repository functions (e.g., Analytics Data Repository Functions (ADRF)): collected data (e.g., event exposure data) and analytics reports.

[0066] As in Figure 3 The non-limiting example of the third node among the first node 111, second node 112, and multiple third nodes 113 depicted herein (where computer system 100 may be a 5G network) may be a corresponding NWDAF. In a particular example, any one of the first node 111, second node 112, and multiple third nodes 113 may be a different Model Training Logical Function (MTLF), where they may store data in different regional NWDAFs. In a particular example of an embodiment herein, the first node 111 may be implemented at an ML Lifecycle Management (LCM) unit, for example, located in a 3GPP NWDAF.

[0067] The first node 111 can be a node with the ability to perform a forgetting process, as will be described later, for example, in conjunction with action 409.

[0068] The second node 112 may be a node that includes data from a target domain in which it may be interested to make predictions, but it may lack a machine learning model for making such predictions.

[0069] Multiple third nodes 113 can be nodes in the corresponding source domain that are being trained or have already trained their respective machine learning models. The corresponding source domains may or may not be the same domain.

[0070] The corresponding domain can be understood as being at least partially different from the target domain in which the second node 112 may have to make predictions.

[0071] In some examples, computer system 100 may include one or more radio network nodes, such as Figure 3 The radio network node 130 is depicted in b). The radio network node 130 may, for example, be included in the radio access network of a telecommunications system. That is, the radio network node 130 may be a transmission point (such as a radio base station, e.g., gNB, eNB) or any other network node with similar characteristics capable of serving wireless devices (such as user equipment or machine-type communication devices) in computer system 100. In a typical example, the radio network node 130 may be a base station, such as a gNB or eNB. In other examples, the radio network node 130 may be a distributed node, such as a virtual node in cloud 115, and may perform its functions entirely on cloud 115 or in part in cooperation with the radio network node.

[0072] The telecommunications system can cover a geographical area that can be divided into cell areas in some embodiments, wherein each cell area can be served by a radio network node 130, although one radio network node 130 may serve one or more cells. Figure 3 In the example, the cell is not depicted for simplicity. Based on the transmission power and thus the cell size, the radio network node 130 can be of different categories, such as, for example, a macro eNodeB, a home eNodeB, or a pico cell. In some examples, the radio network node 130 can serve receiving nodes via a serving beam. The radio network node 130 can be directly connected to one or more core networks.

[0073] Any of the first node 111 and the second node 112 and / or any of the nodes included in the computer system 100 may support one or more communication technologies, and their names may depend on the technologies and terminology used.

[0074] Device 140 may be included in a telecommunications network. Device 140 included in computer system 100 may be a wireless communication device, such as a 5G UE or UE (which may also be referred to as, for example, a mobile terminal, wireless terminal, and / or mobile station), customer premises equipment (CPE), mobile phone, cellular phone, or wireless-capable laptop computer (to name just a few other examples). Device 140 included in the telecommunications network may, for example, be a portable, pocket-sized, handheld, computer-integrated, or vehicle-mounted mobile device capable of transmitting voice and / or data via RAN to another entity (such as a server, laptop computer, personal digital assistant (PDA) or tablet computer, machine-to-machine (M2M) device, device equipped with a wireless interface (such as a printer or file storage device), modem, sensor, IoT device) or any other radio network unit capable of communicating on a radio link in the communication system. In a typical example, device 140 may be or include a microphone. Device 140 included in the telecommunications system may be enabled to perform wireless communication within the telecommunications system. Communication may be performed via, for example, the RAN and one or more core networks that may be included within the telecommunications system.

[0075] It is understood that the telecommunications network may include additional radio network nodes 130 and / or additional devices 140.

[0076] First node 111 can be configured to communicate with second node 112 within computer system 100 via a first link 141 (e.g., a radio link or a wired link). First node 111 can be configured to communicate with multiple third nodes 113 within computer system 100 via corresponding second links 142 (e.g., a radio link or a wired link). First node 111 can be configured to communicate with radio network node 130 within computer system 100 via a third link 143 (e.g., a radio link or a wired link). Second node 112 can be configured to communicate with radio network node 130 within computer system 100 via a fourth link 144 (e.g., a radio link or a wired link). Multiple third nodes 113 can be configured to communicate with radio network node 130 within computer system 100 via corresponding fifth links 145 (e.g., a radio link or a wired link). Radio network node 130 can be configured to communicate with device 140 within computer system 100 via a sixth link 146 (e.g., a radio link).

[0077] Any of the first link 141, the corresponding second link 142, the third link 143, the fourth link 144, the corresponding fifth link 145, and the sixth link 146 may be a direct link, or may include multiple separate links, wherein it may pass through one or more computer systems or one or more core networks in computer system 100. Figure 3 (Not depicted in the text), or it may be via an optional intermediate network. The intermediate network can be one of a public, private, or hosted network, or a combination of more than one; the intermediate network (if any) can be a backbone network or the Internet; in particular, the intermediate network can contain two or more subnets, which is... Figure 3 Not shown in the image.

[0078] Generally, the use of “first,” “second,” “third,” “fourth,” “fifth,” and / or “sixth” here can be understood as any way of referring to different elements or entities, and can be understood as not giving the nouns they modify the characteristics of accumulation or chronological order.

[0079] Some contemplated embodiments will now be described more fully with reference to the accompanying drawings. However, other embodiments are included within the scope of the subject matter disclosed herein, and the disclosed subject matter should not be construed as being limited to the embodiments set forth herein; rather, these embodiments are provided as examples to convey the scope of the subject matter to those skilled in the art.

[0080] Now refer to Figure 4 The flowchart depicted illustrates an embodiment of a computer-implemented method executed by the first node 111. This method is used to process a first machine learning model. The first node 111 operates within a computer system 100.

[0081] This document includes several embodiments. In some embodiments, all actions may be performed. In some embodiments, one or more actions may be optional. Figure 4 In the text, optional actions are indicated by dashed lines. It should be noted that the examples herein are not mutually exclusive. Components from one embodiment may be assumed to exist in another embodiment, and how these components can be used in other exemplary embodiments will be apparent to those skilled in the art. Where applicable, one or more embodiments may be combined. For the sake of simplicity, not all possible combinations have been described.

[0082] In some embodiments, computer system 100 may be a 5G system. In some such embodiments, first node 111 may be an NWDAF.

[0083] Action 401 In action 401, the first node 111 may receive a first instruction from the second node 112 operating in the computer system 100. This first instruction may indicate a request for a first machine learning model used to make predictions in a first domain.

[0084] A domain can be understood as an environment in which a machine learning model can be trained or may have already been trained. As used herein, a domain can be understood as a data domain, which can have unique characteristics such as, for example, data distribution, data features, system characteristics, environmental attributes, data protection rights, etc. These unique characteristics may differ between data domains. In the embodiments described herein, the first domain can be understood as the target domain. That is, the domain in which predictions may be of interest (e.g., by the second node 112).

[0085] The receiving in action 401 can be performed, for example, via the first link 141.

[0086] In some embodiments where computer system 100 may be a 5G system, the second node 112 may be an NWDAF. In some such embodiments, the first node 111 may be a first NWDAF, and the second node 112 may be a second NWDAF.

[0087] The second node 112, also referred to as the "target node," can be understood as a node from a second (target) domain that may require the first machine learning model to make predictions in the first domain. The second node 112 may have data but no model for making the desired predictions. In some non-limiting examples where the second node 112 could be an NWDAF, the second node 112 may query another node operating in the computer system 100 (e.g., a Network Repository Function (NRF)) for the discovery of available NWDAFs before sending the first instruction to the first node 111. The NRF may indicate the availability of the first node 111 to the second node 112, which can then trigger the second node 112 to send the first instruction to the first node 111 in action 111.

[0088] By receiving a first instruction from the second node 112 in action 401, the first node 111 can enable the second node 112 to eventually obtain a predictive model (i.e., a first machine learning model) for making predictions in the first domain, without requiring the second node 112 to possibly train and / or maintain the first machine learning model itself.

[0089] Action 402 In action 402, the first node 111 obtains one or more criteria to be satisfied by the first data, which is to be used as input to obtain a first machine learning model for making predictions in the first domain.

[0090] The acquisition in action 401 may include receiving, for example, from a second node 112 via a first link 141.

[0091] The primary data can also be referred to as the target data. In this article, the target can be understood as the area where a primary machine learning model might be expected or needed to make predictions. It can be used to represent data from the first domain (i.e., the target domain).

[0092] The one or more criteria can be understood as specifications for the first domain. Therefore, the one or more criteria can also be referred to as target specifications. The one or more criteria can be understood as containing meta-information about the first domain, such as: environment type, version, characteristic attributes, sample-level data access rights, distribution-level data access rights, data quality metrics for each sample, etc. As a non-limiting illustrative example, the environment type could be a base station located in a city center area; the version could be LTE; the characteristic attributes could be analytical data related to the performance, configuration, and location of the base station and the core network; data access rights could be restricted for analytical data related to the UE's location, behavior, and mobility, and the data quality could be such that analytical data related to the performance, configuration, and location of the base station and the core network collected during weekends might need to be discarded.

[0093] In this document, It can be used to represent the first data in the first domain (i.e., the target domain).

[0094] In some non-restrictive examples, the acquisition of one or more criteria in action 402 can be performed by a first module included in the first node 111, referred to herein as the “target specification” module.

[0095] In some embodiments, obtaining one or more criteria in action 402 may be in response to a received first instruction.

[0096] In some embodiments, the first node 111 may periodically obtain the latest updates of the one or more guidelines, for example, via the target specification module.

[0097] By obtaining one or more criteria to be satisfied by the first data (which will be used as input to obtain a first machine learning model for making predictions in a first domain), the first node 111 can then be enabled to check whether a suitable machine learning model exists, which may have already been trained with data that can adequately meet the one or more criteria, and is therefore suitable for obtaining the first machine learning model, rather than having to train a new model from scratch.

[0098] Action 403 According to the embodiments herein, a second machine learning model has been trained with second data from a second domain to make predictions in that second domain. The second domain may be referred to herein as the source domain, as it can be understood as the domain on which the second machine learning model may have been trained. The second data may also be referred to herein as source data. The second machine learning model may be referred to herein as the source model. The second machine learning model can be understood as a model that has been trained on data samples from different datasets, each dataset having its own specification for a specific use case. In some examples, the second machine learning model may be a global model that can evolve as new data becomes available. The second machine learning model may belong to different categories of machine learning models, including but not limited to neural networks, random forests, linear models, etc. The only requirement for the second machine learning model can be understood as the category of the model must be differentiable or approximately differentiable. Differentiability can be understood as the ability to compute the derivative of the objective function of the learning problem with respect to the parameters of the learning model. In this context, this can be understood as meaning that for models of this category, they are differentiable in their parameter space, such that the derivative of the learning objective with respect to the parameters of the model can be computed.

[0099] In this document, It can be used to indicate the second data, namely the training data in the second domain.

[0100] In this document, This can indicate the use of second data (i.e., training data). The second machine learning model trained.

[0101] In action 403, the first node 111 obtains information representing the second data. This information representing the second data can also be referred to as the source data specification. In this document, It can be used to indicate the source data specification, where It can represent the nth data sample Source specifications. Source data specifications. This can be understood as containing meta-information about the second domain, such as the type of data features, feature attributes, data quality metrics for each sample, environmental specifications (e.g., environment type, information such as climate, landscape, temperature, density / rural), data access rights (at the sample level), distribution-level data access rights (e.g., who might have access to the data), and the time interval for data collection. Additionally, information about the software and hardware versions of the products receiving the data, as well as the radio access technologies (e.g., 3G, 4G, 5G...) and core network versions. The types of information that may characterize the second data can be understood as equivalent to the one or more criteria, but concerning the second data. Therefore, the information obtained in action 403 can be understood as the second information, i.e., the source specification, while the one or more criteria can be understood as the first information, i.e., the target specification. The information characterizing the second data may at least partially overlap with the one or more criteria.

[0102] In some non-limiting examples, the acquisition of information representing the second data in action 403 can be performed by a second module included in the first node 111, referred to herein as the “source specification” module.

[0103] In some embodiments, the first node 111 (e.g., via the source specification module) may periodically obtain the latest updates to the information representing the second data.

[0104] In some embodiments, the second machine learning model may be one of a plurality of second machine learning models that have been trained with corresponding second data from a second domain to make corresponding predictions in the second domain. In such an embodiment, obtaining information in action 403 may include obtaining corresponding information characterizing the corresponding second data. In other words, if there is more than one available second machine learning model, the first node 111 may obtain information about each of these second machine learning models in action 403.

[0105] In some embodiments where information representing the corresponding second data can be obtained from multiple third nodes 113, the computer system 100 may be a 5G network, and each of the first node 111, the second node 112, and the multiple third nodes 113 may be an NWDAF. The multiple third nodes 113 can be understood as multiple source nodes.

[0106] First node 111 may have access to all second machine learning models existing at different nodes (e.g., NWDAF, regional), and may request additional data specifications from the potential source NWDAFs where second machine learning models being trained on history or multiple second machine learning models may be stored.

[0107] By obtaining information characterizing the second data in action 403, the first node 111 can then be enabled to determine whether the second machine learning model, or at least one of a plurality of second machine learning models, is suitable as a starting point for adaptation, in order to obtain a first machine learning model for prediction in the first domain, as will be explained in subsequent actions 404 and 405. Furthermore, the first node 111 can then be enabled to determine data samples in the second data that may be insufficient to meet the one or more criteria, as will be explained in action 406.

[0108] Action 404 In action 404, the first node 111 can determine whether the level of satisfaction of the information representing the second data with the one or more criteria might exceed a first threshold. In other words, the first node 111 can determine whether the difference between the target specification and the source specification is within an acceptable threshold.

[0109] The first threshold can be understood as a configurable (e.g., user-defined) threshold.

[0110] Determination can be understood as calculation, estimation, deduction or similar means, or as being obtained or received from another node.

[0111] In some embodiments where the second machine learning model can be one of multiple second machine learning models that have been trained with corresponding second data from a second domain to make corresponding predictions in the second domain, and where obtaining information in action 403 can include obtaining corresponding information characterizing the corresponding second data, the determination in action 404 can further include determining whether the corresponding information satisfies a corresponding level of the one or more criteria exceeding a first threshold. Information that may exceed the first threshold can be a set of corresponding information.

[0112] By determining in action 404 whether the level of satisfaction of the information representing the second data with the one or more criteria might exceed a first threshold, the first node 111 can be enabled to determine whether it may be necessary to determine the first machine learning model from scratch. That is, by adapting the second machine learning model, it can be deduced whether the first machine learning model used for prediction in the first domain might not be worthwhile. In an alternative implementation, if the one or more criteria for the first domain are substantially different from the information representing the second data, the first node 111 can learn a new model tailored to the first domain instead of initiating the forgetting process described in action 409.

[0113] When multiple second machine learning models are available, and the level of satisfaction of the information representing the second data with the one or more criteria may exceed a first threshold, the first node 111 may continue to determine which of the multiple second machine learning models is likely to be closest to the one or more criteria (i.e., target specification) and use that second machine learning model.

[0114] If the level of satisfaction of the information representing the second data with the one or more criteria fails to exceed a first threshold, the first node 111 can avoid determining the first machine learning model based on the second machine learning model, and instead can continue to determine the first machine learning model from scratch. In other words, if the difference between the one or more criteria and all available source specifications (i.e., all corresponding information representing the corresponding second data) exceeds a first predefined threshold, a new first machine learning model can be constructed for the target. As previously mentioned, the first threshold can be a configurable (e.g., user-defined) threshold. The first node 111 can then optionally add this newly constructed model to the model's dictionary.

[0115] Action 405 In some embodiments where the second machine learning model can be one of multiple second machine learning models, the first node 111 may select a second machine learning model from the multiple second machine learning models in action 405. The selection in action 405 may be based on the corresponding satisfaction level of the one or more criteria. That is, the first node 111 may select the second machine learning model that may have the highest satisfaction level for the one or more criteria (i.e., may be closer to the one or more criteria).

[0116] If, for at least two of the multiple second machine learning models, the corresponding information representing the corresponding second data satisfies the level of one or more criteria exceeding a first threshold, the first node 111 may perform the selection in action 405.

[0117] Based on the corresponding information representing the corresponding second data and the one or more criteria (i.e., based on data specifications), the first node 111 can perform an evaluation and determine which second machine learning model may be useful for the first domain, and request the corresponding third node 113 in the second domain (e.g., NWDAF source) to send the second machine learning model and corresponding data, or representations that may not be expected to be represented in the first machine learning model.

[0118] In practice, if there is only a single source model, many changes may be necessary for each request. Where changes to the second machine learning model might be too drastic for a specific target domain model group, the first node 111 may need to generate and train multiple source models while maintaining a minimum number of such source models. In some examples of the embodiments herein, the first node 111 may create a model dictionary based on multiple source models, each satisfying certain specifications, i.e., each possessing corresponding information representing the corresponding second data. Once the first node 111 receives a request from the second node 112 in action 401, the first node 111 can select, in action 405, the source model from the model dictionary that is most relevant to the target in terms of data specifications and requirements.

[0119] The first node 111 can perform the selection in action 405 by comparing the one or more criteria (i.e., target specifications) with corresponding information characterizing the corresponding second data (i.e., source specifications). The selection can then be determined as the source specification that is closest to the one or more criteria, and its corresponding second machine learning model and corresponding second data.

[0120] The first node 111 can then proceed to forgetting, as will be described in action 409. This process is schematically depicted later in Figure 9.

[0121] By selecting a second machine learning model from multiple second machine learning models in action 405 based on the corresponding satisfaction level of the one or more criteria, the first node 111 can be enabled to select a second machine learning model that may be closer to the one or more criteria, and thus select a starting point that may require the least adaptation for adapting the second machine learning model to the first domain. This can be understood as making the process more efficient and making the resulting first machine learning model less expensive and more accurate.

[0122] Action 406 Once a second machine learning model has been selected, it is understood that not all data samples in the second dataset used to train it will equally satisfy the one or more criteria at the same level. There may be data samples that have a good match to the one or more criteria, while other data samples may have a poor match or no match at all. In action 406, the first node 111 may attempt to identify which data samples may have a poor match or no match at all.

[0123] In action 406, the first node 111 determines a first data sample in the second data whose satisfaction level for the one or more criteria does not exceed a first threshold.

[0124] The first data sample determined has an impact on the second machine learning model.

[0125] Determination can be understood as calculation, estimation, deduction or similar means, or as being obtained or received from another node.

[0126] First node 111 can use the one or more criteria (i.e., the specifications received from second node 112 in action 402) and the information obtained in action 403 (e.g., the current set of source specifications) The first node 111 can then determine which data samples in the second data might be irrelevant to the first domain, and may generate a list of data samples that may need to be erased. Action 406 can be shown as follows: .

[0127] The first node 111 can perform the determination in action 406 by checking the correspondence between the one or more criteria and the information characterizing the second data, namely the meta-information of the source and target specifications, such as the type of feature attributes, environment, data access rights, etc.

[0128] In some non-limiting examples, the determination of the first data sample in action 406 can be performed by a third module, referred to herein as the “evaluation module”, which is included in the first node 111.

[0129] By determining in action 406 that a first data sample in the second data whose satisfaction level with the one or more criteria does not exceed a first threshold, the first node 111 can identify, for example, unwanted, noisy, biased, erroneous, and privacy-sensitive samples from the fully trained second machine learning model. The first node 111 can then remove the influence of such samples from the second machine learning model without retraining it through a forgetting process, thereby minimizing the potential impact of the first data samples on the second machine learning model. That is, to "erase" their potential influence on the second machine learning model, as will be explained in action 409.

[0130] Action 407 In this document, It can represent a second machine learning model and This can represent the parameters of the second machine learning model. For example, when choosing a neural network as the underlying second machine learning model, This can include all learnable parameters. To ultimately adapt the second machine learning model to obtain the first machine learning model, it may be necessary to first compute the posterior distribution of the second machine learning model. If the second machine learning model is not a probabilistic model represented by the posterior distribution, then the first node 111 may need to approximate it.

[0131] The posterior distribution of the second machine learning can be represented as: It can be obtained from all data. The posterior distribution is learned in the middle. For probabilistic models, such as Gaussian processes, the posterior can be readily available as part of the learning process. However, for non-probabilistic models, such as neural networks, the first node 111 may need to approximate the posterior first. The first node 111 can approximate the posterior through action 407 and the subsequent action 408.

[0132] In some embodiments where the second machine learning model can be a non-probabilistic model, the first node 111 can use a non-parametric sampling process in action 407 to determine the second data sample based on the posterior function of the second machine learning model. This action 407 can also be referred to as non-parametric modeling, where the empirical distribution of these samples can approximately follow the true posterior distribution. The difference between the empirical posterior and the true posterior can be understood as the latter being expressible in terms of a probability density function; in other words, it can be understood as having a known functional form. It may be important to note that at this stage, the first node 111 may only have access to the samples and may only know the empirical posterior distribution, without knowing the parametric functional form of the posterior.

[0133] Determination can be understood as calculation, estimation, deduction or similar means, or as being obtained or received from another node.

[0134] The first node 111 can perform the determination of the second data sample in action 407 by drawing samples that approximately follow an empirical posterior distribution using a nonparametric Markov chain Monte Carlo (MCMC) sampling technique. A well-known technique that can be used is Hamiltonian Markov (HMC) chain sampling. HMC can be understood as applying the energy function... and its gradient As input, that is: , in ,and It can be understood as the first j Monte Carlo samples, and J This can be understood as the total number of Monte Carlo samples. The set of posterior samples can be represented as: .

[0135] In some non-limiting examples, the determination of the second data sample in action 407 can be performed by a fourth module, referred to herein as the “posterior module learner”, which is included in the first node 111.

[0136] By determining the second data sample in action 407 based on the posterior function of the second machine learning model, in events where the second machine learning model can be a non-probabilistic model, the first node 111 can then determine the functional form of the posterior function in the next action 408, specifically, derive the parameter approximation of the posterior function. This enables the first node 111 to subsequently continue deriving the first machine learning model in action 409 according to the embodiments herein. In other words, by performing action 407 and the next action 408, the first node 111 can make the method described herein applicable to instances where the second machine learning model can be a non-probabilistic model.

[0137] Action 408 In action 408, the first node 111 can use a parametric mixture model and determine the parametric approximation of the posterior function based on the second data.

[0138] In some embodiments where the second machine learning model can be a non-probabilistic model, the first node 111 can use a parametric mixture model and the second data sample determined in action 407 to determine a parametric approximation of the posterior function in action 408. Action 408 may also be referred to herein as parametric modeling.

[0139] The first node 111 can determine the approximate parameters of the posterior function in action 408 as follows. Given the second data sample determined in action 407. (i.e., the set of posterior samples), the first node 111 can be approximated by a parametric mixture model to represent the posterior functional form. In one implementation, a Bayesian-Gaussian mixture model can be used to... The model is as follows: , in These can be understood as the parameters of a Gaussian mixture model. These parameters can be optimized using a standard implementation of a Gaussian mixture model. Subsequently, the functional form of the posterior can be given by the predicted posterior distribution. (Unseen) The predicted posterior can then be understood as being given by the marginal distribution as follows: .

[0140] For Gaussian mixture models, the predicted distribution is... It can be understood as a mixture of student-t distributions.

[0141] In some non-restrictive examples, the determination of the parameter approximation in action 408 can be performed by the fourth module (i.e., the posterior module learner) included in the first node 111.

[0142] As previously described, by using the parametric mixture model and the determined second data sample in action 408 to determine the parametric approximation of the posterior function, in the event that the second machine learning model may be a non-probabilistic model, the first node 111 can then be enabled to continue deriving the first machine learning model in action 409 according to the embodiments herein. That is, by performing action 408, the first node 111 can make the method described herein applicable to instances where the second machine learning model may be a non-probabilistic model.

[0143] Action 409 In action 409, the first node 111 obtains the first machine learning model by performing a forgetting process of the second machine learning model. The first node 111 performs the forgetting process by reducing the influence of the determined first data sample on the second machine learning model to more than a second threshold.

[0144] The forgetting process performed in action 409 may use the second machine learning model selected in action 404 (in embodiments where the selection of action 404 may have been performed), as well as the corresponding second data and the selected second machine learning model.

[0145] In some examples of these embodiments, the first node 111 can perform the forgetting process by minimizing the impact of the determined first data sample on the second machine learning model.

[0146] In some embodiments, the forgetting process may include erasing a determined first data sample from the second data to generate third data.

[0147] In this document, A set of indices can be used to indicate data samples that may need to be erased, and It can be used to represent the corresponding erased data. This can be used to represent the remaining data samples after removing erased data samples, i.e., the third data. .

[0148] The first node 111 can use the first data sample determined in action 406 as an evaluation. And can create those marked as and Two data sets, where It can correspond to possible such as The first data sample that needs to be erased, and This can be understood as the third data, that is, the remaining data.

[0149] In some embodiments, the forgetting process may be based on a posterior function of a second machine learning model. Such embodiments may be applicable to situations where the second machine learning model can be a non-probabilistic model.

[0150] The first node 111 can also receive the current posterior distribution of all data samples. The first node 111 can then forget the posterior influence from the second machine learning model. The output of action 409 can be interpreted as the updated posterior distribution, shown as follows. The impact of the first data sample being erased can be understood as having been removed from... It was removed from the text.

[0151] In one implementation, the first node 111 can perform this determination by minimizing the upper boundary of the evidence, which can be defined as: , in This can be understood as the conditional probability distribution of the first data sample, and This can be understood as the target posterior distribution of the remaining data (i.e., the third data) after the first data sample is removed (i.e., the erased data). The minimization described herein can be applied following the description provided in [1].

[0152] The first machine learning model can be understood as being obtained to fit the third data.

[0153] In some non-limiting examples, the acquisition of the first machine learning model in action 409 can be performed by the fifth module, referred to herein as the posterior module forgetter, which is contained in the first node 111.

[0154] By obtaining the first machine learning model through a forgetting process of the second machine learning model in action 409, the first node 111 can then be enabled to customize the second machine learning model according to one or more criteria of the first domain via the model forgetting process (including posterior model learning and a posterior model forgetter, if needed). The first node 111 can then be enabled to provide the first machine learning model as an adaptive, on-demand customized model. The first node 111 can obtain the on-demand customized model in a shorter amount of time than would otherwise take to retrain the second machine learning model or train the first machine learning model from scratch, and therefore does not require high computational resources. Furthermore, the first node 111 can obtain the first machine learning model without requiring all data samples to be available. Therefore, back-and-forth communication between the first node 111 and the third node 113 may not be necessary.

[0155] Custom models can be understood as models that do not require maintenance. Therefore, a minimum number of source models may need to be maintained and continuously evolved. Consequently, the potentially expensive storage requirements and maintenance work that would otherwise exist due to a large number of models can be advantageously reduced.

[0156] This advantage can be particularly relevant in large language models (LLMs). Training such models is extremely expensive. Therefore, forgetting from an LLM rather than learning from scratch or retraining can be understood as an advantage.

[0157] The first node 111 enables the removal of unwanted, noisy, biased, erroneous, and privacy-sensitive samples from a fully trained second machine learning model without requiring retraining of the second machine learning model. Since the second machine learning model can be understood as not requiring retraining but potentially needing maintenance, this can be interpreted as providing potential energy savings.

[0158] One particular advantage of this action 409 may be that the first node 111 can enable the resolution of requirements such as EU directives on AI, which may require compliance with the right of private individuals to have their data removed from databases and systems (such as those containing ML models).

[0159] It can also be further enabled that the first node 111 sends the custom model to the second node 112, now that the custom model has no unwanted representation from the source third node 113, as will be described in the next action 410.

[0160] Action 410 In this action 410, the first node 111 provides an indication of the first machine learning model obtained.

[0161] This instruction can direct a customized posterior regarding the model parameters. .

[0162] In some embodiments where the acquisition of one or more criteria described in action 402 may have been in response to a received first indication, the indication provided in action 410 may be understood as a second indication that may be sent to the second node 112.

[0163] In a specific, non-restrictive example, the first node 111 can provide a customized posterior regarding the model parameters in action 410. Send to the second node 112.

[0164] It's worth noting that the second node, 112, can subsequently use the posterior itself. Alternatively, point estimates can be extracted from the posterior for the purpose of transfer learning. Point estimates from the posterior can be derived from the posterior mean. Provided.

[0165] In some embodiments, the determination of the first data sample in action 406, the initiation of the forgetting process in action 409, and the provision of the instruction in action 410 can be performed when the information exceeds a first threshold.

[0166] By providing instructions in action 410, the first node 111 can enable the use of the first machine learning model, for example, to make predictions in the first domain.

[0167] Action 411 In this action 411, the first node 111 can store the second instruction 411.

[0168] The accuracy of the first machine learning model can be recorded at the first node 111 along with its specifications (i.e., one or more of its criteria). By storing a second instruction in this action 411, the first node 111 can then be enabled to use the stored information for similar future requests.

[0169] Use Case Examples The embodiments described herein can be applied in various contexts. Several non-limiting examples of use cases are described below.

[0170] Diversity in Radio Base Stations: In accordance with the International Telecommunication Union's climate commitments, energy-efficient network modernization is being and will be implemented at a rapid pace in mobile networks. This modernization combines hardware modernization with the development of new software features and network optimization techniques in the RAN and during base station operation. Currently, large volumes of data are being collected from networks, and new developments in deploying to existing mobile networks are expected to cause significant changes in dataset characteristics. For example, datasets collected on base stations where power-saving features are not activated will produce significantly different energy consumption patterns compared to cases where power-saving features are activated. Furthermore, mobile operators are modernizing networks using new hardware, which may require significant effort to retrain existing predictive machine learning models using data collected from new products and software features. This is inherently energy-intensive in terms of time, computation, communication, and storage. Using the embodiments described herein, ML models can be updated to the most recent version without retraining. There is also significant diversity in base station deployments today, with varying radio capabilities depending on geographic environment and consumer profile. Moreover, these base stations can be configured with a wide range of settings. Ideally, the number of machine learning models serving these base stations should preferably be kept low to reduce complexity in lifecycle management.

[0171] Use cases of inconspicuous spatial and temporal data distribution similarity Over time, the distribution of data within a network can also drift. Therefore, a use case might involve shifting models between different time intervals. For example, data collected from base stations years ago during the rainy season, which caused microwave radio signal degradation, might be relevant today, given that this phenomenon is reproducible. Due to climate change, data collected from base stations in recent years may exhibit significant drift in distribution. In other words, the data may be constantly drifting. It's possible that data at some point in the future might be similar to data at some point in the past, allowing for the reuse of pre-existing models if available. However, it's also possible that the data may drift and may never become similar to the old data.

[0172] Multi-tenant Model: The use cases described above, or others similar, can be provided to operators by hardware providers. Some operators may agree to allow the use of their data to train the source model, but others may only allow it on the condition that they forget about their data before sharing the source model with third parties. It is then possible to use all the data to train the source model and remove knowledge from some operators before sharing it. This can be understood as advantageously enabling guidance on better methods (due to access to more data), while still removing any sensitive data before sharing.

[0173] QoE Optimization Use Cases in Open Radio Access Networks (O-RAN): One of these use cases could be QoE optimization, as it can be understood to span a growing number of applications in 5G, such as augmented reality, telephony, and 4K video, as depicted in the ITU slides. Each of these application services may have different levels of requirements, such as bit rate, continuity, and latency. For example, while bit rate and continuity may be important metrics to meet in 4K video, all three can be understood as highly important in augmented reality (AR). This can be understood as requiring many ML models to be (re)trained and (re)deployed using different distributions of the dataset. This can create considerable complexity in model management. Even within a single application service type (such as AR), there may be multiple sub-use cases, such as telephony, personal coaching, and assisted surgery. Therefore, the embodiments described herein can advantageously enable pre-trained ML models to be customized and adapted to different priority choices based on the real-time application of interest (e.g., AR versus telephony versus data download, etc.), by removing irrelevant contributions.

[0174] Context-based dynamic handover management use case: Another O-RAN use case might be context-based handover (HO) management. This use case preferably requires datasets from different entities and domains, such as historical traffic / navigation data (e.g., road conditions), radio / HO data (e.g., data about neighboring cells and their network traffic load), and datasets from the UE (e.g., received signal strength, vehicle speed), to achieve better handover. In some cases, specific entities may stop sharing data, such as due to changed regulations or technical link failures. In such cases, it may be necessary to remove dependencies from the model.

[0175] Figure 5 This is a schematic diagram depicting a non-limiting example of the components that may be included in the first node 111 of an example according to an embodiment of this document. Figure 5 As depicted in the example, the first node 111 in such an example may include a source specification module 501, a target specification module 502, a source model 503, an evaluation module 504, a posterior model learner module 505, and a posterior model forgetter module 506.

[0176] Figure 6 This is a schematic diagram illustrating non-limiting examples of aspects of a method performed by the first node 111 according to embodiments of this document. Specifically, Figure 6 An unrestricted example representation of the output of action 407 is depicted, namely the empirical posterior distribution of the samples. Figure 6 A non-restricted example of the output of action 408 is also depicted, where the empirical distribution of the posterior sample has been approximated by a posterior with a parametric function form. More specifically, panel (A) schematically depicts the distribution of the output of a given data sample. Second machine learning model (i.e., source model) In the case of approximate posterior distribution The actions involved. In some embodiments where the second machine learning model may be a non-probabilistic model, the first node 111 may, based on action 407, use a non-parametric sampling process (such as a Hamiltonian Markov chain) in the first step to sample... And the second machine learning model (i.e., the source model) The second data sample, i.e., the posterior data sample, is determined based on the posterior function of the second machine learning model, using the second data as input. The first node 111 can then determine the parametric approximation of the posterior function based on the parametric mixture model used in the second step of parameter sampling in action 408 and the second data sample determined in action 407. This is in Figure 6 The middle is indicated as the predictor of the posterior. As an example, the first node 111 can be a mixture of Gaussian distributions, which has a prediction distribution represented by a mixture of Student-t distributions. In panel (B), Figure 6 Depicting Here is an example of logistic regression, where These are the regressor parameters. The horizontal axis indicates the first dimension, i.e., the first principal component of the data. The vertical axis indicates the second dimension, i.e., the second principal component of the data. The predicted distribution of the Gaussian mixture model can be understood as a Student's t-distribution. Here... Figure 6 In the diagram, samples from the posterior are shown as contour plots, and the approximate posterior is given by the predicted distribution of the Student-t distribution, where the mode of the Student-t distribution is indicated by crossover and scaled according to its importance.

[0177] Figure 7 This is a schematic block diagram depicting non-limiting examples of aspects of a method performed by the first node 111 according to embodiments of this document. Specifically, Figure 7 The internal workings of the various components of this method are described. Given the target data specifications obtained according to action 402. and the current source data specification obtained based on action 403. The first node 111 can apply the evaluation module and infer the first data sample based on action 406, such as erasing the index set. The first node 111 can obtain the posterior of the model parameters for all data. If regarding the posterior of the model parameters If the value is unknown, then the first node 111 can apply the posterior model learner module and, based on action 408, infer the posterior value of the model parameters. Given the posterior and erase index set The first node 111 can apply the posterior model forgetting module based on action 409 and output... This can be understood as removing and erasing data. The remaining data The posterior on top. The first node 111 can then be used to customize the posterior about the model parameters based on action 410. Send to the second node 112.

[0178] Figure 7 The steps shown in the block diagram are as follows: Figure 8 Signaling diagrams are used to depict this.

[0179] Figure 8 This is a signaling diagram depicting a non-limiting example of a method performed by the first node 111 according to embodiments of this document. Figure 8In this diagram, the first node 111 is designated as a "device node," and the second node 112 is designated as a target node. The first node 111 can be implemented in a centralized or distributed manner, as shown in the signaling diagram. At point 1, the first node 111 can receive a request from the second node 112 according to action 401 to obtain a machine learning model for prediction in a first domain. That is, for a first machine learning model, in this case, a model for the first domain. Along with the request, the first node 111 can receive one or more criteria, i.e., target data specifications, according to action 402. At point 2, the first node 111 can obtain the current source data specification, i.e., the information representing the second data, based on action 403. In an alternative implementation, if the data specification of the target domain is significantly different from the source specification, the first node 111 can learn a new model tailored to the target domain instead of initiating a forgetting process. At point 3, given the target data specification... and current source data specifications The first node 111 can apply the evaluation module based on action 406 and infer the erase index set. At point 4, the first node 111 can obtain the posterior of the model parameters based on action 408. If regarding the posterior of the model parameters If the unknown value is unknown, then the first node 111 can apply the posterior model learner module and infer the posterior value of the model parameters. At 5 locations, given the posterior... and erase index set The first node 111 can apply the posterior model forgetting module based on action 409 and output... This is to remove and erase data. The remaining data The posterior. At point 6, the first node 111 can provide a customized posterior regarding the model parameters. Send to the second node 112.

[0180] Figure 9 is a schematic diagram illustrating a non-limiting example of a method performed by a first node 111 according to an embodiment of this document, where multiple source models, i.e., multiple second machine learning models, may exist. Panel (A) depicts the multiple second machine learning models as including three second machine models that can be stored in a model dictionary. According to action 404, a target specification can be compared by the first node 111 relative to the source specifications. Then, according to action 405, the source specification that is likely closest to the target specification, along with its corresponding model and data, can be selected, highlighted in the figure with a superscript *. If the target specification differs from all available source specifications by more than a predefined threshold, a new source model can be constructed for the target. This step may involve a first threshold (a user-defined threshold). This newly constructed model can optionally be added to the model dictionary. Panel (B) depicts the forgetting process that the first node 111 can then follow given the selected data and second machine learning models, such as combining Figure 7 As described.

[0181] Figure 10 This is a signaling diagram depicting a non-limiting example of a method executed by a first node 111 in an embodiment of computer system 100 that may include a plurality of third nodes 113, and wherein each of the first node 111, the second node 112, and the third node 113 may be an NWDAF. In this particular example, the plurality of third nodes 113 may include N third nodes 113. A first third node 113 is represented as NWDAF source 1, a second third node 113 as NWDAF source 2, and an Nth third node 113 as NWDAF source N. Figure 10 Any of the first node 111, the second node 112, and the plurality of third nodes 113 depicted can be a different Model Training Logical Function (MTLF), where they can store data in different regions of the NWDAF. Figure 10In a non-limiting example, the second node 112 is depicted as an NWDAF in the target domain that has data but no model for making predictions. At point 1, the second node 112 (referred to in this example as the "NWDAF target") can query the corresponding NRF 1001 for the discovery of available NWDAFs by sending a discovery request indicating an identifier (ID) for analysis. At point 2, the NRF 1001 sends a discovery response to the second node 112 indicating the first node 111. The first node 111 is depicted as an NWDAF model learning / forgetting agent that can have access to all models existing in different NWDAFs (e.g., regions). At point 3, the first node 111 can receive a first instruction from the second node 112 indicating a request for a first machine learning model for making predictions in the first domain, according to action 401, and at point 4, the first node 111 can request one or more criteria, i.e., specifications of the target data, from the second node 112, which the first node 111 can then obtain according to action 402. At points 5, 6, and 7, the first node 111 can request additional data specifications from multiple third nodes 113, namely those potential source NWDAFs that may have been trained on history or have already trained models or multiple models. The first node 111 can then obtain the corresponding information representing the corresponding second data, i.e., the source data specifications, from each third node 113 according to action 403. At point 8, based on the data specifications, the first node 111 can perform an evaluation according to action 404 and determine, according to action 405, which source model might be useful for the target domain, and request the corresponding NWDAF source to send the model and the data (or representation) not expected to be represented in the target model. At point 9, the first node 111 can send the request to the third node 113 that holds the selected second machine learning model, and at point 10, the third node 113 of the NWDAF source 2 can send the pre-trained second machine learning model to the first node 111. At point 11, the first node 111 can customize the source model according to the target model via a model forgetting process including posterior model learning of action 409 and posterior model forgetting of action 409 (at point 12). At point 13, the first node 111 can send the customized first machine learning model (now free of unwanted representations from the source NWDAF) to the target NWDAF (i.e., the second node 112), consistent with action 410. At point 14, the second node 112 can send an indication of the accuracy of the first machine learning model to the first node 111, and at point 15, according to action 411, the accuracy of the target model, along with its specification, can be recorded at the first node 111 for use in similar future requests.

[0182] O-RAN implementation The embodiments described herein should be understood as not specific to any particular use case, and many potential Open Radio Access Network (O-RAN) use cases (e.g., QoE optimization, context-based dynamic handover management for Vehicle-to-Everything (V2X) communication, etc.) can support the embodiments described herein. In particular, the embodiments described herein may relate to use cases and overall architecture in O-RAN Working Group 1, use case detailed specifications, technical specifications O-RAN Alliance, 2023, v. 12. Figure 4 An extension of the existing ORAN flow graph described in .4.3-1, which enables the removal of contributions from a specific data provider (e.g., an application server in this case) when the QoE model may be jointly trained by multiple application servers. Figure 11 is a signaling diagram depicting a non-limiting example of such a flow graph, which is extended according to embodiments herein. In panel b), actions corresponding to the extensions according to embodiments herein are enclosed in dashed boxes. Refer to the O-RAN Consortium White Paper "O-RAN Use Cases and Deployment Scenarios: Towards Open and Intelligent RAN" published in February 2020. Figure 1The overall O-RAN logical architecture described herein, in QoE optimization use cases, may involve interfaces such as O1 between the non-real-time RAN Intelligent Controller (non-RTTRIC) and the Open-eNB (O-eNB), A1 between the near-real-time RAN Intelligent Controller (RIC) and the non-RT RIC, and E2 between the near-RT-RIC and the Open Centralized Unit Control Plane (O-CU-CP). In Figure 11, the first node 111 may be included in the non-RT RIC and co-located with the second node 112. The non-RT RIC may be included in the service management and orchestration system 1102 along with the collector 1101. The computer system 100 may further include an O-RAN 1103, which includes a near-RTTRIC 1104 and an Open Centralized Unit / Open Distributed Unit (O-CU / O-DU) 1105. The computer system 100 may further include multiple third nodes 113 as external application servers 1106: application server #1, application server #2, and application server #N. Starting from panel a), at point 1, O-CU / O-DU 1105 can trigger data collection from collector 1101. At point 2, collector 1101 can trigger retrieval of collected data from non-RT RICs 111 and 112. This, in turn, can trigger non-RT RICs 111 and 112 to retrieve application data from application server #1 at point 3, from application server #2 at point 4, and from application server #N at point 5. At point 6, non-RT RICs 111 and 112 can trigger an ML workflow by training an ML model. At point 7, non-RT RICs 111 and 112 can deploy an internal ML model. At point 8, non-RT RICs 111 and 112 can deploy an AI / ML model near RT-RIC 1104. Next, a performance evaluation and optimization phase can be triggered. At point 9, O-CU / O-DU 1105 can trigger data collection from collector 1101 via the O1 interface. <o1>Data collection can be triggered at point 10 by collector 1101 to retrieve data from non-RT RICs 111 and 112. Continuing in panel b), non-RT RICs 111 and 112 can then begin performance monitoring and evaluation at point 11. In some cases, a specific entity, in this example, application server #2, may stop sharing data, for example due to changed regulations or technical link failures. In that case, it may be necessary to remove dependencies from the model. At point 12, the link between application server #2 and non-RT RICs 111 and 112 may contain additional information (besides a dataset that may be intended for ML model training, for learning and forgetting, and may have domain-specific characteristics) regarding the intention to remove data / contributions. Once this intention is registered, non-RT RICs 111 and 112 can trigger the forgetting function at point 13 to remove the contribution of that specific client from the source model, consistent with action 409. Specific client contributions can be identified by the sample_id_vector. At point 14, non-RT RICs 111 and 112 can continue to monitor and evaluate the new version of the model obtained after forgetting. At point 15, non-RT RICs 111 and 112 can perform model version management. This can be understood as performing checkpoints in memory to save forgotten models for later reuse. The model can be stored along with additional corresponding details of the processing steps applied to the previous version of the model. For example, it can explain which sample IDs may have been deleted, and the performance, size, etc., of the new model. At point 16, according to action 410, non-RT RICs 111 and 112 can then... <o1>> or < <a1>The interface shares updated AI / ML models with the near-RT RIC 1102. Information about intent or incentive flags (i.e., the willingness to contribute or not contribute) can be shared from contributors / clients (i.e., application servers in the ORAN architecture) to non-RT RICs 111 and 112. From non-RT RICs 111 and 112, information about updated models and / or possibly notifications (e.g., events) regarding the existence of updated models are sent to external consumers / producers (e.g., application servers and / or to near-RT RICs).

[0183] As a summary overview of the foregoing, the embodiments herein can be understood to relate to a first node 111 that can receive a request for a model from a target node. This first node 111 can use target-specific machine forgetting to create a customized version of the source model, the purpose of which is to ensure that the customized model conforms to the data guidelines and specifications of the target domain. In a specific example of the embodiments herein, this method or the first node 111 may be implemented at an ML Lifecycle Management (LCM) unit, such as in a 3GPP NWDAF.

[0184] Certain embodiments described herein may provide one or more of the following technical advantages. These embodiments can be understood as enabling the provision of adaptive, on-demand customized models. Customized models can be understood as not requiring maintenance. Therefore, only a minimal number of source models may need to be maintained and continuously evolved. This is particularly important considering that these embodiments can reduce the expensive storage requirements and maintenance work associated with existing methods due to a large number of models.

[0185] This benefit can be understood as particularly advantageous in large language models (LLMs). Such models can be understood as being very expensive to train. Therefore, it can be understood as advantageous to forget from an LLM rather than learn or retrain from scratch.

[0186] The embodiments described in this paper can further enable energy savings because they allow the removal of unwanted, noisy, biased, erroneous, and privacy-sensitive samples from a fully trained model without requiring retraining. This can be understood as the source model being understood as not needing retraining, but perhaps only needing maintenance.

[0187] The embodiments described herein can further enable the resolution of new EU directives regarding AI, where, for example, the right of individuals to have their data removed from databases and systems may be relevant.

[0188] Figure 12 The first node 111 is described as potentially capable of performing... Figure 4 and / or Figures 5-1 An example of the arrangement of the method described in section 1. The first node 111 can be understood as being used to handle the first machine learning model. The first node 111 is configured to operate within computer system 100.

[0189] This document includes several embodiments. It should be noted that the examples herein are not mutually exclusive. Where applicable, one or more embodiments may be combined. For the sake of simplicity, not all possible combinations have been described. Components from one embodiment may be assumed by default to exist in another embodiment, and how these components can be used in other exemplary embodiments will be apparent to those skilled in the art. The detailed descriptions of some of the following contents correspond to the same references provided above regarding the actions described for the first node 111, and therefore will not be repeated here. For example, the first machine learning model may also be referred to as the target model.

[0190] The first node 111 is configured to obtain one or more criteria to be satisfied by first data, which is to be used as input to obtain a first machine learning model for making predictions in a first domain. A second machine learning model has been trained with second data from a second domain to make predictions in the second domain.

[0191] The first node 111 is also configured to obtain information configured to represent the second data.

[0192] The first node 111 is further configured to determine a first data sample in the second data whose satisfaction level with the one or more criteria does not exceed a first threshold. The determined first data sample is configured to have an impact on the second machine learning model.

[0193] The first node 111 is further configured to obtain the first machine learning model by performing a forgetting process of the second machine learning model, which is performed by reducing the influence of the determined first data sample on the second machine learning model to more than a second threshold.

[0194] The first node 111 is further configured to provide an instruction that is configured to indicate the first machine learning model to be obtained.

[0195] In some embodiments, the forgetting process may be configured to include erasing a first data sample, configured to be determined, from the second data to generate third data. A first machine learning model may be configured to be obtained and fitted to the third data.

[0196] In some embodiments, the forgetting process can be configured to be based on a posterior function of a second machine learning model, and the first node 111 can be further configured to use a parametric mixture model and determine a parametric approximation of the posterior function based on second data configured to be obtained.

[0197] In some embodiments, the second machine learning model can be configured as a non-probabilistic model prior to the forgetting process, and the first node 111 can be further configured to determine the second data sample using a non-parametric sampling process based on the posterior function of the second machine learning model. The determination of the parameter approximation of the posterior function can be configured to use a parametric mixture model and the second data sample configured to be determined.

[0198] In some embodiments, the first node 111 may be further configured to determine whether the level of satisfaction of the information configured to characterize the second data with the one or more criteria may exceed a first threshold. The determination of the first data sample, the initiation of the forgetting process, and the provision of instructions may be configured to be performed under the condition that the information may exceed the first threshold.

[0199] In some embodiments, the second machine learning model may be one of a plurality of second machine learning models that have been trained with corresponding second data from a second domain to make corresponding predictions in the second domain. In such embodiments, obtaining information may be configured to include obtaining corresponding information configured to characterize the corresponding second data, and the determination may be further configured to include determining whether the corresponding information's corresponding satisfaction level with the one or more criteria is likely to exceed a first threshold. The information that may exceed the first threshold may be configured to be a set of corresponding information.

[0200] In some embodiments, the first node 111 may be further configured to select a second machine learning model from a plurality of second machine learning models. This selection may be configured based on a corresponding satisfaction level of the one or more criteria.

[0201] In some embodiments, the first node 111 may be further configured to have at least one of the following two configurations.

[0202] In some embodiments, the first node 111 may be further configured to receive a first instruction from a second node 112 configured to operate in the computer system 100. The first instruction may be configured to indicate a request for a first machine learning model for making predictions in a first domain. The acquisition of the one or more criteria may be configured in response to the first instruction being received. The instruction may be configured to be a second instruction configurable to be sent to the second node 112.

[0203] In some embodiments, the first node 111 may be further configured to store a second instruction.

[0204] In some embodiments, the corresponding information configured to characterize the corresponding second data can be configured to be obtained from a plurality of third nodes 113, the computer system 100 is configured to be a fifth-generation network, and each of the first node 111, the second node 112 and the plurality of third nodes 113 can be configured to be a network data analysis function.

[0205] The embodiments described in this paper, in the first node 111, can be implemented by one or more processors, such as... Figure 12 The processing circuitry 1201 in the first node 111 depicted herein, along with computer program code for performing the functions and actions of the embodiments herein, is described herein. The processor used herein can be understood as a hardware component. The program code mentioned above can also be provided as a computer program product, for example, in the form of a data carrier carrying the computer program code for performing the embodiments herein when loaded into the first node 111. One such carrier could be a CD-ROM. However, other data carriers such as memory sticks are also feasible. Furthermore, the computer program code can be provided as pure program code on a server and downloaded to the first node 111.

[0206] The first node 111 may further include a memory 1202 comprising one or more memory units. The memory 1202 is configured to store acquired information, data, configurations, schedules, and applications, etc., for execution of the methods herein when executed in the first node 111.

[0207] In some embodiments, the first node 111 may receive information via a receiving port 1203 from, for example, another structure in the second node 112, any of a plurality of third nodes 113, radio network node 130, device 140, and / or computer system 100. In some embodiments, the receiving port 1203 may be connected, for example, to one or more antennas in the first node 111. In other embodiments, the first node 111 may receive information via the receiving port 1203 from another structure in the computer system 100. Since the receiving port 1203 can communicate with the processing circuitry 1201, the receiving port 1203 can then transmit the received information to the processing circuitry 1201. The receiving port 1203 may also be configured to receive other information.

[0208] The processing circuit 1201 in the first node 111 can be further configured to transmit or send information via a transmission port 1204 to, for example, a second node 112, any of a plurality of third nodes 113, a radio network node 130, a device 140, and / or another structure in the computer system 10, which can communicate with the processing circuit 1201 and the memory 1202.

[0209] The first node 111 can be configured to execute in combination. Figure 4 and / or Figures 5-1 1. Any action described, for example, by means of the processing circuitry 1201 within the first node 111 configured to perform any such action.

[0210] Furthermore, in some embodiments, the different units included in the first node 111 may be configured to perform the different actions described above, and may be implemented as one or more applications running on one or more processors (such as processing circuitry 1201).

[0211] Those skilled in the art will also appreciate that the aforementioned units configured to perform different actions within the first node 111 can refer to a combination of analog and digital circuitry, and / or one or more processors configured with software and / or firmware (e.g., stored in memory), which perform the functions described above when executed by one or more processors such as processing circuitry 1201. One or more of these processors, along with other digital hardware, may be included in a single application-specific integrated circuit (ASIC), or multiple processors and various digital hardware may be distributed across several separate components, whether individually packaged or assembled into a system-on-a-chip (SoC).

[0212] Therefore, the method according to the embodiment described herein with respect to first node 111 can be implemented by means of a computer program product 1205, which includes instructions (i.e., software code portions) that, when executed on at least one processing circuit 1201, cause the at least one processing circuit 1201 to perform the actions described herein performed by first node 111. The computer program product 1205 can be stored on a computer-readable storage medium 1206. The computer-readable storage medium 1206 on which the computer program 1205 is stored may include instructions that, when executed on at least one processing circuit 1201, cause the at least one processing circuit 1201 to perform the actions described herein performed by first node 111. In some embodiments, the computer-readable storage medium 1206 may be a non-transitory computer-readable storage medium, such as a CD-ROM or Memory Stick. In other embodiments, the computer program product 1205 may be stored on a carrier containing the computer program 1205 just described, wherein the carrier is one of electronic signals, optical signals, radio signals, or the computer-readable storage medium 1206 as described above.

[0213] The first node 111 may include a communication interface (configured to facilitate) or an interface unit (for facilitating) communication between the first node 111 and other nodes or devices (e.g., the second node 112, the third node, the radio network node 130, the device 140, and / or another structure in the computer system 100). The interface may, for example, include a transceiver configured to transmit and receive radio signals over an air interface according to suitable standards.

[0214] In other embodiments, the first node 111 may include a radio circuit 1207, which may include, for example, a receive port 1203 and a transmit port 1204.

[0215] Radio circuit 1207 can be configured to establish and maintain wireless connections with at least one of the following structures: second node 112, third node, radio network node 130, device 140, and / or computer system 100. In this document, the circuit can be understood as a hardware component.

[0216] Therefore, the embodiments herein also relate to a first node 111 operable in a computer system 100. The first node 111 may include processing circuitry 1201 and a memory 1202 containing instructions executable by the processing circuitry 1201, thereby further operable the first node 111 to perform the instructions herein relating to the first node 111 (e.g., in...). Figure 4 and / or Figures 5-1 The action described in 1).

[0217] References 1. [Nguyen 2020]Nguyen, QP, Low, B.,&Jaill et, P. (2020).Variational Bayesian Unlearning. ArXiv, abs / 2010.12883.F 2. [Ben 2018] Ban, Y., Alameda-Pineda, X., Girin, L., & Horaud, R. (2018). Variational Bayesian Inference for Audio-Visual Tracking of Multiple Speakers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43, 1761-1776. This evolution is aimed at intelligent power settings at ICT sites—Ericsson 3. [Pan 2010]Pan, S.J.,&Yang, Q. (2010). A Survey on TransferLearning. IEEE Transactions on Knowledge and Data Engineering, 22, 1345-1359。

[0218] 4. [Yan 2020]Yan, X., Acuna, D.,&Fidler, S. (2020). Neural DataServer: A Large-Scale Search Engine for Transfer Learning Data. 2020 IEEE / CVFConferenceon Computer Vision and Pattern Recognition (CVPR), 3892-3901。

[0219] 5. [ITU]International Telecommunication Union (ITU) | Department ofEconomic and Social Affairs。

Claims

1. A computer-implemented method executed by a first node (111), the method being used to process a first machine learning model, the first node (111) operating in a computer system (100), the method comprising: - Obtain (402) one or more criteria to be satisfied by first data, which is to be used as input to obtain a first machine learning model for making predictions in a first domain, wherein a second machine learning model has been trained with second data from a second domain to make predictions in the second domain. - Obtain (403) information characterizing the second data, - Determine (406) a first data sample in the second data whose satisfaction level for the one or more criteria does not exceed a first threshold, wherein the determined first data sample has an impact on the second machine learning model, and - The first machine learning model is obtained by performing a forgetting process of the second machine learning model, wherein the forgetting process is performed by reducing the influence of the determined first data sample on the second machine learning model to exceed a second threshold, and - Provide (410) an instruction for the first machine learning model obtained.

2. The method according to claim 1, wherein, The forgetting process includes erasing the determined first data sample from the second data to generate third data, wherein the first machine learning model is obtained to fit the third data.

3. The method according to any one of claims 1-2, wherein, The forgetting process is based on the posterior function of the second machine learning model, and the method further includes: - Using a parametric mixture model and based on the obtained second data, determine the parametric approximation of the posterior function (408).

4. The method according to claim 3, wherein, The second machine learning model is a non-probabilistic model prior to the forgetting process, and the method further includes: - A second data sample is determined (407) using a nonparametric sampling process based on the posterior function of the second machine learning model, wherein the parameters of the posterior function are approximately determined (408) using the parametric mixture model and the determined second data sample.

5. The method according to any one of claims 1-4, wherein, The method further includes: - Determine (404) whether the information characterizing the second data satisfies the level of the one or more criteria exceeding the first threshold, and wherein the determination (406) of the first data sample, the initiation (409) of the forgetting process, and the provision (410) of the indication are performed under the condition that the information exceeds the first threshold.

6. The method according to claim 5, wherein, The second machine learning model is one of a plurality of second machine learning models that have been trained with corresponding second data from the second domain to make corresponding predictions in the second domain, wherein the acquisition (403) of the information includes acquiring corresponding information characterizing the corresponding second data, and wherein the determination (404) further includes determining whether the corresponding information satisfies the corresponding level of the one or more criteria exceeding the first threshold, and wherein the information exceeding the first threshold is a set of the corresponding information.

7. The method of claim 6, further comprising: - Select (405) a second machine learning model from the plurality of second machine learning models, the selection (405) being based on the corresponding satisfaction level of the one or more criteria.

8. The method according to any one of claims 1-7, further comprising at least one of the following: - Receive (401) a first instruction from a second node (112) operating in the computer system (100), the first instruction indicating a request for the first machine learning model for making predictions in the first domain, wherein, The acquisition (402) of one or more criteria is in response to a received first indication, wherein the indication is a second indication sent to the second node (112), and - Store (411) the second instruction.

9. The method according to any one of claims 6-8, wherein, The corresponding information representing the corresponding second data is obtained from a plurality of third nodes (113), wherein the computer system (100) is a fifth-generation network, and wherein each of the first node (111), the second node (112) and the plurality of third nodes (113) is a network data analysis function.

10. A first node (111) for processing a first machine learning model, the first node (111) being configured to operate in a computer system (100), the first node (111) being further configured to: - Obtain one or more criteria to be satisfied by first data, which is used as input to obtain a first machine learning model for making predictions in a first domain, wherein a second machine learning model has been trained with second data from a second domain to make predictions in the second domain. - Obtain information configured to characterize the second data. - Determine a first data sample in the second data whose satisfaction level with the one or more criteria does not exceed a first threshold, wherein the first data sample is configured to be determined to have an impact on the second machine learning model, and - The first machine learning model is obtained by performing a forgetting process on the second machine learning model, said forgetting process being performed by reducing the influence of the determined first data sample on the second machine learning model to exceed a second threshold, and - Provides an instruction configured to indicate the first machine learning model obtained by multiplying the configuration.

11. The first node (111) according to claim 10, wherein, The forgetting process is configured to include erasing a first data sample, configured to be determined, from the second data to generate third data, wherein the first machine learning model is configured to be obtained to fit the third data.

12. The first node (111) according to any one of claims 10-11, wherein, The forgetting process is configured to be based on the posterior function of the second machine learning model, and wherein the first node (111) is further configured to: - Using a parametric mixture model and based on the second data configured to be obtained, determine the parametric approximation of the posterior function.

13. The first node (111) according to claim 12, wherein, The second machine learning model is configured as a non-probabilistic model prior to the forgetting process, and wherein the first node (111) is further configured to: - A second data sample is determined using a nonparametric sampling process based on the posterior function of the second machine learning model, wherein the determination of the parameters of the posterior function is configured to use the parametric mixture model and the second data sample to be determined.

14. The first node (111) according to any one of claims 10-13, wherein, The first node (111) is further configured as follows: - Determine whether the information configured to characterize the second data satisfies the level of one or more criteria exceeding a first threshold, and wherein the determination of the first data sample, the initiation of the forgetting process, and the provision of the indication are configured to be performed under the condition that the information exceeds the first threshold.

15. The first node (111) according to claim 14, wherein, The second machine learning model is one of a plurality of second machine learning models that have been trained with corresponding second data from the second domain to make corresponding predictions in the second domain, wherein the acquisition of the information is configured to include acquiring corresponding information configured to characterize the corresponding second data, and wherein the determination is further configured to include determining whether the corresponding information satisfies a corresponding level of the one or more criteria exceeding a first threshold, and wherein the information exceeding the first threshold is configured as a set of the corresponding information.

16. The first node (111) according to claim 15 is further configured as follows: - Select a second machine learning model from the plurality of second machine learning models, the selection being configured based on the corresponding satisfaction level of the one or more criteria.

17. The first node (111) according to any one of claims 10-16 is further configured to perform at least one of the following: - Receive a first instruction from a second node (112) configured to operate in the computer system (100), the first instruction being configured to indicate a request for the first machine learning model for making predictions in the first domain, wherein, The acquisition of one or more criteria is configured to respond to a first indication configured to be received, wherein the indication is configured to be a second indication configured to be sent to the second node (112), and - Store the second instruction.

Images