EP4690715A1

EP4690715A1 - Machine learning model feature selection in a communication network

Info

Publication number: EP4690715A1
Application number: EP24781392.6A
Authority: EP
Inventors: Athanasios KARAPANTELAKIS; Abdulrahman ALABBASI; Konstantinos Vandikas
Original assignee: Telefonaktiebolaget LM Ericsson AB
Current assignee: Telefonaktiebolaget LM Ericsson AB
Priority date: 2023-03-31
Filing date: 2024-02-28
Publication date: 2026-02-11
Also published as: WO2024205464A1; CN120898407A

Abstract

Embodiments of the present disclosure provide machine learning model feature selection in a communication network The method comprises: in response to a feature selection trigger for a first machine learning model, determining a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task based on a current set of input features, the current set of input features being different from the target set of input features; and causing a second machine learning model to be provisioned to perform the analytic task based on the determined target set of input features. In this way, a machine learning model can be provisioned with an optimized set of input features that is applicable to the current network context and providing an acceptable level of model performance.

Description

MACHINE LEARNING MODEL FEATURE SELECTION IN A COMMUNICATION NETWORK

FIELD

[0001] Embodiments of the present disclosure generally relate to the field of telecommunication and in particular, to machine learning model feature selection in a communication network.

BACKGROUND

[0002] In the telecommunication industry, artificial intelligence/machine learning (AI/ML) models have been employed in communication systems to improve the performance of communications systems. A machine learning model may be considered as logic or an algorithm that can be “trained” by data and human expert input as examples to replicate a decision an expert would make when provided that same information. In the communication scenarios, different machine learning models may be provided for various network analytic tasks such as network congestion prediction, malicious behavior detection, radio resource allocation/scheduling, device positioning, and/or the like. A machine learning model for a specific analytic task is configured to process a set of input features that are believed to be relevant to the analytic task and provide a prediction output for the analytic task.

[0003] Currently, the features used to create a prediction per analytics task are known a priori to a model provider in a communication network. In particular, it has been defined in some 3^rd Generation Partnership Project (3GPP) communication specifications a number of use cases, each use case having a pre-defined set of input features and the prediction output. For example, in a use case of slice load level related network data analytics, the input features to be collected are defined as including user equipment (UE) registered in a network slice or network slice instance, packet data unit (PDU) session established on a network slice or network slice instance, load of network functions (NFs) associated with a network slice instance, UE registers/de-registers to a network slice or network slice instance, a number of UEs served by an Access and Mobility Management Function (AMF), PDU session established/released on a network slice, current number of UEs registered in a network slice, current number of PDU sessions established in a network slice, and load of NFs associated to network slice instance.

[0004] However, given the dynamicity of communication networks and the external factors affecting them (e.g., behavior of terminal devices such as traffic, cardinality and mobility patterns, environmental effects, power supply issues, and the like), it is not ideal to use a static collection of input features for an analytic task in the communication networks. SUMMARY

[0005] As machine learning models relying on more relevant input features can provide more accurate predictions, the exact input features that are used for different use cases or analytic tasks may differ. It is thus expected to optimize input features of the machine learning models to the condition of individual communication networks in an efficient way. Embodiments of the present disclosure provide a solution for machine learning model feature selection, to explore input feature selection for machine learning models deployed in a communication network.

[0006] The feature selection in accordance with the embodiments of the present disclosure allows to explore the input feature space for an analytic task in an intelligent and dynamic or realtime way. As a result, a machine learning model can be provisioned with an optimized set of input features that is applicable to the current network context and providing an acceptable level of model performance. In addition, through the feature selection process, it can save bandwidth resources by reducing the number of input features and thus the amount of data needed to be collected for inferencing. Potentially, it can also save computational resources by reducing the number of model layers as a result of the lower number of input features. On the other hand, it allows for better performance of provisioned models, as a set of input features are selected to provide more accurate predictions than the baseline set.

[0007] In a first aspect of the present disclosure, a computer-implemented method for machine learning model feature selection in a communication network is proposed. The method comprises: in response to a feature selection trigger for a first machine learning model, determining, by a network node, a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task based on a current set of input features, the current set of input features being different from the target set of input features; and causing, by the network node, a second machine learning model to be provisioned to perform the analytic task based on the determined target set of input features.

[0008] In some embodiments of the first aspect, the feature selection trigger is initiated based on at least one of the following: an expiration of a timer started from a last feature selection trigger, the number of analytic requests for the analytic task exceeding a threshold number, an evaluation level of the first machine learning model being below a threshold level, detection of a change in at least one feature for the first machine learning model, detection of a change in context information related to the analytic task, reception of a feature selection request from a network manager, detection of a new feature for enhancing the first machine learning model, or detection of an update of a use case context. [0009] In some embodiments of the first aspect, the target set of input features are determined further based on a reference set of input features for the analytic task.

[0010] In some embodiments of the first aspect, the reference set of input features are determined to allow a baseline evaluation level of the first machine learning model, or the reference set of input features comprises the current set of input features for the first machine learning model.

[0011] In some embodiments of the first aspect, determining the target set of input features comprises: determining at least one candidate set of input features from the reference set of input features; determining, given a condition of the context information, at least one predicted evaluation level of a machine learning model with the at least one candidate set of input features as a model input; and selecting the target set of input features from the at least one candidate set of input features based on the at least one predicated evaluation level.

[0012] In some embodiments of the first aspect, the target set of input features are determined through at least one of the following: a reinforcement learning, RL, process, a contextual bandit process, a Monte-Carlo tree search process, or a collaborative filtering process using a recommendation engine.

[0013] In some embodiments of the first aspect, the communication network comprises a core network, CN, and the network node comprises a model training logical function, MTLF, in a network data and analytics function, NWDAF, or a separate network function outside of the MTLF.

[0014] In some embodiments of the first aspect, the feature selection trigger is received from an analytical logical function, AnLF, the method further comprising: notifying the AnLF of the target set of input features.

[0015] In some embodiments of the first aspect, the communication network comprises an open radio access network, O-RAN, and the network node is deployed at a non-real time RAN intelligence controller, non-RT RIC, or at a near RT RIC within the O-RAN.

[0016] In some embodiments of the first aspect, causing the second machine learning model to be provisioned comprises: causing the first machine learning model to be retrained with the target set of input features, to obtain the second machine learning model; causing a trained machine learning model to be reconfigured with the target set of input features as the second machine learning model; or causing the second machine learning model to be selected from a plurality of candidate trained machine learning models. [0017] In some embodiments of the first aspect, the context information comprises at least one of the following: context information received from one or more further network nodes within the communication network, or context information received from a third party.

[0018] In some embodiments of the first aspect, the target set of input features is used as a recommendation for determining a set of input features for the second machine learning model.

[0019] In a second aspect, a network node is proposed. The network node comprises: at least one processor; at least one memory coupled to the at least one processor, the at least one memory comprising instructions that when executed by the at least one processor implement a method for machine learning model feature selection in a communication network, the method comprising: in response to a feature selection trigger for a first machine learning model, determining a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task, the current set of input features being different from the target set of input features; and causing a second machine learning model to be provisioned to perform the analytic task based on the target set of input features.

[0020] In a third aspect, an apparatus is proposed. The apparatus comprises means for in response to a feature selection trigger for a first machine learning model, determining a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task, the current set of input features being different from the target set of input features; and means for causing a second machine learning model to be provisioned to perform the analytic task based on the target set of input features.

[0021] In a fourth aspect, a computer readable medium is proposed. The computer readable medium has instructions stored thereon, the instructions when executed by at least one processor causing the at least one processor to perform the method according to any of the embodiments of the first aspect.

[0022] In a fifth aspect, a computer program is proposed. The computer program comprises instructions which, when executed on at least one processor, cause the at least one processor to carry out the method according to any of the embodiments of the first aspect.

[0023] It is to be understood that the Summary section is not intended to identify key or essential features of embodiments of the present disclosure, nor is it intended to be used to limit the scope of the present disclosure. Other features of the present disclosure will become easily comprehensible through the following description.

BRIEF DESCRIPTION OF THE DRAWINGS [0024] Through the more detailed description of some example embodiments of the present disclosure in the accompanying drawings, the above and other objects, features and advantages of the present disclosure will become more apparent, where:

[0025] FIG. 1 illustrates a schematic diagram for a simplified communication environment in which embodiments of the present disclosure can be implemented;

[0026] FIG. 2A illustrates a schematic diagram for a communication environment with a feature selection function in accordance with some embodiments of the present disclosure;

[0027] FIG. 2B illustrates a schematic diagram for a further communication environment with a feature selection function in accordance with some further embodiments of the present disclosure;

[0028] FIG. 3 illustrates a signaling chart for feature selection within the communication environment in FIG. 2A or FIG. 2B in accordance with some further embodiments of the present disclosure;

[0029] FIG. 4A illustrates a schematic diagram for a core network (CN) with the feature selection function for a machine learning model in accordance with some embodiments of the present disclosure;

[0030] FIG. 4B illustrates a schematic diagram for a CN with the feature selection function for a machine learning model in accordance with some further embodiments of the present disclosure;

[0031] FIG. 5 illustrates a signaling chart for feature selection among network nodes within the CN in FIG. 4A or FIG. 4B in accordance with some embodiments of the present disclosure;

[0032] FIG. 6A illustrates a schematic diagram for an open radio access network (O-RAN) with the feature selection function for a machine learning model in accordance with some embodiments of the present disclosure;

[0033] FIG. 6B illustrates a schematic diagram for an O-RAN with the feature selection function for a machine learning model in accordance with some further embodiments of the present disclosure;

[0034] FIG. 7 illustrates a schematic diagram for an environment with the feature selection function provided as a standalone function in accordance with some embodiments of the present disclosure;

[0035] FIG. 8 illustrates a flowchart of a method for machine learning model feature selection in accordance with some embodiments of the present disclosure;

[0036] FIG. 9 illustrates an example of a communication system in accordance with some embodiments of the present disclosure; and [0037] FIG. 10 illustrates a simplified block diagram of a network node that is suitable for implementing embodiments of the present disclosure.

[0038] Throughout the drawings, the same or similar reference numerals represent the same or similar element.

DETAILED DESCRIPTION

[0039] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. The embodiments set forth below represent information to enable those skilled in the art to practice the embodiments and illustrate the best mode of practicing the embodiments. Upon reading the following description in light of the accompanying figures, those skilled in the art will understand the concepts of the disclosure and will recognize applications of these concepts not particularly addressed herein. It should be understood that these concepts and applications fall within the scope of the disclosure.

[0040] Generally, all terms used herein are to be interpreted according to their ordinary meaning in the relevant technical field, unless a different meaning is clearly given and/or is implied from the context in which it is used. All references to a/an/the element, apparatus, component, means, step, etc. are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. The steps of any methods disclosed herein do not have to be performed in the exact order disclosed, unless a step is explicitly described as following or preceding another step and/or where it is implicit that a step must follow or precede another step. Any feature of any of the embodiments disclosed herein may be applied to any other embodiment, wherever appropriate. Likewise, any advantage of any of the embodiments may apply to any other embodiments, and vice versa. Other objectives, features, and advantages of the enclosed embodiments will be apparent from the following description.

[0041] As used herein, the term “includes” and its variants are to be read as open terms that mean “includes, but is not limited to.” The term “based on” is to be read as “at least in part based on.” The term “one embodiment” and “an embodiment” are to be read as “at least one embodiment.” The term “another embodiment” is to be read as “at least one other embodiment.” The terms “first,” “second,” and the like may refer to different or same objects. Other definitions, explicit and implicit, may be included below.

[0042] As used herein, the term “terminal device” refers to any device having wireless or wired communication capabilities. Examples of the terminal device include, but not limited to, user equipment (UE), personal computers, desktops, mobile phones, cellular phones, smart phones, personal digital assistants (PDAs), portable computers, tablets, wearable devices, internet of things (loT) devices, Ultra-reliable and Low Latency Communications (URLLC) devices, Internet of Everything (loE) devices, machine type communication (MTC) devices, devices on vehicle for V2X communication where X means pedestrian, vehicle, or infrastructure/network, devices for Integrated Access and Backhaul (IAB), Space home vehicles or Air borne vehicles in Non-terrestrial networks (NTN) including Satellites and High Altitude Platforms (HAPs) encompassing Unmanned Aircraft Systems (UAS), extended Reality (XR) devices including different types of realities such as Augmented Reality (AR), Mixed Reality (MR) and Virtual Reality (VR), the unmanned aerial vehicle (UAV) commonly known as a drone which is an aircraft without any human pilot, devices on high speed train (HST), or image capture devices such as digital cameras, sensors, gaming devices, music storage and playback appliances, or Internet appliances enabling wireless or wired Internet access and browsing and the like. The “terminal device” can further have “multicast/broadcast” feature, to support public safety and mission critical, V2X applications, transparent IPv4/IPv6 multicast delivery, IPTV, smart TV, radio services, software delivery over wireless, group communications and loT applications. It may also incorporate one or multiple Subscriber Identity Module (SIM) as known as Multi-SIM. The term “terminal device” can be used interchangeably with a UE, a mobile station, a subscriber station, a mobile terminal, a user terminal or a wireless device.

[0043] As used herein, the term “network node” may also be referred to as a network function (NF), a network entity, or a network device, and refers to a physical, virtual or hybrid function or entity which is deployed at a network side and provides one or more services to clients/consumers. For example, an NF may be arranged at a device in a radio access network (RAN) or a core network (CN) of a communication system. The network node may be implemented in hardware, software, firmware, or some combination thereof. Examples of a network node in a RAN include, but not limited to, a Node B (NodeB or NB), an evolved NodeB (eNodeB or eNB), a next generation NodeB (gNB), a transmission reception point (TRP), a remote radio unit (RRU), a radio head (RH), a remote radio head (RRH), an IAB node, a low power node such as a femto node, a pico node, a reconfigurable intelligent surface (RIS), and the like. Examples of a network node in a CN include, but not limited to, a Mobility Management Entity (MME), a Packet Data Network Gateway (P-GW), a Service Capability Exposure Function (SCEF), a Home Subscriber Server (HSS), or the like. Some other examples of a core network node include a node implementing an Access and Mobility Management Function (AMF), a User Plane Function (UPF), a Session Management Function (SMF), an Authentication Server Function (AUSF), a Network Slice Selection Function (NSSF), a Network Exposure Function (NEF), a Network Function (NF) Repository Function (NRF), a Policy Control Function (PCF), a Unified Data Management (UDM), or the like.

[0044] Communications in a communication environment may conform to any suitable standards including, but not limited to, Global System for Mobile Communications (GSM), Long Term Evolution (LTE), LTE-Evolution, LTE-Advanced (LTE-A), New Radio (NR), Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access (CDMA), GSM EDGE Radio Access Network (GERAN), Machine Type Communication (MTC) and the like. The embodiments of the present disclosure may be performed according to any generation communication protocols either currently known or to be developed in the future. Examples of the communication protocols include, but not limited to, the first generation (1G), the second generation (2G), 2.5G, 2.75G, the third generation (3G), the fourth generation (4G), 4.5G, the fifth generation (5G) communication protocols, 5.5G, 5G-Advanced networks, or the sixth generation (6G) networks and beyond and the like, wireless local network communication protocols such as Institute for Electrical and Electronics Engineers (IEEE) 802.11 and the like, and/or any other protocols currently known or to be developed in the future. Moreover, the communication may utilize any proper wireless communication technology, comprising but not limited to: Code Division Multiple Access (CDMA), Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Frequency Division Duplex (FDD), Time Division Duplex (TDD), Multiple-Input Multiple-Output (MIMO), Orthogonal Frequency Division Multiple (OFDM), Discrete Fourier Transform spread OFDM (DFT-s-OFDM) and/or any other technologies currently known or to be developed in the future.

[0045] As used herein, the term “model” is referred to as an association between an input and an output learned from training data, and thus a corresponding output may be generated for a given input after the training. The generation of the model may be based on machine learning (ML) techniques. The machine learning techniques may also be referred to as artificial intelligence (Al) techniques. In general, a machine learning model can be built, which receives input information and makes predictions based on the input information. For example, a classification model may predict a class of the input information among a predetermined set of classes. As used herein, “model” may also be referred to as “machine learning model”, “learning model”, “machine learning network”, or “learning network,” which are used interchangeably herein.

[0046] Generally, machine learning may usually involve three stages, i.e., a training stage, a validation stage, and an application stage (also referred to as an inference stage). At the training stage, a given machine learning model may be trained (or optimized) iteratively using a great amount of training data until the model can obtain, from the training data, consistent inference similar to those that human intelligence can make. During the training, a set of parameter values of the model is iteratively updated until a training objective is reached. Through the training process, the machine learning model may be regarded as being capable of learning the association between the input and the output (also referred to an input-output mapping) from the training data. At the validation stage, a validation input is applied to the trained machine learning model to test whether the model can provide a correct output, so as to determine the performance of the model. Generally, the validation stage may be considered as a step in a training process, or sometimes may be omitted. At the inference stage, the trained machine learning model may be used to process a real-world model input based on the set of parameter values obtained from the training process and to determine the corresponding model output.

[0047] FIG. 1 illustrates a schematic diagram for a simplified communication environment 100 in which embodiments of the present disclosure can be implemented. In this communication environment 100, a machine learning model 130 is provisioned following the consumer-producer paradigm. A communication entity which provisions the machine learning model 130 is referred to as a model provider 110, and a communication entity which requests from the model provider 110 model related services is referred to as a model consumer 120.

[0048] The model related services provisioned by the model provider 110 may include an analytic task to be performed by the machine learning model 130. The analytic task may be any prediction/analysis tasks that are to be performed within a communication environment. The machine learning model 130 is a data driven method which learns knowledge from a large amount of data and provide inference for the analytic task based on the leamt knowledge.

[0049] The machine learning model 130 is configured to process its model input to provide a model output. The model input of the machine learning model 130 may include a set of input features that are believed to be relevant to the analytic task, and the model output of the machine learning model 130 may be a prediction output for the analytic task. As an example, in a scenario where the analytic task is to predict network congestion, a set of input features to the machine learning model 130 may comprise indications such as the total number of active and idle mobile subscribers, aggregated uplink (UL) and downlink (DL) throughput, the load of the baseband board, and the like. As another example, in a scenario where the analytic task is to predict whether a behavior of a terminal device is expected or not, then the set of input features may comprise parameters such as trajectory and location of the terminal device as well as its traffic profile.

[0050] The model provider 110 and the model consumer 120 may be different entities in different communication networks. In some embodiments of a core network (CN), the model provider 110 may comprise a network node in the CN which is responsible for model maintenance, and the model consumer 120 may comprise any network node that requests for the model-related services from the model provider 110. For example, the model provider 110 may be a Network Data Analytics Function (NWDAF) in the CN, and the model consumer 120 may be any core network functions (referred to as “NF consumer”) which request analytics services from NWDAF. NWDAF may comprises a Model Training Logical Function (MTLF) and an Analytical Logical Function (AnLF), where MTLF is responsible for training and serving models for inference to AnLF. AnLF interfaces with the consumer and responds to the requests from the consumer. In the use case of CN, machine learning models may be provisioned for any end-to-end services in the CN, e.g., policy and subscription management and the like.

[0051] In addition to NWDAF, which concerns mostly the CN, another communication part where machine learning takes place is the radio access network (RAN), wherein Open-RAN (O- RAN) architecture defines the RAN Intelligence Controller (RIC). RIC is a function that controls radio resource management (RRM) in the O-RAN. In some embodiments of O-RAN, the model provider 110 may be a non-real time RAN intelligence controller (non-RT RIC), and the model consumer 120 may be a near real time (NT) RIC or any other entities within the O-RAN which accesses the model-related services (i.e., performance of an analytic task). In some other embodiments, the model provider 110 in the O-RAN may be a near NT RIC within the O-RAN, and the model consumer 120 may be any other entities within the O-RAN. In the use case of O- RAN, machine learning models may be provisioned for any radio access services, such as radio resource management.

[0052] In some embodiments, the model provider 110 may be an external function, for example, an Application Function (AF) or an Application Server (AS) which trains a machine learning model 130 with data from the communication network for a prediction. The model consumer 120 may be any entities within a communication network, e.g., in a CN, a RAN, or an O-RAN, which requests an analytic task to be performed by the trained machine learning model 130.

[0053] It is to be understood that although one machine learning model for one analytic task is illustrated in the model provider 110, there may be more than one machine learning model provisioned for the same or different analytic tasks. It is also to be understood that the elements shown in the communication environment 100 are intended to represent main functions provided within the environment. As such, the blocks shown in FIG. 1 refer to specific elements in communication networks that provide these main functions. However, other network elements may be used to implement some or all of the main functions represented. Also, it is to be understood that not all functions of a communication environment are depicted in FIG. 1. Rather, functions that facilitate an explanation of illustrative embodiments are represented.

[0054] Currently, input feature selection for the models trained in a communication network is not explored, as it is assumed that the features used to create a prediction per analytics task are known a priori to the model provider. As mentioned, it has been defined in some 3^rd Generation Partnership Project (3GPP) communication specifications a number of use cases, each use case having a pre-defined set of input features and the prediction output. [0055] However, given the dynamicity of communication networks and the external factors affecting them (e.g., behavior of terminal devices such as traffic, cardinality and mobility patterns, environmental effects, power supply issues, and the like), it is not ideal to use a static collection of input features for an analytic task in the communication networks.

[0056] For example, in the scenario where a machine learning model is supposed to predict network congestion, in addition to the indications such as the total number of active and idle mobile subscribers, aggregated UL and DL throughput, the load of the baseband board as mentioned above, other features may also be useful depending on the location and the status of the power grid, such as the voltage reading on the power supply unit (PSU), the current state of charge and rate of discharge of the battery subsystem. Such feature information may be used by the machine learning model to determine whether an external power fluctuation that cannot be rectified by the battery subsystem may produce a congestion, due to the radio units and baseband board being under-powered (or not powered at all). In another example of predicting whether a behavior of a terminal device is expected or not by a machine learning model, in addition to the parameters such as the trajectory, location, and traffic profile as mentioned above, other parameters such as Channel State Information (CSI) reporting, power level, and power class can also be considered in some cases.

[0057] Typically, machine learning models relying on more relevant input features can provide more accurate predictions. While the exact input features that are used for different use cases or analytic tasks may differ, the challenge lies on how generic machine learning models, shipped as features or parts of a feature in the communication network, can be optimized to the condition of individual communication networks in an efficient way. Specifically, it is expected that the model provider could efficiently explore the input feature space to provide models that can make accurate predictions based on the current state of the network but also do not contain any unnecessary features. In addition, while an all-inclusive feature strategy could work, such model will be overly complex at the cost of computational resources, as well as bandwidth to transfer the data required for model training and inference.

[0058] Example embodiments of the present disclosure provide a solution for machine learning model feature selection by enabling feature selection for a machine learning model in a communication network. The feature selection allows to explore the input feature space for an analytic task in an intelligent and dynamic or real-time way. As a result, a machine learning model can be provisioned with an optimized set of input features that is applicable to the current network context and providing an acceptable level of model performance.

[0059] According to some example embodiments of the present disclosure, a logical function, namely Feature Selection Function (FSF) is introduced to implement the machine learning model feature selection. FSF may be a physical part of a model provider or may be a standalone component that separates from and is communicatively coupled with the model provider. Reference can be made to FIG. 2A and FIG. 2B. FIG.2A illustrates that FSF 210 is introduced as a part of the model provider 110 in the communication environment 100. In this case, signaling overhead is reduced which saves network bandwidth. FIG.2B illustrates that FSF 210 is deployed as a separate component outside of the model provider 110 in the communication environment 100. The model provider 110 may subscribe to FSF 210 for feature selection. In the standalone case of FIG. 2B, FSF 210 may be subscribed by one or more other entities that would potentially desire to optimize their machine learning models.

[0060] In some embodiments, FSF 210 proposed herein may be implemented as a network node or a part of a network node (e.g., the model provider) in a communication network. In different communication networks with machine learning enabled analytics, FSF 210 proposed herein may be deployed within any suitable entities in those communication networks that provide machine learning models or otherwise communicatively coupled with the corresponding entities. In the following, example embodiments of the feature selection by FSF 210 in FIG. 2A and FIG. 2B are first discussed in detail, followed by example embodiments of deployments of FSF in different use cases, including the CN use case (with reference to FIGS. 4 A, 4B, and 5), the O-RAN use case (with reference to FIGS. 6A-6B), and a standalone use case outside of a communication network (with reference to FIG 7).

[0061] FIG. 3 illustrates a signaling chart 300 for feature selection within the communication environment in FIG. 2A or FIG. 2B in accordance with some further embodiments of the present disclosure. As shown in FIG. 3, the signaling chart 300 involves FSF 210 and the model provider 110 in FIG. 2A and FIG. 2B.

[0062] FSF 210 is configured to provision feature selection for one or more machine learning models maintained by the model provider 110 for one or more analytic tasks. It is assumed that the machine learning model 130 is currently provisioned by the model provider 110 for a specific analytic task based on a current set of input features. That is, in response to an analytic request of the analytic task from the model consumer 120, the model provider 110 may utilize the current machine learning model 130 to produce an analytic result for the model consumer 120.

[0063] The feature selection for the machine learning model 130 is triggered by a feature selection trigger. As an external trigger 301, FSF 210 may receive 305 the feature selection trigger from an entity external to FSF 210, e.g., from the model provider 110. As an internal trigger 302, the feature selection trigger may be initiated 310 in FSF 210 internally.

[0064] In some embodiments, the feature selection may be triggered by a time constraint, e.g., periodical execution. For example, FSF 210 or the model provider 110 may start a timer from a last feature selection trigger for the analytic task. A feature selection trigger may be initiated at an expiration of the timer. In the case that the timer is set at the model provider 110, it may transmit a feature selection trigger to FSF 210.

[0065] In some embodiments, the feature selection trigger may be initiated based on the number of analytic requests for the analytic task. Analytic requests from the model consumer 120 may include an identifier (ID) of the analytic task so that the model provider 130 can determine which machine learning model is now provisioned for execution of the analytic task. The number of analytic requests can thus be counted for a certain analytic task. If the number of analytic requests exceeds a threshold number (which may be (pre)configured for the analytic task), a feature selection trigger may be initiated. That is, the feature selection trigger may be initiated, for example, every X analytic tasks (where X may be (pre)configured as any suitable number). In the case that the number of analytic requests is counted at the model provider 110, it may transmit a feature selection trigger to FSF 210. It is generally not necessary that the feature selection is called for every analytic request, but the threshold number for triggering the feature selection may be dependent on different factors such as the type of the analytic, the instance ID of the model consumer originating the analytic tasks, the type of the model consumer, the type of the machine learning model that is currently provisioning for the analytic task, and/or the like. In some embodiments, the counter of the analytic task may be cleared upon the initiation of the current feature selection trigger.

[0066] In some embodiments, the feature selection trigger may be initiated by detection of poor model performance for the analytic task. In particular, if an evaluation level of the machine learning model 130 is detected to be below a threshold level, then a feature selection trigger may be initiated in order to select a set of input features to improve the performance. The evaluation level may be determined in such a way that a higher evaluation level indicates better model performance. Therefore, if the evaluation level is detected to fall below the threshold level, it may indicate that the current set of input features may be not optimized given the current context of the communication network.

[0067] The evaluation level of the currently provisioned machine learning model 130 may be measured with any suitable model performance indicators. As for some examples, the evaluation level may be determined based on an accuracy level (e.g., the number of correct predictions made with respect to all predictions made), a log loss, an area under curve (AUC), a mean absolute error and mean squared error, etc. Any other metrics that are known for evaluating a machine learning model may also be applicable. Similar to the threshold number above, the threshold level for triggering the feature selection may be dependent on different factors such as the type of the analytic, the instance ID of the model consumer originating the analytic tasks, the type of the model consumer, the type of the machine learning model that is currently provisioning for the analytic task, and/or the like.

[0068] In some embodiments, FSF 210 may internally monitor the model performance for a certain analytic task by checking whether the currently provisioned machine learning model 130 can provide satisfied prediction outputs for the analytic task. FSF 210 may perform the model performance monitoring by subscribing to the data collection service from, e.g., Data Collection Coordination Function (DCCF) and/or Messaging Framework Adaptor Function (MFAF) in the communication network, to retrieve current model inputs and perform model interference using the currently provisioned machine learning model 130. FSF 210 may monitor the evaluation level (e.g., some key performance indicators) for a period of time using the same data collection service, to identify whether the evaluation level is below the threshold level (e.g., the model outputs were accurate or not). In the case of an unsatisfied evaluation level, FSF 210 may trigger feature selection for the machine learning model 130. In some other embodiments, the model provider 110 or other network nodes may perform the model performance monitoring and transmit an external feature selection trigger to FSF 210 upon detection of deteriorated model performance.

[0069] In some embodiments, the feature selection trigger may be initiated by the detection of a change of one or more features that are used in the currently provisioned machine learning model 130, the detection of one or more new features for enhancing the machine learning model 130, and/or the detection of a change of context information related to the analytic task. In some embodiments, if one or more features for the machine learning model 130 are extracted from time series data or periodic data (such features may be referred to as “seasonality” features), those features may be changed in different contexts, which implies that the original features may not be suitable for the analytic task and/or one or more new features can be determined from the time series data. In some embodiments, the new features may be determined, for example, through a model drift analysis. For a detected model drift, a root cause analysis may be applied, to translate the cause of the detected drift into one or more new features which could be used for enhancing the model performance. In some embodiments, the change of context information may also imply that the feature space is to be explored for the changed context. In some embodiments, the feature change, context change and/or detection of new feature(s) may be indicated by the model consumer 120, by the model provider 110, by FSF 210 own initiative, a service application initiating the analytic task, and/or an analysis agent (such as a root cause analysis agent for model drift).

[0070] In some embodiments, the feature selection by FSF 210 may be triggered on the detection of an update of a use case context. A use case context may indicate the context of use of the analytics. In some examples, a parameter indicating a use case context may be provided by the model consumer 120 (e.g., NF consumer) when subscribing to the analytics from the model provider 110 (e.g., NWDAF). The model provider 110 can use the parameter to select the relevant machine learning model for the requested analytics, for example, if some specific models are configured for this specific use case. In such a case, if a new use case context is detected (for example, power failure or some other type of situation which affects the use case), then the feature selection by FSF 210 may be triggered to select anew set of input features for the analytic task. In some embodiments, the model provider 110 may monitor the use case context and transmit an external feature selection trigger to FSF 210 upon detection of the update of the use case context. In some embodiments, FSF 210 may receive or otherwise detect the use case context from the model consumer 120 to determine whether feature selection is to be initiated.

[0071] In some embodiments, FSF 210 may initiate the feature selection for the machine learning model 130 in response to a specific request from a network manager. The network manager may transmit a feature selection trigger to FSF 210 when needed to invoke the feature selection. In this way, it is feasible to achieve flexible and convenient model management via FSF 210 in the communication network.

[0072] It is noted that although some conditions for initiating the feature selection trigger are provided individually above, some of those conditions may be combined to trigger the feature selection at FSF 210. For example, the feature selection is triggered in the case when the number of analytic requests for the analytic task exceeds a threshold number and the timer is expired.

[0073] In response to the feature selection trigger, FSF 210 performs 315 feature selections for the machine learning model 130. Given different network contexts, FSF 210 aims to optimize machine learning models provisioned in communication networks by reducing computational and bandwidth requirements for model training and inferencing, while at the same time ensuring a satisfactory level of model performance.

[0074] Specifically, FSF 210 determines, based on context information related to the analytic task, a target set of input features for the analytic task that is currently performed by the machine learning model 130. The target set of input features is determined to be adapted to the current context related to the analytic task, in order to avoid the usage of redundant and irrelevant input features and/or missing some important and relevant features in the current context. As a result, the target set of input features can be determined as different from the current set of input features for the machine learning model 130.

[0075] The context information related to the analytic task may comprise any suitable information that can indicate a context or state of the communication network in relation to the analytic task. The context information may be varied over time and from network to network. With the change in the context, a different set of input features may become relevant to derive accurate results for the analytic task. In some embodiments, the context information may be received from the communication network, e.g., one or more network nodes deployed within the communication network. For example, the context information may include performance monitor (PM) counters originating from the baseband network node of the radio access network, network function status information, alarms, logs, and/or the like. Additionally, or as an alternative, the context information may be received from a third party outside of the communication network, such as temperature predictions retrieved by a weather service, power predictions retrieved by a power supplier, and/or the like.

[0076] In some embodiments, FSF 210 may determine the target set of input features further based on a reference set of input features for the analytic task. The reference set of input features may be the current set of input features for the machine learning model 130. As an alternative, the reference set of input features may be a baseline set of input features that are determined to allow a baseline evaluation level of a machine learning model for the analytic task. FSF 210 may be provided with a description of the reference set of input features.

[0077] In some embodiments, during the feature selection process, FSF 210 may retrieve the currently provisioned machine learning model as well as the context information and a description of the reference set of input features. Given a condition of the context information, FSF 210 may modify the reference set of input features by deleting one or more input features from the reference set. In some embodiments, during the feature selection process, the reference set of input features may be modified not only by reduction of its cardinality but also replacing incumbent input features with others and/or adding a new input feature(s), to create a set of more relevant features. [0078] In some embodiments, the target set of input features may be determined to have a relatively small (or minimum) number of input features for the analytic task and can still yield a machine learning model with an acceptable evaluation level. FSF 210 may apply, for example, statistical-based solutions such as ANOVA F-test to genetic algorithm-inspired solutions, to select a suitable set of input features in the current context. In some embodiments, FSF 210 may construct one or more candidate sets of input features from the reference set of input features, with each candidate set comprising different combinations of input features from the reference set. FSF 210 may determine, given a condition of the context information, a predicted evaluation level (e.g., accuracy, loss, AUC or mean absolute error and mean squared error) of a machine learning model with each candidate set of input features as a model input. Then FSF 210 may select the target set of input features from the candidate sets of input features based on the predicated evaluation levels, for example, select the candidate set of input features with the highest predicted evaluation level or with a predicted evaluation level higher than an expected level. [0079] There are various specific implementations to explore the deletion, addition, and/or replacement of features given the reference set of input features (e.g., the baseline set or the current set), some examples of which will be described in detail below.

[0080] With the target set of input features determined, FSF 210 causes a further machine learning model to be provisioned to perform the analytic task based on the determined target set of input features. For the purpose of discussion herein, the current machine learning model 130 for the analytic task is sometimes referred to as a “first” machine learning model for the analytic task, the further machine learning model based on the target set of input features is sometimes referred to as a “second” machine learning model for the same analytic task. The second machine learning model may be reprovisioned by the model provider 110 to respond to future analytic requests for the analytic task.

[0081] In some embodiments, 321 where FSF 210 is a separate component outside of the model provider 110, FSF 210 may transmit 320 a model reprovisioning request to the model provider 110. The model reprovisioning request indicates the target set of input features for the analytic task (e.g., identified with its analytic ID), to cause the model provider 110 to provision the second machine learning model with the target set of input features for performing the analytic task. In some embodiments, the model reprovisioning request may be considered as a response to a feature selection request transmitted by the model provider 110 to FSF 210. Upon reception 325 of the model reprovisioning request, the model provider 110 may perform 330 model reprovisioning for the analytic task by determining the second machine learning model for performing the analytic task based on the target set of input features. In some embodiments, where FSF 210 is integrated within the model provider 110, the model provider 110 may internally trigger the provisioning of the second machine learning model based on the determined target set of input features.

[0082] In some embodiments, the second machine learning model may be a retrained (or refined) version of the first machine learning model. That is, the first machine learning model may be retrained or refined with the target set of input features, to generate the second machine learning model. The model provider 110 may collect training data for the retraining or refining process. For example, MTLF in NWDAF of CN, which operates as a model provider, may collect training data from other network nodes, namely the Operations Support System (OSS) via DCCF and MFAF which retrieve current training data from other NFs and historical training data from Analytics Data Repository Function (ADRF). In the use case of O-RAN, the RIC, which operates as a model provider, may collect training data from the RIC itself and also from other nodes within the O- RAN, such as from O-RAN Distributed Unit (O-DU), O-RAN Control Unit (O-CU) and/or O- RAN Radio Unit (O-RU). After the model retraining or refining, the model input of the obtained second machine learning model may be set as the target set of input features.

[0083] In some embodiments, the model provider 110 may maintain a plurality of candidate trained machine learning models for the analytic task and thus may select one of the candidate trained machine learning models which is configured with the target set of input features as the second machine learning model. In some embodiments, in addition to retraining and selection of a trained machine learning model, the model provider 110 may reconfigure an available machine learning model (which may be the first machine learning model or other trained machine learning model with a current set of input features) for the analytic task, to function as the second machine learning model with the determined target set of input features as the second machine learning model.

[0084] In implementations, the model provider 110 may have various ways to provision the second machine learning model with the determined target set of input features to perform the analytic task. In some embodiments, the target set of input features determined by FSF 210 may be used as a recommendation for determining a set of input features for the second machine learning model to be reprovisioned for the analytic task, for example, if FSF 210 is an external function consulted by the model provider 110. In this case, FSF 210 is considered as providing a recommendation service on what features to use for the model and updating with new features in case of changes in context.

[0085] Some example implementations have been discussed above. The model provider 110 may utilize any of those implementations or other implementations to provision the second machine learning model. In some other embodiments, FSF 210 may perform the retraining, selection and/or reconfiguration of the second machine learning model itself in case it has access to training data or the available trained machine learning models for the analytic task.

[0086] With the assistance of FSF 210, the model provider 110 can provide model optimization services for the analytic task. The second machine learning model with the target set of input features can be optimized to the current context in the communication network. Through the feature selection process, it can save bandwidth resources by reducing the number of input features and thus the amount of data needed to be collected for inferencing. Potentially, it can also save computational resources by reducing the number of model layers as a result of the lower number of input features. On the other hand, it allows for better performance of provisioned models, as it finds a set of input features that provide more accurate predictions than the baseline set.

[0087] Some example embodiments of the feature selection process for a machine learning model have been discussed above. As discussed above, the feature selection process is to determine the target set of input features based on a reference set of input features given context information related to the analytic task. Some implementations to explore the deletion, addition, and/or replacement of features given the reference set of input features (e.g., the baseline set or the current set) will now be set forth in more detail.

[0088] In some embodiments, FSF 210 may apply a reinforcement learning (RL) process to select the target set of input features. The RL process is about learning the optimal behavior in an environment to obtain a maximum reward. This optimal behavior is learned through interactions with the environment and observations of how it responds. The basic elements in the RL process are the environment, the RL agent, the action a, the states, and the reward r. When an action is executed in the environment, the RL process observes the state and the reward of the action. While the observed reward r is higher than or equal to a threshold consistently (i.e., for the last m iterations), then a new action is executed based on heuristic exploration of the state space. Then the reward and the next state are observed. The current state, action, reward, and the next state < s, a, r, ,s+ 1 > may be stored in an experience buffer. After a number of iterations, the RL agent is trained using data stored in the experience buffer. Finally, the RL agent is returned to provide the final action as the target set of input features.

[0089] In the RL process for feature selection, the reference set of input features may be initialized with, for example, /features (which may from the baseline set of input features or the current set of input features from the currently provisioned machine learning model). The RL process performs constrained exploration of the broader feature set, taking the action a as selection of input features comprising the reference set of input features with some features added or removed. That is, each action may select a candidate set of input features for evaluation.

[0090] The action is selected in accordance with a state. A state .s' in the feature selection indicates the context information of the communication network in relation to the analytic task. As a specific example, the context information may describe the alarms raised by the baseband board and are accessible to the model provider via DCCF/MFAF and OSS. A heuristic-based exploration of the state space can take place. For example, if the alarm originated from the Power Subsystem MO, then features relating to power measurements can be considered, constraining the action space. In this example, such features may be PM counters from the Power Supply Unit (PSU) and battery backup unit (BBU). To avoid the creation of multiple models with all possible permutations of such PM counters, FSF 210 may apply explainable Al techniques such as Shapely Additive Explanation (SHAP) and local interpretable model-agnostic explanations (LIME). The idea is that features relating to the state (e.g., PM counters under a managed object, MO) may be added into the reference set of input features and then the explainable Al techniques may be utilized to find relative feature importance. Then the set of features that are indicated as most important may be selected and monitored for the reward. [0091] The reward function comprises of two parts. The first part incentivizes the RL agent to predict events accurately, while the second part penalizes the RL agent for selecting a larger feature set. The idea behind this design is that the feature selection process is to have the least number of features selected while at the same time making the machine learning model to output as accurate predictions as possible. The two constituents may also be weighed based on which one is more important than the other. Accuracy estimation can be based e.g. on the time where an event is predicted to occur versus the predicted time of occurrence, while the feature set constituent can be calculated as the ratio of selected features to the total number of features.

[0092] In some embodiments, FSF 210 may apply a contextual bandit process to select the target set of input features. The contextual bandit process is to leam the number of input features (or more specifically which subset of features) to select depending on a given context and for a specific reward function. In the scenario of feature selection, the state or context is considered as the context information related to the specific analytic task, for example, the context information describing specific alarms as in the aforementioned example, or context information without alarms and other contexts. The action space may be the powerset of k, where k is a set containing all input features that the corresponding machine learning model for the analytic task can consider. Thus, the total number of candidate sets may be the number of different selections from the k input features (represented as “k”). In some examples, k may be configured as a reasonably small number to avoid an extreme number. Given the state for each bandit and the number of features, FSF 210 leams though exploration (i.e., via the e-greedy approach) which candidate set of features yields the highest reward for a given state. In some embodiments, the reward function may be the weighted sum between the accuracy of the machine learning model trained with the selected features plus the normalized sum of input features that are marked as important using an explainability technique such as SHAP or LIME. In an example, the reward function may be represented as follows: where R(s, k) represents the reward of the state .s' with the features k, \k\ represents a selection from the k input features, model_accuracy represents the accuracy of the machine learning model, important ^_eatures represents the importance of the selected features, wl and w2 represent the weights for the model accuracy and the feature importance, respectively. It is noted that the reward function for the contextual bandit process can be defined in different ways which are not limited in the scope of the present disclosure. [0093] In some embodiments, FSF 210 may apply a Monte-Carlo tree search process to select the target set of input features. The RL process as discussed above may be considered as having a state space for a specific analytic task. In the Monte-Carlo tree search process, it is assumed that an RL agent has a large state space for a plurality of analytic tasks (e.g., a plurality of analytics IDs). In the exploration phase, the RL agent may build a map between the newly selected input features and the improvement of the targeted analytic task. Such a map may be in form of M(A_xy) which may for example indicate Acc(A_21)= 0.9, Acc(A_22)= 0.95, Acc(A_34)= 0.95 (note that those numerical values are provided for purpose of illustration only). In this example, it is assumed that A_lx is the accuracy of all x inputs to a target analytics task (“Analytics- 1”), where A_22 represents the accuracy of the second input of the second analytic task when checked during the training of the target analytic task “Analytics- 1”. After completing this map M(A_xy), it is compared to a threshold, and a segment of this map is left for the RL agent to explore and select from in order to use it in enhancing the accuracy of the target analytic task A l . If the analytic task index i (A_ij) is negative, it means that this is out of the existing analytics inputs, and it can be considered as external input. After creating such map, using e.g., Monte-Carlo Tree search Al mechanisms (like Alpha Zero), the RL agent may train a neural network (NN) to guide its search via the trained neural network to pick which input to other analytics that helped in increasing accuracy of the target analytics task (“Analytics- 1”).

[0094] In some embodiments, FSF 210 may apply a collaborative filtering process using a recommendation engine to select the target set of input features. In the collaborative filtering process, the context information may be mapped with a set of input features using the recommendation engine. Various recommendation mechanisms may be applied by the recommendation engine. The target set of input features may be selected from the recommended set of input features for the current context related to the analytic task.

[0095] It is noted that although some example processes for exploring the target set of input features are provided above individually, some of those processes may be applied in combination to facilitate the feature selection process. Variations to any of the above mentioned processes and any other suitable exploring processes are also applicable for feature selection.

[0096] The overall functionalities of FSF 210 have been introduced above. As mentioned, FSF 210 may be applied in different use cases, including the CN and 0-RAN, or as a standalone use case outside of a communication network.

[0097] FIG. 4 A and FIG. 4B illustrate example CN use cases with FSF 210 introduced. As illustrated in FIG. 4A and FIG. 4B, in the CN use case, NWDAF 410 functions as a model provider and comprises AnLF 412 for interfacing with a NF consumer 420 and MTLF 414 for model training and serving for inference to AnLF 412. The NF consumer 420 functions as a model consumer and may be any NF in the CN which subscribes the model related services from NWDAF 410.

[0098] In normal operation, upon reception of a subscription request (e.g., an Nnwdaf AnafyticsSubscription Subscribe request) from the NF consumer 420, AnLF 412 requests MTLF 414 to provision a machine learning model for inferencing via a model provisioning request (e.g., an Nnwda '_ModelProvision_Subscri.be request). If the requested machine learning model exists (e.g., the machine learning model 130), then MTLF 414 may provide a Universal Resource Identifier (URI) or other address of the machine learning model to AnLF 412 in a model provisioning response (e.g., an Nnwdaf 'ModelProvision Notify response). If the machine learning model does not exist, then MTLF 414 may initiate a data collection process of current and historical training data which is subsequently input to a model training algorithm (e.g., based on gradient descent). The process of data collection may involve other nodes, e.g., the OSS via DCCF and/or MFAF 430 which retrieve current training data from other network functions and historical training data from ADRF. With the model trained, in some embodiments, MTLF 414 may deploy a microservice and expose the trained machine learning model to AnLF 412, e.g., by responding with the model address in the Nnwdaf ' ModelProvision Notify response. [0099] AnLF 412 may use the machine learning model for inference, including collecting input features using DCCF and/or MFAF 430 and executing the machine learning model (e.g., doing a forward pass in case the model is a feed forward neural network). The inference process may return an analytic result which may be encapsulated to a response (e.g., Nnwdaf AnafyticsSubscription Notify response) to the originating NF consumer 420.

[0100] On top of the normal model provisioning in NWDAF 410, FSF 210 is introduced for feature selection related to a machine learning model or the associated analytic task. In a CN 400A as illustrated in FIG. 4A, FSF 210 may be a physical part of NWDAF 410, e.g., integrated within MTLF 414. It is noted that in some other embodiments, FSF 210 may be a physical part of AnLF 412. In a CN 400B as illustrated in FIG. 4B, FSF 210 may be implemented as an external network node outside ofNWDAF 410. In the CN 400B, FSF 210 may be subscribed by NWDAF 410 (more specifically, by MTLF 414) and/or one or more other network nodes that would potentially desire to optimize their machine learning models. In the scenario of FIG. 4B, FSF 210 may also interact with DCCF and/or MFAF 430, to retrieve data or information needed for the feature selection process. In the CN 400A and CN 400B, FSF 210 can operate as discussed in any of the above embodiments.

[0101] FIG. 5 illustrates a signaling chart 500 for feature selection among network nodes within the CN in FIG. 4A and FIG. 4B in accordance with some embodiments of the present disclosure. [0102] In the signaling chart 500, a feature selection process for a certain analytic task may be triggered via a feature selection trigger, which may be received 510 by FSF 210 as an external trigger 501 or may be initiated 515 internally in FSF 210 as an internal trigger 502. As an external trigger 501, the feature selection trigger may be received from AnLF 412, the NF consumer 420, and/or any other network nodes. The conditions for triggering the feature selection may be similar to those discussed above.

[0103] In response to the feature selection trigger, FSF 210 performs 520 feature selections for the machine learning model 130 (referred to as a “first machine learning model”) which is currently provisioned for performing the analytic task. FSF 210 may obtain context information related to the analytic task and determine a target set of input features for the analytic task that is currently performed by the machine learning model 130. The detailed implementations for feature selection may be similar to those discussed above.

[0104] In some embodiments, 521 where FSF 210 is a separate component outside of MTLF 414, FSF 210 may transmit 525 a model reprovisioning request to MTLF 414, to cause MTLF 414 to provision a further machine learning model (referred to as a “second machine learning model”) to perform the analytic task based on the determined target set of input features. Upon reception 530 of the model reprovisioning request, MTLF 414 may perform 535 model reprovisioning for the analytic task by determining the second machine learning model for performing the analytic task based on the target set of input features. In some embodiments, where FSF 210 is integrated within the model provider 110, MTLF 414 may internally trigger the provisioning of the second machine learning model based on the determined target set of input features. The model reprovisioning may comprise retraining/refining of the currently provisioned machine learning model, reconfiguring an available trained machine learning model, or selecting from a plurality of candidate trained machine learning models for the analytic task, as discussed above.

[0105] In some embodiments, 541 where FSF 210 is a separate component outside of MTLF 414, MTLF 414 may transmit 540 a model reprovisioning notification to FSF 210, as a response to the model reprovisioning request. Upon reception 545 of the notification, FSF 210 may confirm whether the model reprovisioning is completed. In some embodiments, if the feature selection trigger is received from AnLF 412, FSF 210 may notify AnLF 412 of the target set of input features determined for the analytic task by transmitting 550 a feature selection notification to AnLF 412. Upon reception 555 of the feature selection notification, AnLF 412 may be aware of the input features to be collected for inferences based on the second machine learning models.

[0106] In addition to NWDAF, another use case of FSF 210 is the O-RAN architecture. FIG. 6A and FIG. 6B illustrate example O-RAN use cases with FSF 210 introduced. The O-RAN architecture aims at intelligent RAN automation by applying artificial intelligence, machine learning, and advanced analytics to manage complex, ever changing network, device and end-user requirements for optimal efficiency.

[0107] As illustrated in FIG. 6A and FIG. 6B, the O-RAN architecture defines the RAN Intelligence Controllers (RICs). RIC is a function that controls radio resource management (RRM) in RAN. RIC works in two layers, a non-real time (non-NT) layer (also referred to as a service management and orchestration (SMO) layer) with a non-NT RIC 610, and a near real time (RT) layer with a near NT RIC 620.

[0108] In the O-RAN 600A of FIG. 6A, the non-NT RIC 610 may function as both a model provider and a model consumer of one or more machine learning models including the machine learning model 130. The non-NT RIC 610 may collect training data or model input data for the machine learning models from O-CU 622, O-DU 624, and/or O-RU 626, for example, via 01 interface. FSF 210 may be integrated within the non-RT RIC 610, for example, as a non-RT RIC application (called “rAPP”). In this case, FSF 210 provides services to the machine learning models of the non-RT RIC 610 using R1 interface. The machine learning models can be exposed to FSF 210 via rAPP service exposure. FSF 210 in the O-RAN 600A may operate in a similar way as discussed above. In some embodiments, FSF 210 may also have access to model training data via 01 interface and to the near NT RIC 620 via Al interface.

[0109] In some embodiments of the O-RAN 600A in FIG. 6A, the near RT RIC 620 may also access FSF 210 in the non-RT RIC 610, for example, using the Al interface, to request feature selection for machine learning models of the near RT RIC 620.

[0110] FIG. 6B illustrates another scenario of O-RAN 600B, where FSF 210 is integrated within the near RT RIC 620. In this case, the near RT RIC 620 may function as both a model provider and a model consumer of its machine learning model(s). The near RT RIC 620 may collect training data or model input data for the machine learning models from O-CU 622, O-DU 624, and/or O-RU 626, for example, via 01 interface. In some embodiments, FSF 210 may use E2 interface and specifically E2 Service Model (E2SM) Key performance matrix (KPM) to retrieve data needed for feature selection. FSF 210 may provide services to other xApps 630 in the near RT RIC 620 via an open messaging infrastructure 640 contained in the near RT RIC 620.

[0111] In addition to NWDAF and O-RAN, FSF 210 may be provided as a standalone function, either internal to a communication network or external to it, for example, hosted in a third-party cloud. FIG. 7 illustrates a schematic diagram for an environment 700 where FSF 210 is provided as a standalone function external to a communication network 715. The communication network 715 may include a CN, a RAN, an O-RAN, or other mobile networks.

[0112] As illustrated, a service exposure function such as the Service Capability Exposure Function (SCEF) or Network Exposure Function (NEF) 710 of the communication network 705 may be configured to request a feature selection service from FSF 210. In some embodiments, if an external function (AF or AS 720) of the communication network 715 is configured to train a machine learning model with data from the network for analytics, FSF 210 may also provide feature selection for such machine learning model, for example, through SCEF/NEF 710.

[0113] In the use case of FIG. 7, AF or AS 720 may function as a model provider, and the SCEF or NEF 710 or any other network node within the communication network 715 may function as a model consumer. In some embodiments, the result of the feature selection by FSF 210 may be used as a recommendation on what features to use for the model and update with new features in case of changes in context.

[0114] FIG. 8 illustrates a flowchart of a method 800 for machine learning model feature selection in accordance with some embodiments of the present disclosure. For the purpose of discussion, the method 800 will be described from the perspective of FSF 210 in FIG. 2A and FIG. 2B.

[0115] At block 810, in response to a feature selection trigger for a first machine learning model, FSF 210 determines a target set of input features for an analytic task based on context information related to the analytic task. The first machine learning model is currently provisioned for performing the analytic task based on a current set of input features, the current set of input features being different from the target set of input features. In some embodiments, the context information may comprise at least one of the following: context information received from one or more further network nodes within the communication network, or context information received from a third party.

[0116] In some embodiments, the feature selection trigger may be initiated based on at least one of the following: an expiration of a timer started from a last feature selection trigger, the number of analytic requests for the analytic task exceeding a threshold number, an evaluation level of the first machine learning model being below a threshold level, detection of a change in at least one feature for the first machine learning model, detection of a change in context information related to the analytic task, reception of a feature selection request from a network manager, detection of a new feature for enhancing the first machine learning model, or detection of an update of a use case context.

[0117] In some embodiments, the target set of input features may be determined further based on a reference set of input features for the analytic task. In some embodiments, the reference set of input features may be determined to allow a baseline evaluation level of the first machine learning model, or the reference set of input features comprises the current set of input features for the first machine learning model. [0118] In some embodiments, FSF 210 may determine at least one candidate set of input features from the reference set of input features, and determine, given a condition of the context information, at least one predicted evaluation level of a machine learning model with the at least one candidate set of input features as a model input. FSF 210 may select the target set of input features from the at least one candidate set of input features based on the at least one predicated evaluation level.

[0119] In some embodiments, the target set of input features may be determined through at least one of the following: a reinforcement learning (RL) process, a contextual bandit process, a Monte-Carlo tree search process, or a collaborative filtering process using a recommendation engine.

[0120] At block 820, FSF 210 causes a second machine learning model to be provisioned to perform the analytic task based on the determined target set of input features.

[0121] In some embodiments, the first machine learning model may be caused to be retrained with the target set of input features, to obtain the second machine learning model. In some embodiments, a trained machine learning model may be caused to be reconfigured with the target set of input features as the second machine learning model. In some embodiments, the second machine learning model may be caused to be selected from a plurality of candidate trained machine learning models.

[0122] In some embodiments, the target set of input features may be used as a recommendation for determining a set of input features for the second machine learning model. [0123] In some embodiments, the communication network may comprise a core network (CN). In some embodiments, FSF 210 may comprise a model training logical function (MTLF) in a network data and analytics function (NWDAF) or a separate network function outside of the MTLF.

[0124] In some embodiments, the feature selection trigger may be received from an analytical logical function (AnLF). In this case, FSF 210 may notify the AnLF of the target set of input features.

[0125] In some embodiments, the communication network may comprise an open radio access network (O-RAN). In some embodiments, FSF 210 may be deployed at a non-real time RAN intelligence controller (non-RT RIC) or at a near RT RIC within the O-RAN.

[0126] FIG. 9 shows an example of a communication system 900 in accordance with some embodiments of the present disclosure. In the example, the communication system 900 includes a telecommunication network 902 that includes an access network 904, such as a radio access network (RAN), and a core network 906, which includes one or more core network nodes 908. The access network 904 includes one or more access network nodes, such as network nodes 910a and 910b (one or more of which may be generally referred to as network nodes 910), or any other similar 3^rd Generation Partnership Project (3GPP) access node or non-3GPP access point. The network nodes 910 facilitate direct or indirect connection of user equipment (UE), such as by connecting terminal devices 912a, 912b, 912c, and 912d (one or more of which may be generally referred to as terminal devices 912) to the core network 906 over one or more wireless connections. [0127] Example wireless communications over a wireless connection include transmitting and/or receiving wireless signals using electromagnetic waves, radio waves, infrared waves, and/or other types of signals suitable for conveying information without the use of wires, cables, or other material conductors. Moreover, in different embodiments, the communication system 900 may include any number of wired or wireless networks, network nodes, terminal devices, and/or any other components or systems that may facilitate or participate in the communication of data and/or signals whether via wired or wireless connections. The communication system 900 may include and/or interface with any type of communication, telecommunication, data, cellular, radio network, and/or other similar type of system.

[0128] The terminal devices 912 may be any of a wide variety of communication devices, including wireless devices arranged, configured, and/or operable to communicate wirelessly with the network nodes 910 and other communication devices. Similarly, the network nodes 910 are arranged, capable, configured, and/or operable to communicate directly or indirectly with the terminal devices 912 and/or with other network nodes or equipment in the telecommunication network 902 to enable and/or provide network access, such as wireless network access, and/or to perform other functions, such as administration in the telecommunication network 902.

[0129] In the depicted example, the core network 906 connects the network nodes 910 to one or more hosts, such as host 916. These connections may be direct or indirect via one or more intermediary networks or devices. In other examples, network nodes may be directly coupled to hosts. The core network 906 includes one more core network nodes (e.g., core network node 908) that are structured with hardware and software components. Features of these components may be substantially similar to those described with respect to the terminal devices, network nodes, and/or hosts, such that the descriptions thereof are generally applicable to the corresponding components of the core network node 908. Example core network nodes include functions of one or more of a Mobile Switching Center (MSC), Mobility Management Entity (MME), Home Subscriber Server (HSS), Access and Mobility Management Function (AMF), Session Management Function (SMF), Authentication Server Function (AUSF), Subscription Identifier De-concealing function (SIDF), Unified Data Management (UDM), Security Edge Protection Proxy (SEPP), Network Exposure Function (NEF), and/or a User Plane Function (UPF). [0130] The host 916 may be under the ownership or control of a service provider other than an operator or provider of the access network 904 and/or the telecommunication network 902, and may be operated by the service provider or on behalf of the service provider. The host 916 may host a variety of applications to provide one or more service. Examples of such applications include live and pre-recorded audio/video content, data collection services such as retrieving and compiling data on various ambient conditions detected by a plurality of terminal devices, analytics functionality, social media, functions for controlling or otherwise interacting with remote devices, functions for an alarm and surveillance center, or any other such function performed by a server.

[0131] As a whole, the communication system 900 of FIG. 9 enables connectivity between the terminal devices, network nodes, and hosts. In that sense, the communication system may be configured to operate according to predefined rules or procedures, such as specific standards that include, but are not limited to: Global System for Mobile Communications (GSM); Universal Mobile Telecommunications System (UMTS); Long Term Evolution (LTE), and/or other suitable 2G, 3G, 4G, 5G standards, or any applicable future generation standard (e.g., 6G); wireless local area network (WLAN) standards, such as the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standards (WiFi); and/or any other appropriate wireless communication standard, such as the Worldwide Interoperability for Microwave Access (WiMax), Bluetooth, Z-Wave, Near Field Communication (NFC) ZigBee, LiFi, and/or any low-power wide-area network (LPWAN) standards such as LoRa and Sigfox.

[0132] In some examples, the telecommunication network 902 is a cellular network that implements 3GPP standardized features. Accordingly, the telecommunications network 902 may support network slicing to provide different logical networks to different devices that are connected to the telecommunication network 902. For example, the telecommunications network 902 may provide Ultra Reliable Low Latency Communication (URLLC) services to some terminal devices, while providing Enhanced Mobile Broadband (eMBB) services to other terminal devices, and/or Massive Machine Type Communication (mMTC)/Massive loT services to yet further terminal devices.

[0133] In some examples, the terminal devices 912 are configured to transmit and/or receive information without direct human interaction. For instance, a terminal device may be designed to transmit information to the access network 904 on a predetermined schedule, when triggered by an internal or external event, or in response to requests from the access network 904. Additionally, a terminal device may be configured for operating in single- or multi -RAT or multi-standard mode. For example, a terminal device may operate with any one or combination of Wi-Fi, NR (New Radio) and LTE, i.e. being configured for multi-radio dual connectivity (MR-DC), such as E- UTRAN (Evolved-UMTS Terrestrial Radio Access Network) New Radio - Dual Connectivity (EN-DC).

[0134] In the example, the hub 914 communicates with the access network 904 to facilitate indirect communication between one or more terminal devices (e.g., terminal device 912c and/or 912d) and network nodes (e.g., network node 910b). In some examples, the hub 914 may be a controller, router, content source and analytics, or any of the other communication devices described herein regarding terminal devices. For example, the hub 914 may be a broadband router enabling access to the core network 906 for the terminal devices. As another example, the hub 914 may be a controller that sends commands or instructions to one or more actuators in the terminal devices. Commands or instructions may be received from the terminal devices, network nodes 910, or by executable code, script, process, or other instructions in the hub 914. As another example, the hub 914 may be a data collector that acts as temporary storage for terminal device data and, in some embodiments, may perform analysis or other processing of the data. As another example, the hub 914 may be a content source. For example, for a terminal device that is a VR headset, display, loudspeaker or other media delivery device, the hub 914 may retrieve VR assets, video, audio, or other media or data related to sensory information via a network node, which the hub 914 then provides to the terminal device either directly, after performing local processing, and/or after adding additional local content. In still another example, the hub 914 acts as a proxy server or orchestrator for the terminal devices, in particular in if one or more of the terminal devices are low energy loT devices.

[0135] The hub 914 may have a constant/persistent or intermittent connection to the network node 910b. The hub 914 may also allow for a different communication scheme and/or schedule between the hub 914 and terminal devices (e.g., terminal device 912c and/or 912d), and between the hub 914 and the core network 906. In other examples, the hub 914 is connected to the core network 906 and/or one or more terminal devices via a wired connection. Moreover, the hub 914 may be configured to connect to an M2M service provider over the access network 904 and/or to another terminal device over a direct connection. In some scenarios, terminal devices may establish a wireless connection with the network nodes 910 while still connected via the hub 914 via a wired or wireless connection. In some embodiments, the hub 914 may be a dedicated hub - that is, a hub whose primary function is to route communications to/from the terminal devices from/to the network node 910b. In other embodiments, the hub 914 may be a non-dedicated hub - that is, a device which is capable of operating to route communications between the terminal devices and network node 910b, but which is additionally capable of operating as a communication start and/or end point for certain data channels. [0136] FIG.10 illustrates a simplified block diagram of a network node 1000 that is suitable for implementing embodiments of the present disclosure. As used herein, network node refers to equipment capable, configured, arranged and/or operable to communicate directly or indirectly with a terminal device and/or with other network nodes or equipment, in a telecommunication network. Examples of network nodes include, but are not limited to, access points (APs) (e.g., radio access points), base stations (BSs) (e.g., radio base stations, Node Bs, evolved Node Bs (eNBs) and NR NodeBs (gNBs)). In some embodiments of the present disclosure, the FSF 210 and/or the model provider 110 as discussed above may be implemented as or included in the network node 1000.

[0137] Base stations may be categorized based on the amount of coverage they provide (or, stated differently, their transmit power level) and so, depending on the provided amount of coverage, may be referred to as femto base stations, pico base stations, micro base stations, or macro base stations. A base station may be a relay node or a relay donor node controlling a relay. A network node may also include one or more (or all) parts of a distributed radio base station such as centralized digital units and/or remote radio units (RRUs), sometimes referred to as Remote Radio Heads (RRHs). Such remote radio units may or may not be integrated with an antenna as an antenna integrated radio. Parts of a distributed radio base station may also be referred to as nodes in a distributed antenna system (DAS).

[0138] Other examples of network nodes include multiple transmission point (multi-TRP) 5G access nodes, multi-standard radio (MSR) equipment such as MSR BSs, network controllers such as radio network controllers (RNCs) or base station controllers (BSCs), base transceiver stations (BTSs), transmission points, transmission nodes, multi-cell/multicast coordination entities (MCEs), Operation and Maintenance (O&M) nodes, Operations Support System (OSS) nodes, Self-Organizing Network (SON) nodes, positioning nodes (e.g., Evolved Serving Mobile Location Centers (E-SMLCs)), and/or Minimization of Drive Tests (MDTs).

[0139] The network node 1000 includes a processing circuitry (including a processor(s)) 1002, a memory 1004, a communication interface 1006, and a power source 1008. The network node 1000 may be composed of multiple physically separate components (e.g., a NodeB component and a RNC component, or a BTS component and a BSC component, etc.), which may each have their own respective components. In certain scenarios in which the network node 1000 comprises multiple separate components (e.g., BTS and BSC components), one or more of the separate components may be shared among several network nodes. For example, a single RNC may control multiple NodeBs. In such a scenario, each unique NodeB and RNC pair, may in some instances be considered a single separate network node. In some embodiments, the network node 1000 may be configured to support multiple radio access technologies (RATs). In such embodiments, some components may be duplicated (e.g., separate memory 1004 for different RATs) and some components may be reused (e.g., a same antenna 1010 may be shared by different RATs). The network node 1000 may also include multiple sets of the various illustrated components for different wireless technologies integrated into the network node 1000, for example GSM, WCDMA, LTE, NR, WiFi, Zigbee, Z-wave, LoRaWAN, Radio Frequency Identification (RFID) or Bluetooth wireless technologies. These wireless technologies may be integrated into the same or different chip or set of chips and other components within the network node 1000.

[0140] The processing circuitry 1002 may comprise a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application-specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, software and/or encoded logic operable to provide, either alone or in conjunction with other network node 1000 components, such as the memory 1004, to provide network node 1000 functionality.

[0141] In some embodiments, the processing circuitry 1002 includes a system on a chip (SOC). In some embodiments, the processing circuitry 1002 includes one or more of radio frequency (RF) transceiver circuitry 1012 and baseband processing circuitry 1014. In some embodiments, the radio frequency (RF) transceiver circuitry 1012 and the baseband processing circuitry 1014 may be on separate chips (or sets of chips), boards, or units, such as radio units and digital units. In alternative embodiments, part or all of RF transceiver circuitry 1012 and baseband processing circuitry 1014 may be on the same chip or set of chips, boards, or units.

[0142] The memory 1004 may comprise any form of volatile or non-volatile computer- readable memory including, without limitation, persistent storage, solid-state memory, remotely mounted memory, magnetic media, optical media, random access memory (RAM), read-only memory (ROM), mass storage media (for example, a hard disk), removable storage media (for example, a flash drive, a Compact Disk (CD) or a Digital Video Disk (DVD)), and/or any other volatile or non-volatile, non-transitory device-readable and/or computer-executable memory devices that store information, data, and/or instructions that may be used by the processing circuitry 1002. The memory 1004 may store any suitable instructions, data, or information, including a computer program, software, an application including one or more of logic, rules, code, tables, and/or other instructions capable of being executed by the processing circuitry 1002 and utilized by the network node 1000. The memory 1004 may be used to store any calculations made by the processing circuitry 1002 and/or any data received via the communication interface 1006. In some embodiments, the processing circuitry 1002 and memory 1004 is integrated.

[0143] The communication interface 1006 is used in wired or wireless communication of signaling and/or data between a network node, access network, and/or terminal device. As illustrated, the communication interface 1006 comprises port(s)/terminal(s) 1016 to send and receive data, for example to and from a network over a wired connection. The communication interface 1006 also includes radio front-end circuitry 1018 that may be coupled to, or in certain embodiments a part of, the antenna 1010. Radio front-end circuitry 1018 comprises filters 1020 and amplifiers 1022. The radio front-end circuitry 1018 may be connected to an antenna 1010 and processing circuitry 1002. The radio front-end circuitry may be configured to condition signals communicated between antenna 1010 and processing circuitry 1002. The radio front-end circuitry 1018 may receive digital data that is to be sent out to other network nodes or terminal devices via a wireless connection. The radio front-end circuitry 1018 may convert the digital data into a radio signal having the appropriate channel and bandwidth parameters using a combination of filters 1020 and/or amplifiers 1022. The radio signal may then be transmitted via the antenna 1010. Similarly, when receiving data, the antenna 1010 may collect radio signals which are then converted into digital data by the radio front-end circuitry 1018. The digital data may be passed to the processing circuitry 1002. In other embodiments, the communication interface may comprise different components and/or different combinations of components.

[0144] In certain alternative embodiments, the network node 1000 does not include separate radio front-end circuitry 1018, instead, the processing circuitry 1002 includes radio front-end circuitry and is connected to the antenna 1010. Similarly, in some embodiments, all or some of the RF transceiver circuitry 1012 is part of the communication interface 1006. In still other embodiments, the communication interface 1006 includes one or more ports or terminals 1016, the radio front-end circuitry 1018, and the RF transceiver circuitry 1012, as part of a radio unit (not shown), and the communication interface 1006 communicates with the baseband processing circuitry 1014, which is part of a digital unit (not shown).

[0145] The antenna 1010 may include one or more antennas, or antenna arrays, configured to send and/or receive wireless signals. The antenna 1010 may be coupled to the radio front-end circuitry 1018 and may be any type of antenna capable of transmitting and receiving data and/or signals wirelessly. In certain embodiments, the antenna 1010 is separate from the network node 1000 and connectable to the network node 1000 through an interface or port.

[0146] The antenna 1010, communication interface 1006, and/or the processing circuitry 1002 may be configured to perform any receiving operations and/or certain obtaining operations described herein as being performed by the network node. Any information, data and/or signals may be received from a terminal device, another network node and/or any other network equipment. Similarly, the antenna 1010, the communication interface 1006, and/or the processing circuitry 1002 may be configured to perform any transmitting operations described herein as being performed by the network node. Any information, data and/or signals may be transmitted to a terminal device, another network node and/or any other network equipment.

[0147] The power source 1008 provides power to the various components of the network node 1000 in a form suitable for the respective components (e.g., at a voltage and current level needed for each respective component). The power source 1008 may further comprise, or be coupled to, power management circuitry to supply the components of the network node 1000 with power for performing the functionality described herein. For example, the network node 1000 may be connectable to an external power source (e.g., the power grid, an electricity outlet) via an input circuitry or interface such as an electrical cable, whereby the external power source supplies power to power circuitry of the power source 1008. As a further example, the power source 1008 may comprise a source of power in the form of a battery or battery pack which is connected to, or integrated in, power circuitry. The battery may provide backup power should the external power source fail.

[0148] Embodiments of the network node 1000 may include additional components beyond those shown in FIG.10 for providing certain aspects of the network node’s functionality, including any of the functionality described herein and/or any functionality necessary to support the subject matter described herein. For example, the network node 1000 may include user interface equipment to allow input of information into the network node 1000 and to allow output of information from the network node 1000. This may allow a user to perform diagnostic, maintenance, repair, and other administrative functions for the network node 1000.

[0149] In some example embodiments, an apparatus capable of performing any of the method 800 (for example, FSF 210 in FIG. 2A and/or FIG. 2B) may comprise means for performing the respective operations of the method 800. The means may be implemented in any suitable form. For example, the means may be implemented in a circuitry or software module. The apparatus may be implemented as or included in FSF 210 or the model provider 110.

[0150] In some embodiments, the apparatus comprises means for, in response to a feature selection trigger for a first machine learning model, determining a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task based on a current set of input features, the current set of input features being different from the target set of input features; and means for causing a second machine learning model to be provisioned to perform the analytic task based on the determined target set of input features.

[0151] In some embodiments, the feature selection trigger is initiated based on at least one of the following: an expiration of a timer started from a last feature selection trigger, the number of analytic requests for the analytic task exceeding a threshold number, an evaluation level of the first machine learning model being below a threshold level, detection of a change in at least one feature for the first machine learning model, detection of a change in context information related to the analytic task, reception of a feature selection request from a network manager, detection of a new feature for enhancing the first machine learning model, or detection of an update of a use case context.

[0152] In some embodiments, the target set of input features are determined further based on a reference set of input features for the analytic task.

[0153] In some embodiments, the reference set of input features are determined to allow a baseline evaluation level of the first machine learning model, or the reference set of input features comprises the current set of input features for the first machine learning model.

[0154] In some embodiments, the means for determining the target set of input features comprises: means for determining at least one candidate set of input features from the reference set of input features; means for determining, given a condition of the context information, at least one predicted evaluation level of a machine learning model with the at least one candidate set of input features as a model input; and means for selecting the target set of input features from the at least one candidate set of input features based on the at least one predicated evaluation level.

[0155] In some embodiments, the target set of input features are determined through at least one of the following: a reinforcement learning, RL, process, a contextual bandit process, a Monte- Carlo tree search process, or a collaborative filtering process using a recommendation engine.

[0156] In some embodiments, the communication network comprises a core network, CN, and the apparatus comprises a model training logical function, MTLF, in a network data and analytics function, NWDAF, or a separate network function outside of the MTLF.

[0157] In some embodiments, the feature selection trigger is received from an analytical logical function, AnLF, the apparatus further comprising: means for notifying the AnLF of the target set of input features.

[0158] In some embodiments, the communication network comprises an open radio access network, O-RAN. In some embodiments, the apparatus is deployed at a non-real time RAN intelligence controller, non-RT RIC, or at a near RT RIC within the O-RAN.

[0159] In some embodiments, the means for causing the second machine learning model to be provisioned comprises: means for causing the first machine learning model to be retrained with the target set of input features, to obtain the second machine learning model; means for causing a trained machine learning model to be reconfigured with the target set of input features as the second machine learning model; or means for causing the second machine learning model to be selected from a plurality of candidate trained machine learning models. [0160] In some embodiments, the context information comprises at least one of the following: context information received from one or more further network nodes within the communication network, or context information received from a third party.

[0161] In some embodiments, the target set of input features is used as a recommendation for determining a set of input features for the second machine learning model.

[0162] In some example embodiments, the apparatus further comprises means for performing other operations in some example embodiments of the method 800 or FSF 210. In some example embodiments, the means comprises at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the performance of the apparatus.

[0163] In some embodiments, a computer readable storage medium having instructions stored thereon is provided. The instructions when executed by at least one processor can cause the at least one processor to carry out the functionality in accordance with any one of the embodiments described herein. In some embodiments, the computer readable medium may be a non-transitory computer readable storage medium. A computer readable storage medium may include but not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the computer readable storage medium would include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a read-only memory (ROM), a random access memory (RAM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0164] In some embodiments, a computer program comprising instructions is provided. The instructions, when executed by at least one processor, cause the at least one processor to carry out the functionality in accordance with any one of the embodiments described herein. In one embodiment, a carrier containing the aforementioned computer program product is provided. The carrier is one of an electronic signal, an optical signal, a radio signal, or a computer-readable storage medium (e.g., a non-transitory computer-readable medium).

[0165] Although the computing devices described herein (e.g., network nodes) may include the illustrated combination of hardware components, other embodiments may comprise computing devices with different combinations of components. It is to be understood that these computing devices may comprise any suitable combination of hardware and/or software needed to perform the tasks, features, functions and methods disclosed herein. Determining, calculating, obtaining or similar operations described herein may be performed by processing circuitry, which may process information by, for example, converting the obtained information into other information, comparing the obtained information or converted information to information stored in the network node, and/or performing one or more operations based on the obtained information or converted information, and as a result of said processing making a determination. Moreover, while components are depicted as single boxes located within a larger box, or nested within multiple boxes, in practice, computing devices may comprise multiple different physical components that make up a single illustrated component, and functionality may be partitioned between separate components. For example, a communication interface may be configured to include any of the components described herein, and/or the functionality of the components may be partitioned between the processing circuitry and the communication interface. In another example, non- computationally intensive functions of any of such components may be implemented in software or firmware and computationally intensive functions may be implemented in hardware.

[0166] In certain embodiments, some or all of the functionality described herein may be provided by processing circuitry executing instructions stored on in memory, which in certain embodiments may be a computer program product in the form of a non-transitory computer- readable storage medium. In alternative embodiments, some or all of the functionality may be provided by the processing circuitry without executing instructions stored on a separate or discrete device-readable storage medium, such as in a hard-wired manner. In any of those particular embodiments, whether executing instructions stored on a non-transitory computer-readable storage medium or not, the processing circuitry can be configured to perform the described functionality. The benefits provided by such functionality are not limited to the processing circuitry alone or to other components of the computing device, but are enjoyed by the computing device as a whole, and/or by end users and a wireless network generally.

[0167] Any appropriate steps, methods, features, functions, or benefits disclosed herein may be performed through one or more functional units or modules of one or more virtual apparatuses. Each virtual apparatus may comprise a number of these functional units. These functional units may be implemented via processing circuitry, which may include one or more microprocessor or microcontrollers, as well as other digital hardware, which may include DSPs, special-purpose digital logic, and the like. The processing circuitry may be configured to execute program code stored in memory, which may include one or several types of memory such as ROM, RAM, cache memory, flash memory devices, optical storage devices, etc. Program code stored in memory includes program instructions for executing one or more telecommunications and/or data communications protocols as well as instructions for carrying out one or more of the techniques described herein. In some embodiments, the processing circuitry may be used to cause the respective functional unit to perform corresponding functions according one or more embodiments of the present disclosure. [0168] While processes in the figures may show a particular order of operations performed by certain embodiments of the present disclosure, it should be understood that such order is exemplary (e.g., alternative embodiments may perform the operations in a different order, combine certain operations, overlap certain operations, etc.). [0169] Those skilled in the art will recognize improvements and modifications to the embodiments of the present disclosure. All such improvements and modifications are considered within the scope of the concepts disclosed herein.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method (800) for machine learning model feature selection in a communication network, the method (800) comprising: in response to a feature selection trigger for a first machine learning model, determining (810), by a network node (1000), a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task based on a current set of input features, the current set of input features being different from the target set of input features; and causing (820), by the network node (1000), a second machine learning model to be provisioned to perform the analytic task based on the determined target set of input features.

2. The method (800) according to claim 1, wherein the feature selection trigger is initiated based on at least one of the following: an expiration of a timer started from a last feature selection trigger, the number of analytic requests for the analytic task exceeding a threshold number, an evaluation level of the first machine learning model being below a threshold level, detection of a change in at least one feature for the first machine learning model, detection of a change in context information related to the analytic task, reception of a feature selection request from a network manager, detection of a new feature for enhancing the first machine learning model, or detection of an update of a use case context.

3. The method (800) according to claim 1 or 2, wherein the target set of input features are determined further based on a reference set of input features for the analytic task.

4. The method (800) according to claim 3, wherein the reference set of input features are determined to allow a baseline evaluation level of the first machine learning model, or the reference set of input features comprises the current set of input features for the first machine learning model.

5. The method (800) according to claim 3 or 4, wherein determining (810) the target set of input features comprises: determining at least one candidate set of input features from the reference set of input features; determining, given a condition of the context information, at least one predicted evaluation level of a machine learning model with the at least one candidate set of input features as a model input; and selecting the target set of input features from the at least one candidate set of input features based on the at least one predicated evaluation level.

6. The method (800) according to any of claims 1 to 5, wherein the target set of input features are determined through at least one of the following: a reinforcement learning, RL, process, a contextual bandit process, a Monte-Carlo tree search process, or a collaborative filtering process using a recommendation engine.

7. The method (800) according to any of claims 1 to 6, wherein the communication network comprises a core network, CN, and wherein the network node (1000) comprises a model training logical function, MTLF, in a network data and analytics function, NWDAF, or a separate network function outside of the MTLF.

8. The method (800) according to claim 7, wherein the feature selection trigger is received from an analytical logical function, AnLF, the method (800) further comprising: notifying the AnLF of the target set of input features.

9. The method (800) according to any of claims 1 to 6, wherein the communication network comprises an open radio access network, O-RAN, and wherein the network node (1000) is deployed at a non-real time RAN intelligence controller, non-RT RIC, or at a near RT RIC within the O-RAN.

10. The method (800) according to any of claims 1 to 9, wherein causing (820) the second machine learning model to be provisioned comprises: causing the first machine learning model to be retrained with the target set of input features, to obtain the second machine learning model; causing a trained machine learning model to be reconfigured with the target set of input features as the second machine learning model; or causing the second machine learning model to be selected from a plurality of candidate trained machine learning models.

11. The method (800) according to any of claims 1 to 10, wherein the context information comprises at least one of the following: context information received from one or more further network nodes within the communication network, or context information received from a third party.

12. The method (800) according to any of claims 1 to 11, wherein the target set of input features is used as a recommendation for determining a set of input features for the second machine learning model.

13. A network node (1000) comprising: at least one processor (1002); at least one memory (1004) coupled to the at least one processor (1002), the at least one memory (1004) comprising instructions that when executed by the at least one processor (1002) implement a method (800) for machine learning model feature selection in a communication network, the method comprising: in response to a feature selection trigger for a first machine learning model, determining (810) a target set of input features for an analytic task based on context information related to the analytic task, the first machine learning model being currently provisioned for performing the analytic task, the current set of input features being different from the target set of input features; and causing (820) a second machine learning model to be provisioned to perform the analytic task based on the target set of input features.

14. The network node (1000) according to claim 13, wherein the feature selection trigger is initiated based on at least one of the following: an expiration of a timer started from a last feature selection trigger, the number of analytic requests for the analytic task exceeding a threshold number, an evaluation level of the first machine learning model being below a threshold level, detection of a change in at least one feature for the first machine learning model, detection of a change in context information related to the analytic task, reception of a feature selection request from a network manager, detection of a new feature for enhancing the first machine learning model, or detection of an update of a use case context.

15. The network node (1000) according to claim 13 or 14, wherein the target set of input features are determined further based on a reference set of input features for the analytic task.

16. The network node (1000) according to claim 15, wherein the reference set of input features are determined to allow a baseline evaluation level of the first machine learning model, or the reference set of input features comprises the current set of input features for the first machine learning model.

17. The network node (1000) according to claim 15 or 16, wherein determining (810) the target set of input features comprises: determining at least one candidate set of input features from the reference set of input features; determining, given a condition of the context information, at least one predicted evaluation level of a machine learning model with the at least one candidate set of input features as a model input; and selecting the target set of input features from the at least one candidate set of input features based on the at least one predicated evaluation level.

18. The network node (1000) according to any of claims 13 to 17, wherein the target set of input features are determined through at least one of the following: a reinforcement learning, RL, process, a contextual bandit process, a Monte-Carlo tree search process, or a collaborative filtering process using a recommendation engine.

19. The network node (1000) according to any of claims 13 to 18, wherein the communication network comprises a core network, CN, and wherein the network node comprises a model training logical function, MTLF, in a network data and analytics function, NWDAF, or a separate network function outside of the MTLF.

20. The network node (1000) according to claim 19, wherein the feature selection trigger is received from an analytical logical function, AnLF, the method further comprising: notifying the AnLF of the target set of input features.

21. The network node (1000) according to any of claims 13 to 20, wherein the communication network comprises an open radio access network, O-RAN, and wherein the network node is deployed at anon-real time RAN intelligence controller, non- RT RIC, or at a near RT RIC within the O-RAN.

22. The network node (1000) according to any of claims 13 to 21, wherein causing the second machine learning model to be provisioned comprises: causing the first machine learning model to be retrained with the target set of input features, to obtain the second machine learning model; causing a trained machine learning model to be reconfigured with the target set of input features as the second machine learning model; or causing the second machine learning model to be selected from a plurality of candidate trained machine learning models.

23. The network node (1000) according to any of claims 13 to 22, wherein the context information comprises at least one of the following: context information received from one or more further network nodes within the communication network, or context information received from a third party.

24. The network node (1000) according to any of claims 13 to 23, wherein the target set of input features is used as a recommendation for determining a set of input features for the second machine learning model.

25. A computer readable medium having instructions stored thereon, the instructions when executed by at least one processor (1002) causing the at least one processor (1002) to perform a method (800) according to any of claims 1 to 12.

26. A computer program comprising instructions which, when executed on at least one processor, cause the at least one processor (1002) to carry out a method (800) according to any of claims 1 to 12.