Consistency between training and inference of AI / ML models

Abstract

Network condition information from two different time periods is used for a predictive model. This may involve comparing new conditions with the historic data used for training a model, and may trigge

Description

TECHNICAL FIELD Various example embodiments of this disclosure relate to one or more methods, devices and / or computer programs relating to consistency between training and inference of AI / ML models. BACKGROUND In general, the present disclosure relates to predictive models or predictors, which are typically known as machine-learning models. While it is commonly referred to machine-learning models, the present disclosure is also applicable for other predictive models that output or infer a prediction based on input data like e.g. different filters such as a Kalman filter, a Linear quadratic estimator, probabilistic estimators or predictors like a Bayes estimator or a Bayes optimal classifier, or other data driven methods. Example embodiments of the present disclosure although not limited thereto, relate to the problem of ensuring consistency between the conditions (for collection of the input data) during training and inference of machine-learning (ML) models. Ideally, conditions during inference, such as network (NW) conditions (e.g., reference signal configuration), should be the same as the conditions during training or at least similar in order to achieve high performance. However, in practice, this is hardly ever likely. For example, a scenario could be considered, in which a device trains an ML model with measurements and / or labels collected in the past under a set X of NW conditions (e.g., reference signal configuration, Transmission Reception Point (TRP) IDs, reference signal configuration of candidate TRPs, spatial direction information, resources sets etc.,). Afterwards, the device wants to use the trained ML model for inference. However, at the time of inference, the network may have a different set X' of conditions. Ideally, X should be the same as X'. However, in practice, it is hardly ever likely that X will be the same as X'. That is, a network may change its conditions dynamically overtime, e.g. due to the timevarying nature of network load or user demand, as well as availability of network resources, for example with respect to positioning service frequency and requirements, user traffic, etc. Thus, it is an object of the present disclosure how to check and ensure such consistency between conditions assumed during training and inference. SUMMARY Various example embodiments of this disclosure aim at addressing at least part of the issue and / or problems and drawbacks either explicitly described herein or otherwise apparent to a person skilled in the relevant art(s) to provide methods, devices, computer programs, and / or systems by which, in particular but not exclusively, consistency between training and inference of AI / ML models can be ensured. According to at least some example embodiments of the present disclosure, network conditions information is stored in association with time information including at least one timestamp. Thus, it is possible to retrieve respective network conditions under which training data for training a machine-learning model was collected at a later point in time. Based on the retrieved network conditions, it is possible to check whether the same or similar network conditions as for collecting the training data are prevailing when the trained machine-learning is to be used for inference, e.g. by feeding data collected at a later point in time into the trained machine-learning model. According to a first aspect, there is provided an apparatus (200), comprising: at least one processor (210), and at least one memory (220) storing computer program codes that, when executed by the at least one processor, cause the apparatus at least to: record network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn); retrieve recorded network conditions information associated with at least one timestamp (ti) indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; compare the retrieved network conditions information (Xi) with network conditions information (X') of a second time (tj) being different from the first time; transmit, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time. According to an example aspect, the transmitted result of comparing comprises at least one of the following: - information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to receive time information comprising the at least one timestamp (ti) indicative of the first time from the terminal device. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to: transmit network configuration information for obtaining the data, with which the predictive model is trained, to the terminal device; and store at least one timestamp (ti) at the time of transmitting the network configuration information as the at least one timestamp (ti) indicative of the first time. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to: perform retrieving the recorded network conditions information, comparing, and transmitting the result of comparing to the terminal device in response to receiving the time information. According to an example aspect, the network conditions information (Xo, ..., Xn) is recorded further in association with the at least one timestamp (ti) indicative of the first time. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to: compare the retrieved network conditions information (Xi) with network conditions information (X') of the second time, and in response to determining that the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are different, transmit, to the terminal device, the result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to: receive, from a third-party device or second node, the data, with which the predictive model is trained, and transmit the received data to the terminal device. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to: receive information about the trained predictive model, and store the received information about the trained predictive model. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to a second aspect, there is provided an apparatus (100), comprising: at least one processor (110), and at least one memory (120) storing computer program codes that when executed by the at least one processor cause the apparatus at least to: receive, from a network entity, a result of comparing network conditions information (XQ associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; and decide, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the received result of comparing comprises at least one of the following: - information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to perform at least one of model monitoring or model retraining before feeding the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, in response to determining that at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (Xz) of the second time are different, - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (Xz) of the second time indicate a difference not included in a predefined set of differences, - the amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (Xz) of the second time is more than a predetermined threshold, or - the received result of comparing includes an indication not to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to decide to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, in response to determining that at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (Xz) of the second time are substantially the same, or - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time do not indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and the parameters of the network conditions information (X') of the second time is less than a predetermined threshold, or - the received result of comparing includes an indication to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to transmit, to the network entity, time information comprising the at least one timestamp (ti) indicative of the first time. According to an example aspect, transmitting the time information to the network entity is performed in response to training the predictive model. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to receive the data used for training the predictive model from the network entity. According to an example aspect, the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to transmit, in addition to the time information, information about the predictive model to the network entity. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to an example aspect, the data, with which predictive model is trained, obtained at the first time comprises at least one of: - a downlink reference signal configuration, - a measurement regarding a transmitted downlink reference signal, - time information indicating a time of performing the measurement regarding the transmitted downlink reference signal, - a quality indicator indicating a quality of the measurement, or - ground truth labels. According to a third aspect, there is provided a method, comprising: recording network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn); retrieving recorded network conditions information associated with at least one timestamp (ti) indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; comparing the retrieved network conditions information (XQ with network conditions information (Xz) of a second time, the second time being different from the first time; transmitting, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time. According to an example aspect, the transmitted result of comparing includes at least one of the following: information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the method further comprises receiving time information comprising the at least one timestamp (ti) indicative of the first time. According to an example aspect, the method further comprises: transmitting network configuration information for obtaining the data, with which the predictive model is trained, to the terminal device; and storing at least one timestamp (ti) at the time of transmitting the network configuration information as the at least one timestamp (ti) indicative of the first time. According to an example aspect, retrieving the recorded network conditions information, comparing, and transmitting the result of comparing to the terminal device are performed in response to receiving the time information. According to an example aspect, the network conditions information (Xo, ..., Xn) is recorded further in association with the at least one timestamp (ti) indicative of the first time. According to an example aspect, the method further comprises: comparing the retrieved network conditions information (Xi) with network conditions information (X') of the second time, and in response to determining that the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are different, transmitting, to the terminal device, the result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time in response to the comparing. According to an example aspect, the method further comprises: receiving, from a third-party device or second node, the data, with which the predictive model is trained, and transmitting the received data to the terminal device. According to an example aspect, the method further comprises: receiving information about the trained predictive model, and recording the received information about the trained predictive model. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to a fourth aspect there is provided a method, comprising: receiving, from a network entity, a result of comparing network conditions information (Xi) associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with the data obtained under the network conditions of the first time; and deciding, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the method further comprises transmitting, to the network entity, time information including the at least one timestamp (ti) indicative of the first time. According to an example aspect, the received result of comparing includes at least one of the following: - information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, or - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, or - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the method further comprises at least one of model monitoring or model retraining before feeding the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, if at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are different, or - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time is more than a predetermined threshold, or - the received result of comparing includes an indication not to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the method further comprises deciding to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, if at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, or - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time do not indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and the parameters of network conditions information (Xz) of the second time is less than a predetermined threshold, or - the received result of comparing includes an indication to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, transmitting the time information to the network entity is performed in response to training the predictive model. According to an example aspect, the method further comprises receiving the data used for training the predictive model from the network entity. According to an example aspect, the method further comprises transmitting, in addition to the time information, information about the predictive model to the network entity. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to an example aspect, the data, with which the predictive model is trained, obtained under the retrieved network conditions information (Xi) comprises at least one of: - a downlink reference signal configuration, or - a measurement regarding a transmitted downlink reference signal, or - time information indicating a time of performing the measurement regarding the transmitted downlink reference signal, or - a quality indicator indicating a quality of the measurement, or - ground truth labels. According to an fifth aspect, there is provided an apparatus, comprising: storing means for recording network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn); processing means for retrieving recorded network conditions information associated with at least one timestamp (ti) indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; wherein the processing means is further configured for comparing the retrieved network conditions information (Xi) with network conditions information (X') of a second time, the second time being different from the first time; and means for transmitting, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time. According to an example aspect, the transmitted result of comparing includes at least one of the following: information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the apparatus further comprises means for receiving time information comprising the at least one timestamp (b) indicative of the first time. According to an example aspect, the means for transmitting is further configured for transmitting network configuration information for obtaining the data, with which the predictive model is trained, to the terminal device; and the means for storing is further configured for storing at least one timestamp (ti) at the time of transmitting the network configuration information as the at least one timestamp (ti) indicative of the first time. According to an example aspect, retrieving the recorded network conditions information, comparing, and transmitting the result of comparing to the terminal device are performed in response to receiving the time Information. According to an example aspect, the storing means is further configured for recording the network conditions information (Xo, ..., Xn) further in association with the at least one timestamp (ti) indicative of the first time. According to an example aspect, the processing means is configured for comparing the retrieved network conditions information (Xi) with network conditions information (X') of the second time, and in response to determining that the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are different, the means for transmitting is further configured for transmitting, to the terminal device, the result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time in response to the comparing. According to an example aspect, the means for receiving is further configured for receiving, from a third-party device or second node, the data, with which the predictive model is trained, and transmitting the received data to the terminal device. According to an example aspect, the means for receiving is further configured for receiving information about the trained predictive model, and recording the received information about the trained predictive model. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to a sixth aspect there is provided an apparatus, comprising: means for receiving, from a network entity, a result of comparing network conditions information (Xi) associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with the data obtained under the network conditions of the first time; and processing means for deciding, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the apparatus further comprises means for transmitting, to the network entity, time information including the at least one timestamp (ti) indicative of the first time. According to an example aspect, the received result of comparing includes at least one of the following: - information indicating whether the retrieved network conditions information (Xi) and the network conditions information (Xz) of the second time are substantially the same, or - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, or - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the processing means is configured for performing at least one of model monitoring or model retraining before feeding the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, if at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are different, or - the information indicative of the differences between the retrieved network conditions Information (Xi) and the network conditions Information (X') of the second time indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time is more than a predetermined threshold, or - the received result of comparing includes an indication not to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, the processing means is configured for deciding to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, if at least one of the following conditions is met: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, or - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time do not indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and the parameters of network conditions information (X') of the second time is less than a predetermined threshold, or - the received result of comparing includes an indication to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. According to an example aspect, transmitting the time information to the network entity is performed in response to training the predictive model. According to an example aspect, the means for receiving is further configured for receiving the data used for training the predictive model from the network entity. According to an example aspect, the means for transmitting is further configured for transmitting, in addition to the time information, information about the predictive model to the network entity. According to an example aspect, the at least one timestamp (ti) indicates a predefined time slot. According to an example aspect, the data, with which the predictive model is trained, obtained under the retrieved network conditions information (Xi) comprises at least one of: - a downlink reference signal configuration, or - a measurement regarding a transmitted downlink reference signal, or - time information indicating a time of performing the measurement regarding the transmitted downlink reference signal, or - a quality indicator indicating a quality of the measurement, or - ground truth labels. According to a seventh aspect, there is provided a computer program product comprising computer-executable computer program code, which when the program is run on a computer, is configured to cause the computer to perform any method according to aspects above. According to an eighth aspect, there is provided non-transitory computer-readable medium comprising program instructions that, when executed by an apparatus, cause the apparatus to perform any method according to aspects above. According to a ninth aspect, there is provided system, comprising: an apparatus according to the first or fifth aspect; and an apparatus according to the second or sixth aspect. According to another aspect, at least one timestamp (ti), that is, one or more or all timestamps (ti) may be in at least one format of UTC or GNSS or network time. A predictive model can be a machine-learning model. By virtue of at least some of the above aspects, one or more of the following advantages can be obtained: - consistency between training and inference of AI / ML models can be ensured; - in particular, consistency between (conditions for data collection of) training data and data fed into a machine-learning model for inference can be ensured; - high performance of a machine-learning model can be achieved; - signaling overhead can be reduced; - redundant data storage of multiple (terminal) devices storing the same data can be avoided; - choosing the best or most appropriate action regarding employing a trained machine-learning model; - reporting overhead can be reduced; - required data storage size can be reduced; - reducing communication overhead; - allowing a network entity to proactively check respective network conditions; - simplifying encoding of exchanged and transmitted information; - differentiating between set of configurations or parameters that are provided by the network for collection of training data and data to be fed into the trained machine-learning model for inference; Further details, features, and advantages will be apparent to those skilled in the relevant art(s) from this disclosure. BRIEF DESCRIPTION OF THE DRAWINGS Some example embodiments will be described in greater detail, by way of nonlimiting and illustrative examples, with reference to the drawings (Figs.), in which: Fig. 1 illustrates a signaling diagram according to at least some example embodiments of the present disclosure; Fig. 2 illustrates a signaling diagram according to at least some example embodiments of the present disclosure; Fig. 3 illustrates a signaling diagram according to at least some example embodiments of the present disclosure; Fig. 4 illustrates a signaling diagram according to at least some example embodiments of the present disclosure; Fig. 5 illustrates a signaling diagram according to at least some example embodiments of the present disclosure; Fig. 6 (a) illustrates recorded network conditions according to at least some example embodiments of this disclosure; Fig. 6 (b) illustrates recorded network conditions according to at least some example embodiments of this disclosure; Fig. 6 (c) illustrates recorded network conditions according to at least some example embodiments of this disclosure; Fig. 7 (a) illustrates an example for implementing a terminal device according to at least some example embodiments of this disclosure; Fig. 7 (b) illustrates another example for implementing a terminal device according to at least some example embodiments of this disclosure; Fig. 8 (a) illustrates an example for implementing a network entity according to at least some example embodiments of this disclosure; and Fig. 8 (b) illustrates another example for implementing a network entity according to at least some example embodiments of this disclosure. DETAILED DESCRIPTION Herein below, certain example embodiments are described in detail with reference to the accompanying drawings, wherein at least some features of example embodiments can be freely combined with features of other embodiments, unless described otherwise. However, it is to be understood that the description of certain embodiments Is given by way of example only, and that It is by no way intended to be understood as limiting aspects of embodiments to the disclosed details. Moreover, it is to be understood that a device or an apparatus is configured to perform a corresponding method, and a respective method can be executed by a correspondingly configured device or apparatus although in some cases, only the device or apparatus or only the method may be described. Various example embodiments will be described with respect to certain aspects. These aspects are not intended to indicate key or essential features of the various example embodiments, nor are they intended to be used to otherwise limit the scope of this disclosure. Other features, aspects and elements of the various example embodiments will be readily apparent to a person skilled in the art in view of the disclosure. At least some of the following embodiments are described with respect to machinelearning models. Machine-learning models are a particularly common examples for predictive models that have to be trained using training data. A trained predictive model can then be used to infer a prediction by inputting or feeding data other than the training data into the trained predictive model like e.g. a machine-learning model. Other examples for predictive models include any data driven models like e.g. time series analysis models or simulation models, different filters such as a Kalman filter, a Linear quadratic estimator, probabilistic estimators or predictors like a Bayes estimator or a Bayes optimal classifier, or the like. In the context of the present disclosure, the term input data refers to any kind of data to be inputted or fed into a trained predictive model. Data fed into an un-trained predictive model for training is typically referred to as the training data. The term data (both the training data and the input data), may in the context of this disclosure typically refer to channel measurements (e.g., using DL PRS), ground truth labels of what is being estimated, e.g., a UE location or timing information (e.g., TOA), as well as any associated quality information and time stamp with them. This data is typically collected and used by a terminal device like a UE, e.g., to train UE-side models. At least some of the following embodiments are described with respect to positioning of a terminal device like e.g. a UE. However, it is to be understood that the present disclosure is not limited to use cases of positioning and can be applied to various other use case, for which e.g. the input data and / or collection thereof depends on network conditions and / or for which the input data is collected for example by measuring signals transmitted in or by a network, such as e.g. timebased, power-based, phase-based or angle-based measurements. That is, the present disclosure may for example also be applied to using machine-learning models for beam management, for CSI feedback or CSI compression, for power control or for various other use cases of machine-learning. For example, in use cases of machine-learning models for beam management, the network-side entity may be a gNB, whereas in use cases of machine-learning models for positioning, the network-side entity may be a Location Management Function (LMF). Embodiments of the present disclosure may for example be employed in the Radio Access Network (RAN) domain and / or in the Core Network (Core) domain. For example, in the Core Network, a device like the terminal device training and employing a respective machine-learning model might include or be replaced by a functional entity like e.g. a Network Data Analytics Function (NWDAF). On the other hand, a terminal device training and employing a respective machinelearning model in the Radio Access Network (RAN) domain might be for example a user equipment (UE). That is, it is to be understood that a UE is only an example for a device or entity and the present disclosure may be applied to various other entities and devices employing a predictive model like e.g. a machine-learning model. For example, known solutions are based on addressing this challenge via an associated identifier (ID) that is indicative of certain set of configurations or parameters that are provided by the network, so that the UE can differentiate network conditions it uses for training and inference. However, associating an ID to a set of configurations leads to several challenges as follows. First, for example, the size of an ID needs to be too large in order to uniquely identify the values set for configuration parameters given a large number of parameters and large number of values that each parameter can take (e.g., configuration parameters of DL PR.S). In the end, indicating an associated ID is not much different than indicating the actual values of the parameters, e.g., via assistance data provided by NW as in legacy, in terms of signaling and processing overhead. For example, when Associated ID#1 corresponds to a set of parameters with values {Pl=valuel, P2=value2, P3=value3, ...} Then a new ID would be needed for {Pl=value3, P2=value2, P3=value3, ...} thus the number of possible IDs grows factorially. Second, the ID has to remain consistent over time such that after using it for training, the UE can use it later to verify the change in conditions during inference, which may take place at any point of time in the future. More importantly, the IDs needs to be globally aligned across cells, gNBs, and PLMNs, so that the UE can uniquely identify the set of parameter values irrespective of its location or serving network / cell. E.g., while gNB#l assigns ID#1 to {Pl=valuel, P2=value2, P3=value3, the same ID#1 should not be used again by a different gNB for a different set of parameter values. An associated ID may help a terminal device to ensure consistency within a cell, however there are still no guarantees that the associated IDs can be consistent across cells, gNBs, service areas, and PLMNs. This poses a great challenge especially for the positioning use case since the UEs will likely reside not only in a single cell but rather move across the network(s) to a great extent. According to the present disclosure, there are provided example embodiments, according to which a device or a terminal device like for example a UE or another entity like an NWDAF merely stores time information indicating a time for collecting data used as the training data. According to at least some example embodiments, a network entity like e.g. a Location Management Function (LMF) or some base station like e.g. a next Generation Node B a gNB, or a 6G base station logs, that is, records and stores network conditions information in association with a first time information indicating a time of recording said network conditions. The network entity may determine the first time information based on the indication the time for collecting data used as training data received from the device like the UE. Alternatively, the network entity may also determine the first time differently. Subsequently, the network entity can compare network conditions information at the time of collecting or obtaining the training data with current network conditions that are prevailing and measured and / or recorded when a machine-learning model is to be used for inference or shortly before such that it can be assumed that said current network conditions will still prevail at the time when the machine-learning model is used for inference. This way, consistency between (conditions for data collection of) the training data and the input data for inference of a machine-learning model can be ensured without introducing large data storage and communication overheads. In the context of this disclosure, consistency means whether or how the network conditions are different between the network conditions for obtaining the training data and the current conditions that the terminal device will use to perform inference (by conducting measurements). Such difference is an important factor that determines the generalization performance of a predictive model like e.g. an AI / ML model across different settings. Network conditions in the context of this disclosure may include but are not limited to at least one of: • Reference signal (e.g., DL PRS) configuration including signal bandwidth, periodicity, power, duration, beam information (incl. pattern, angle, ID, spatial direction information etc.), mapping to TRP / Antenna Reference Point (ARP) information, etc., or • TRP / ARP information including IDs, location, antenna information, transmit (Tx) and receive (Rx) timing error information, phase error information, etc., or • TRP Line-Of-Sight (LOS) / Non-Line-Of-Sight (NLOS) state information and relative time difference information, or • Geographical / network area and / or site information, gNB / cell list (including serving and neighboring ones), public land mobile network (PLMN) ID, vendor ID or information, etc., or • NW synchronization status and / or error, or • Model input format, or • Measurement data quality range (e.g., in terms of S(I)NR, RSRP, etc.), or • Label quality range (e.g., mean label positioning error), or • Validity area and / or time of above information, or • label data quality range (e.g. mean label positioning error), or • Time range of collecting data, or • Network Synchronization Error, or • Antenna pattern information, or • beam pattern information, or • synchronization signal blocks (SSB) information of TRPs; Furthermore, it is to be noted that respective parts or information included in network conditions can be grouped into classes, e.g. based on importance or functionality, such as geographical conditions, data quality conditions, hardware conditions etc. That is, network conditions, denote the conditions assumed at the NW-side when the data is generated or obtained, such as a PRS configuration, TRP IDs / locations, as well as any network-proprietary (non-standardized) features, e.g., antenna or power amplifier properties, etc. These are set by the network entity (which might be based on some preferences indicated by the UE), and the network entity does not necessarily disclose all details of the conditions to the terminal device or the UE, especially the proprietary ones. The network (NW) conditions can be also referred to as NW-side conditions, additional conditions, NW-side additional conditions, (NW) configurations, or the like. In the following, an example embodiment is described with reference to Fig. 1. In an example embodiment relating to positioning, data to be input or fed into the machine-learning model can for example comprise channel measurements (e.g., using DL PRS), ground truth labels of what is being estimated, e.g., UE location or timing information (e.g., TOA), as well as any associated quality information and time stamp with them. However, it is to be understood that the present disclosure is not limited to a respective sort of data (input data and / or training data) or machine-learning model functionality and various other applications and sorts of data may be used within the scope of this disclosure. In Fig. 1, a UE 100 is an example for a terminal device. The network entity 200 may 200 be for example a location management function (LMF) when positioning is considered as a use case for applying machine-learning. In other example embodiments, the network entity 200 may for example be a gNB. In the example illustrated in Fig. 1, it is assumed that the UE 100 collects the data for training its model in step Sil. Collecting data may also be referred to as obtaining data. It is to be noted that respective data may also be obtained by other devices or entities and the present disclosure is not limited to examples, in which the terminal device obtains data fortraining a machine-learning model. In particular, obtaining respective data may include any measurements using signals transmitted by the network entity, e.g., DL PRS, such as time-based, power-based, phase-based or angle-based measurements for a positioning use case. The measurements could be further associated with other data such as time stamps, quality indicators (e.g., in terms of accuracy, confidence level, etc.), as well as any ground truth labels (or their approximation) associated with measurements, e.g., estimated location information obtained by non-radio access technologies (RAT) positioning techniques, together with their quality indicators. If the collected data itself does not contain any time stamps, then UE 100 logs the time during which the data is collected. According to at least some example embodiments, the time at which the data for training is collected in step Sil may be referred to as a first time. Thus, data for training is collected or obtained under network conditions of or prevailing at the first time. Afterwards, in step S12, the UE 100 performs training of its model using the collected data. It is to be noted that the training may also be performed at an external server of a UE vendor for example, to which the collected data with associated time stamps is provided. In step S13, the network entity 200 logs, that is, records and stores information about the NW-side conditions it provided for data collection, e.g., the DL PRS configuration, together with associated time information (e.g. timestamps when the respective information was recorded) as network conditions information. Such recording of network conditions information may be performed periodically or repeatedly with arbitrary and / or changing timing resolution. That is, the network entity 200 may record network conditions information periodically, e.g. once every minute or once every five minutes, or the network entity 200 may record network conditions information only in response to a specific action or event like for example receiving a message from the UE or sending some information to the UE that might trigger data collection. In this regard, it is to be noted that the network entity 200 may perform step S13 before, while and / or after the terminal device performs steps Sil and S12. It is assumed that at some later point in time, the UE 100 wants to use the trained machine-learning model for interference. For inference, typically data fed into the machine-learning model is collected at the time inference is to be performed. That is, the data for inference may be collected much later than data used for training the machine-learning model. Thus, according to at least some example embodiments, before performing inference, UE 100 indicates (i.e. transmits) to the network entity 200 time information of when the data utilized fortraining the machine-learning model was collected. Such time information is indicative of at least one timestamp (ti) of the collected data. According to at least some example embodiments, the time information may comprise a timestamp indicating the beginning of collecting data for training and another timestamp indicating the end of collecting data for training. According to at least some example embodiments, the time information may comprise one or more timestamps indicating a pre-defined time slot each. Upon or in response to receiving the time information, the network entity 200 retrieves network conditions information (Xi) for the time indicated in the received time information in step S15. Next, in step S16, the network entity 200 compares the retrieved network conditions information (Xi) with current network conditions information (Xz) or network conditions of a second time, e.g. tj. Such current network conditions information (X') may for example be obtained through measuring or recording respective network conditions in response to receiving the time information in step S14 or in response to retrieving the network conditions information (Xi) in step S15. Accordingly, the current network conditions information represent network conditions information (X') of a second time, that is, associated with a second timestamp, wherein the second timestamp may be indicative of a (second time shortly before or infinitively close to the time, when the trained machine-learning model is to be used for inference of a prediction. For example, the second timestamp may indicate (that is, the second time may be) a time of receiving the time information in step S14 or a time of retrieving the network conditions information (Xi) or the like. In the context of this disclosure, the word time refers to an actual time and the word timestamp refers to encoded information indicating a time, e.g. when some Information is or was recorded. That Is, a timestamp may be considered as a sequence of characters or encoded information identifying or indicating a respective point or period in time. Furthermore, the time of inference, i.e., the second time can be also depend on Quality of Service (QoS) requirements associated with the positioning request: For example, QoS may indicate a maximum response time for a positioning or measurement request, and / or a scheduled location time, around which the measurements should take place. Thus, the network entity may arrange the configurations and / or resources accordingly to make the measurements. Hence, inference may happen close to the scheduled location time (if any) or within a maximum response time, if any. Based on this, the network entity may be aware of conditions it will be able to provide not only currently but also around this time, e.g. a scheduled time in the (near) future. Thus, "current" conditions may also relate to the conditions in the near future that inference is expected to happen according to network and accordingly, the second time may be a time in the future. That is, the current or possible or scheduled future network conditions information (X') represent network conditions information that were prevailing just or shortly before using them for the comparison in step S16 such that it can be assumed that these current or future network conditions information (X') are still prevailing at the time the network entity performs step S16. That is, in step S16, the network entity 200 checks whether it can provide (substantially) the same or similar network conditions for inference that is about to take place, as the network conditions under which data used for training the machine-learning model was collected. It is to be understood that the second time may be infinitely close to the present or to the time of comparing network conditions in step S16 or to the time when the machine-learning model trained with data obtained under the network conditions of the first time is to be used. That is, it can be assumed that the network conditions of the second time and the network conditions when obtaining data to be fed into the trained machine-learning model for inference of a prediction are substantially the same or at least similar. Thus, it may also be assumed that data to be fed into the trained machine-learning model for inference of a prediction is also obtained at the second time. In this regard, comparing the retrieved network conditions information (Xi) with currently prevailing or possible or scheduled future network conditions is only an example of comparing the retrieved network conditions information (XQ with network conditions information (X') of a second time. According to at least some example embodiments, steps S15 and S16 may also be combined to one single step. Afterwards, in step S17 the network entity 200 transmits to the UE 100 a result of the comparison in step 16 or information about the result of the comparison. For example, in step S17, the network entity 200 may transmit at least one of the following: - information indicating whether the retrieved network conditions information (Xi) and the network conditions (X') of the second time are substantially the same, or - information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, or - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or - an indication whether to feed data for inference, e.g. collected or obtained under the network conditions (X') of the second time, into the trained machinelearning model trained with the data collected under the retrieved network conditions information (Xi). In this regard, it may be assumed that data to be fed into the machine-learning model for inference is or will be obtained under network conditions (X') of the second time. Furthermore, the transmitted information about the result of the comparison can also depend on some conditions. That is, for example, the information about the result of the comparison may be an indication that the retrieved network conditions information (X,) and the network conditions (X') of the second time are the substantially same if this is the case. For example, additional information about respective differences or amounts of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time may be transmitted only in the case that the retrieved network conditions information (Xi) and the network conditions (X') of the second time are not substantially the same or similar, that is in the case that the retrieved network conditions information (Xi) and the network conditions (X') of the second time are different. That is, the network entity 200 may indicate to the UE 100 whether it can provide substantially the same or similar conditions that the UE used for training its model, or any differences in the conditions, e.g., any different parameter settings as per the current or available (future) conditions that the network entity can provide for the inference to take place. Depending on differences between the network conditions prevailing when collecting data used for training the machine-learning model and network conditions of the second time (i.e. network conditions prevailing when possibly collecting data for inference), e.g. the following actions may be performed in step S18: (a) If the conditions are substantially the same or similar to some extent (the extend being up to the implementation of the terminal device or the network entity), the terminal device or the network entity may decide to directly use the trained model for the inference. (b) If the network conditions largely differ (the extend being up to the implementation of the terminal device or the network entity), the UE or the network entity may trigger monitoring the performance of the model, e.g., with the new conditions, after which the terminal device or the network entity may decide to either use the same model, switch to a different model, or switch to fallback solution, e.g., non-AI / ML method, depending on the performance. Alternatively, or in addition, the UE may request from the network entity more similar conditions to the ones it assumed fortraining, or the UE may request and / or trigger new data collection for further training. Furthermore, it is also conceivable that several machine-learning models were trained at the terminal device and that in case the network conditions largely differ, the terminal device switches to a different trained or untrained machine-learning model. In such case, abovedescribed steps S14 to S18 may be repeated for the other machine-learning model after switching. For example, conditions that actions mentioned in item (a) are performed include at least one of the following: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, or - the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time do not indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time is less than a predetermined threshold, or - the received result of comparing includes an indication to feed data for inference, e.g. collected or obtained under the network conditions (X') of the second time into the trained machine-learning model. For example, conditions that actions mentioned in item (b) are performed include at least one of the following: - the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are not substantially the same, that is different, or - the information Indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X7) of the second time indicate a difference not included in a predefined set of differences, or - the amount of differences between parameters of the retrieved network conditions information (Xi) and the parameters of network conditions information (X') of the second time is more than a predetermined threshold, or - the received result of comparing includes an indication not to feed data collected under the network conditions (X') of the second time into the trained machine-learning model. For example, the terminal device may store e.g. a list indicating a predefined set of differences between network conditions that are acceptable for feeding, into a trained predictive model, the input data obtained under network conditions including such acceptable differences compared to the network conditions under which the training data for training the predictive model was obtained. By way of such example embodiments, consistency between training and inference of AI / ML models can be ensured. In particular, consistency between (conditions for data collection of) the training data and the input data for inference of a machine-learning model can be ensured. Thus, also a high performance of the trained machine-learning model can be achieved, and it is possible to choose the best or most appropriate action regarding employing a trained machine-learning model. In Fig. 2, a modified example embodiment is illustrated. Steps already described with respect to Fig. 1 (Sil to S18) are assumed to be the same as in Fig. 1 if not indicated or described differently and a repeated description will thus be omitted. In addition to the steps described with respect to Fig. 1 or alternatively to step S17, the UE 100 may transmit (in step S21) time information to the network entity 200 in response to training the machine-learning model or even in response to collecting or obtaining the data to be used fortraining the machine-learning model. That is, the UE 100 may transmit the time information to the network entity well before inference using the trained machine-learning model is planned to take place. Based on the received time information, the network entity can store the recorded network conditions information also in association with times during which measurements for collecting or obtaining the data for training the machinelearning model were performed. If these times of collecting or obtaining the data for training the machine-learning model are also stored at the network entity 200, step S14 may be omitted, and the network entity may proactively handle changes in network conditions. For example, the network entity may repeatedly (e.g. periodically or in response to events like changing or setting network configurations) perform steps S15 and S16 as described with respect to Fig. 1. Based thereon, the network entity can inform the UE 100 in response to noticing changes or significant changes when comparing retrieved network conditions information with respective current or future network conditions information, e.g. by transmitting information indicating specific differences between the retrieved network conditions information and the respective current or future network conditions information. That is, transmitting time information to the network entity 200 in response to training the machine-learning model or even in response to collecting or obtaining the data to be used for training the machine-learning model enables allows the network entity to accurately and proactively check and compare respective network conditions in steps S15 to S17 without requiring step S14. By means of such example embodiments, the reporting overhead and transmission overhead can be reduced. In Fig. 3, another modified example embodiment is illustrated. Steps already described with respect to figures 1 and 2 (in particular, Sil to S18 and S21) are assumed to be the same as in figures 1 and 2 if not indicated or described differently, and a repeated description will thus be omitted. In Fig. 3 a signaling diagram according to at least some example embodiments of the present disclosure is illustrated for an example of applying a machine-learning model to positioning. In step S31, a terminal device like the UE 100 may request necessary configurations, e.g., downlink (DL) positioning reference signal (PRS) configuration, for data collection, from NW. Within the request, the terminal device may optionally indicate any preferences for the configuration setting, e.g., the UE 100 may send an on-demand PRS request, in which the UE provides desired parameter settings for a DL PRS configuration. In particular, in an example relating to positioning, the UE 100 requests a PRS configuration to the network entity 200 in step S31. However, in the context of this disclosure, the machine-learning model may also be applied for different use cases, in which a different configuration might be requested, e.g. for different downlink reference signals. In step S32 in response to step S31, the network entity 200 provides (transmits) network configuration information to the UE 100, which can be used for data collection, e.g., by indicating a DL PRS setting via assistance data. The network entity 200 may or may not take into account any preferences of the UE that UE may have indicated before. Such network configuration information can be at least one of, that is, one or more of a reference signal (e.g., DL PRS) configuration, a measurement configuration (e.g., RSRPP for positioning), and a reporting configuration (e.g., periodicity of reporting). Thus, in general, said network configuration information may for example be used for collecting or obtaining the data for training the machinelearning model and / or data for inference using the machine-learning model. Some time later, i.e. after training the machine-learning model is completed, according to at least some example embodiments, the trained machine-learning model may be used e.g. for performing or for assisting position estimation. In a use case relating to position estimation, the network entity 200 may request location information from the UE 100, e.g., upon receiving a Location Service (LCS) positioning request at a location management function (LMF) that is an example for the network entity 200. For example, location information may comprise absolute and / or relative location of the UE, e.g., in 2D / 3D coordinates, as well as any positioning-related intermediate feature that is used for the final position estimate such as timing information (e.g., Time of Arrival, Time Difference of Arrival, Rx-Tx time difference), angle-based information (e.g., Angle of Arrival), a LOS / NLOS indicator, etc. Furthermore, it is to be understood that according to at least some example embodiments, one or both of steps S21 and S14 may be omitted in the example illustrated in Fig. 3. That is, If the UE does not indicate any time information in steps S21 and / or S14, the network entity 200 can assume a time it has provided necessary configurations for data collection (e.g., PRS configuration, measurement configuration) to the UE 100. That is, the network entity can assume the time of sending a configuration in step S32 as the time for collecting or obtaining the data fortraining the machine-learning model. Based on this assumption, the network entity can store a timestamp indicating the time of transmitting the configuration to the terminal device in step S32 as the time information. Thus, according to at least some example embodiments, when retrieving network conditions information (Xi) in step S15, the network entity can retrieve the network conditions information (Xi) for which the associated time equals the stored timestamp indicating the time of transmitting the configuration to the terminal device in step S32. By means of omitting steps S21 and / or S14, the reporting overhead and transmission overhead can be reduced. According to at least some example embodiments, a third entity or also referred to as third-party entity or second node may collect or obtain the data for training the machine-learning model. In the context of this disclosure, a third-party entity refers to any different device or entity than the one (UE) the data is utilized for. Such third-party device may accordingly also be called e.g. a "second device", a "second node", a "second entity" or similar. This allows outsourcing the process of obtaining data from the terminal device. For example, such third-party entity or second node may be a Positioning Reference Unit (PRU) 300 or another UE. A PRU 300 can for example be a UE with a known position. An example embodiment including a PRU is described below with respect to Fig. 4. Fig. 4 illustrates a signaling diagram according to at least some example embodiments. Steps already described with respect to figures 1, 2 and 3 (in particular, Sil to S18, S21 and S31 to S33) are assumed to be the same as in figures 1, 2 and 3 if not indicated or described differently, and a repeated description will thus be omitted. In case, the third-party or second node entity collects or obtains the data for training the machine-learning model, the terminal device does not need to collect data in step Sil, which can thus be omitted. In the example illustrated in Fig. 4, it is assumed that data collection is triggered in that the network entity 200 requests data in step S41. In this regard, the network entity 200 may similar as in step S32 provide configurations to the PRU 300, which can be used for data collection, e.g., by indicating a DL PRS setting via assistance data. In response to receiving a request for data, the PRU 300 collects or obtains the data for training the machine-learning model, similar to how the UE collects data in step Sil described with respect to Fig. 1. That is, in step S42, the PRU 300 may perform any measurements using signals transmitted by the network entity, e.g., DL PRS, such as time-based, power-based, phase-based or angle-based measurements for a positioning use case. The measurements could be further associated with other data such as time stamps, quality indicators (e.g., in terms of accuracy, confidence level, etc.), as well as any ground truth labels (or their approximation) associated with measurements, e.g., estimated location information obtained by non- radio access technologies (RAT) positioning techniques, together with their quality indicators. Typically, the collected data includes time stamps indicating the time of collecting the data. Alternatively or in addition, the UE 100 may log the time during which the data is collected or obtained. The network entity may additionally indicate these timestamps to the UE in Step S44. Afterwards, in step S43, the PRU 300 transmits the collected data to the network entity 200 and in step S44, the network entity 200 forwards the collected data to the UE 100, which may then use the received data to train the machine-learning model in step S12. It is to be understood that also according to at least some example embodiments in line with the example illustrated in Fig. 4, one or both of steps S21 and S14 may be omitted. That is, if the UE 100 does not indicate any time information in steps S21 and / or S14, the network entity 200 can assume a time it has provided necessary configurations for data collection (e.g., PRS configuration, measurement configuration) or the request for data collection to the PRU 300. That is, the network entity 200 can assume the time of sending a configuration or the request to the PRU 300in step S41 as the time for collecting or obtaining the data for training the machine-learning model. Alternatively, the network entity 200 can assume the time of sending a configuration, e.g. the PRS configuration in step S32 of the example illustrated in Fig. 3 as the time for collecting or obtaining the data fortraining the machine-learning model. Alternatively, the network entity 200 can assume a time span, which starts at the time of sending a configuration or the request to the PRU 300 in step S41, and which ends at the time of receiving the collected data from the PRU in step S43, as the time for collecting or obtaining the data for training the machine-learning model. Based on this assumption, the network entity 200 can store one or more timestamps indicating the time of performing step S41 and / or step S43 or all timestamps between the time of performing steps S41 and S43 as the time information. Thus, according to at least some example embodiments, when retrieving network conditions information (Xi) in step S15, the network entity can retrieve the network conditions information (Xi), for which the associated time equals the one or more timestamps stored based on the time of performing step S41 and / or step S43. By means of omitting steps S21 and / or S14, the reporting overhead and transmission overhead can be reduced. Another example embodiment is described with reference to Fig. 5. Fig. 5 illustrates a signaling diagram according to at least some example embodiments. According to at least some example embodiments, the network entity may proactively handle changes in network conditions. That is, in a case of the network entity proactively handling changes in network conditions, the network entity may be updated about collected data (e.g. by receiving respective time information) upon data collection. Thus, the network entity can continuously or repetitively compare network conditions prevailing when the training data was collected with network conditions of the second time. Therefore, when the network conditions change or significantly change, the network entity can transmit a message to the terminal device indicating a change of conditions and / or differences between network conditions of the second time and the network conditions prevailing when the training data was collected. Following this approach, i.e. the network entity proactively handling changes in network conditions, the terminal is only notified when the change occurs, and the network entity only shares the conditions change (and not all the conditions information). This has the advantage of optimizing the related exchanges and reporting overhead as well as alleviating this task from the terminal device, with no need to continuously check if there is any condition change. Based on a respective change notification (e.g. indicating the differences between X' and X), the terminal device or the network entity can decide the best action to be realized accordingly (for example triggering a model monitoring or also a new round of data collection and model retraining) as in the previously explained example embodiments. In the example illustrated in Fig. 5, it is assumed that the UE 100 collects or obtains the data used fortraining the machine-learning model. However, as evident from the description of Fig. 4, this may also be replaced or complemented by one or more other devices collecting respective training data. In the example illustrated in Fig. 4, also step S21 is optional and as describe above, the network entity may also assume the time of sending a configuration in step S32 as the time of collecting or obtaining the data for training the machine-learning model. According to at least some example embodiments, the UE 100 indicates (i.e. transmits a notification) to the network entity 200 that a new model is trained using data collected during the period (T1-T2) (at a time t). Furthermore, the UE 100 may also transmit information about a machine-learning model or a functionality of the machine-learning model, and / or feature ID in step S21. Based thereon (e.g. the indicated machine-learning model functionality, and / or feature ID), the network entity 200 may also know which trained machine-learning model is to be used for inference before the network transmits a request that might trigger inference. For example, when the network entity plans to request location information from the UE 100, it can infer based on stored machine-learning model functionality, and / or feature ID, which trained machine-learning model is to be used for inference at the UE 100. Thus, the network entity may automatically perform step S15 (retrieving network conditions information) and step S16 (comparing retrieved network conditions information with network conditions information of the second time) before transmitting a request to the UE 100 that will lead to inference using a respectively trained machine-learning model. That is, the network entity may perform steps S15 (retrieving network conditions information), S16 (comparing retrieved network conditions information with network conditions information of a second time), and S17 (transmitting the result of the comparison) based on or before or in response to requesting a service, which will lead to usage of the trained machine-learning model for inference. Transmitting such information about the machine-learning model allows improving the proactive case. That is, the network entity can proactively check or compare respective network conditions before a specific model will (likely) be used, e.g. if a certain service using a previously indicated machine-learning model functionality will be requested. Alternatively, steps S15 and S16 may be performed regularly (periodically) and / or in response to network related events like e.g. changing or setting respective network configurations and / or settings. It is to be noted that for implementing a proactive checking of network conditions by the network entity, several options are possible. For example, the network entity can record and store a time, at which it sends network condition information for collecting data to the terminal device or to a third device as a first time for obtaining data for training a machine-learning model. This option reduces the communication overhead and the amount of data to be stored at a terminal device like a UE. In another example, the terminal device may transmit time information comprising at least one timestamp indicative of the (first) time of collecting or obtaining data used fortraining a machine-learning model in response to training the machine-learning model. This option allows the network entity to accurately store the time for obtaining the training data and may thus also improve the accuracy when proactively comparing network conditions information (network conditions for obtaining the training data and for performing inference using the trained machine-learning model). According to at least some example embodiments, the time information can be expressed in terms of pre-defined intervals or time slots for each item of network conditions information like e.g. network configuration parameters, network configuration ID, or any other specification. For example, the network entity may divide the time as Tl: MORNING (05:00-12:00), T2: AFTERNOON (12:00-17:00), and T3: EVENING (17:00-21:00), and T4: NIGHT (21:00-05:00), and indicate such division to the terminal device. In turn, such time slots can be utilized between the terminal device and the network entity to indicate the time when the training data was collected, e.g., instead of absolute time. Using such time slots for encoding time information and timestamps may simplify the encoding of the time information and may also reduce the communication overhead since such time slots may allow to use short codes. For example, if four time slots per day are predefined, the time slots can be encoded using only two bits. According to at least some alternative embodiments, the time information can be indicated in the format of UTC, GNSS or network time, as in existing LPP TS 37.355. With reference to Fig. 6 (a), Fig. 6 (b) and Fig. 6 (c), some examples for recorded network conditions are illustrated. According to at least some example embodiments, the network entity 200 records (that is, measures and stores) network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn) of the network conditions (Xo, ..., Xn) . For readability, the network conditions information is simply indicated with xo, ..., Xn in Fig. 6 (a), Fig. 6 (b) and Fig. 6 (c). Examples that could replace respective entries denoted with Xi may include: reference signal configuration (e.g., bandwidth / periodicity of DL PRS), TRP information (e.g., TRP IDs, TRP locations, antenna information), beam information for each TRP. Note that the features could be both standardized and non-standardized, e.g., for the latter, NW may associate ID for some of its proprietary features, e.g., power amplifier characteristics, etc., and may only let the UE know about the ID rather than what it actually corresponds to internally at NW-side. In Fig. 6 (a), an example for recorded network conditions is illustrated, in which the network entity stores specific network configuration parameters, e.g. indicating a Reference signal (e.g., DL PRS) configuration including signal bandwidth, periodicity, power, duration, beam information (incl. pattern, angle, ID, etc.), mapping to TRP / ARP information, etc. In the example illustrated in Fig. 6 (a), it is assumed that the network entity periodically records respective network configuration parameters every five minutes. In Fig. 6 (b), an example of stored network condition information is illustrated, in which an associated identifier (ID) of a network configuration that is indicative of certain set of configurations or parameters that are provided by the network is stored as network conditions information. For example, every recorded item of network conditions information be assigned a different ID, such that based on the assigned ID, it is possible to uniquely identify a set of parameter values and / or the configuration of the network. Also in the example illustrated in Fig. 6 (b), it is assumed that the network entity periodically records respective network configuration parameters every five minutes. However, it is to be understood that various different sampling rates for recording network condition information may be applicable depending on various use cases and network configurations. That is, the network entity may record respective network configuration parameters for example every 5 or 10 seconds, every minute, twice or once per hour or in some cases only once every day. In the examples illustrated in Fig. 6 (a) and Fig. 6 (b), a respective timestamp is in the format "YYYY-MM-DD-HH-MM". However, various other formats of possible timestamps may be applicable as well. In Fig. 6 (c), another example of stored network condition information is illustrated, in which an associated ID of a network configuration that is indicative of certain set of configurations or parameters that are provided by the network is stored as network conditions information. However, different from the example of Fig. 6 (a) and Fig. 6 (b), in the example illustrated in Fig. 6 (c), predefined intervals or time slots are used for defining timestamps. For example, four different time slots may be used like e.g.: Tl: MORNING (05:00-12:00), T2: AFTERNOON (12:00-17:00), T3: EVENING (17:00-21:00), and T4: NIGHT (21:00-05:00). Alternatively, also more or less time slots may be predefined with respective shorter or longer intervals. In the example illustrated in Fig. 6 (c), the time slot NIGHT (21:00-05:00) is associated with the earlier day of the two days it partly covers. That is, the time slot night is associated with the date of the start of the time slot in this example. However, the time slot night may alternatively also be associated with the date of the end of the time slot for example. The respective network conditions information may be recorded for example at the beginning of a time slot, in the middle of a time slot, at the end of a time slot, or at some arbitrary time within a time slot. Furthermore, several modifications and additions are conceivable as described below. According to at least some example embodiments, the collected data during data collection phase (e.g. step Sil or S42) comprises at least one or all of the following: DL RS configuration, or Any measurement regarding the transmitted DL RS, or Time stamp, or Quality indicator, or Ground truth labels. According to at least some example embodiments, the terminal device may log the time during which the data is collected. However, according to at least some other example embodiments, the network entity may log the time during which the data is collected. According to at least some example embodiments, the data is first collected at a second node (or third-party entity), like e.g. a PR.U and then provided by the network entity to the terminal device for model training. According to at least some example embodiments, the terminal device may send time frame information to the network entity after training the model. The time frame information may for example include two timestamps, one timestamp indicating a start time of collecting the data used for training the machine-learning model and the second time stamp indicating an end time of collecting the data used for training the machine-learning model. Based thereon, multiple items of network conditions information may be retrieved in step S15 and compared to network condition information of the second time. For example, the data used for training the machine-learning model may be collected over a period of time during which network conditions changed. In this case, a respectively trained machine-learning model might be applicable for inference by feeding data into the trained machine-learning model if the data for inference was collected under network conditions similar to any of the network conditions that were prevailing when the data used for training the machinelearning model was collected. According to at least some example embodiments, the terminal device may transmit, to the network entity, information about a functionality or about some features of the AI / ML model, for which the training data was collected by performing measurements in the network. In this regard, the terminal device may transmit, to the network entity, information about the trained machine-learning model or parameters thereof. According to at least some example embodiments, the network entity logs the conditions it provided for data collection together with the associated time information. According to at least some example embodiments, the associated time information comprises: • The time at which the network entity has provided the conditions (e.g. a time of setting a respective configuration or a time of changing settings or parameters), • The time during which a PRU or the UE performed measurements, • The time information indicated by the UE. According to at least some example embodiments, the network entity checks the network conditions information during the training phase and the conditions of the second time and indicates to the terminal device that substantially the same network conditions can be provided. According to at least some example embodiments, in case the respective network conditions information items are (substantially) the same or an amount of differences between parameters of the respective network conditions items' parameters are within a threshold, the terminal device may continue with the inference by feeding the input data into the trained machine-learning model. According to at least some example embodiments, in case the respective network conditions information items are not substantially the same, that is, different during the inference and data collection for training the machine-learning model, the terminal device may either start model monitoring or model retraining before continuing with inference by feeding data collected under network conditions of the second time (X') into the trained machine-learning model. According to at least some example embodiments, the terminal device receives a request for a report together with an information regarding the change of conditions during the training phase and the network conditions of the second time from the network. According to at least some example embodiments, the requested report is the terminal device's location. According to at least some example embodiments, the time information is expressed in terms of pre-defined time slots, such as MORNING (05:00-12:00), T2: AFTERNOON (12:00-17:00). According to at least some example embodiments, the time information is indicated in the format of UTC, GNSS or network time. Furthermore, it is to be understood that when it is stated that the processor (or some other means) is configured to perform some function, such function is to be construed to be equivalently implementable by specifically configured circuitry (e.g., the expression "processor configured to [cause the device to] perform xxx-ing" is construed to be equivalent to an expression such as "xxx-circuitry configured to perform xxx-ing"). According to at least some example embodiments as illustrated in Fig. 7(a), an apparatus representing a device, entity or terminal device like 100 e.g. a UE or an NWDAF comprises at least one processor 110, at least one memory 120. According to at least some example embodiments, the terminal device may further comprise at least one interface 130 configured for communication with at least another apparatus. The processor (i.e. the at least one processor 110, with the at least one memory 120 and the computer program code and with and with the at least one Interface 130) Is configured to receive, from a network entity, a result of comparing network conditions information (XQ associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; and decide, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction. Alternatively, according to at least some example embodiments as illustrated in Fig. 7(b), an apparatus representing a device, entity or terminal device 100' like e.g. a UE or an NWDAF comprises means for receiving e.g. from a network entity, means for transmitting e.g. to a network entity, processing means e.g. for training a machine-learning model and storing means for storing information like e.g. time information of collecting or obtaining data or the trained machine-learning model, AtT1 v* LU ■ According to at least some example embodiments as illustrated in Fig. 8(a), an apparatus representing the network entity 200 like e.g. an LMF or a gNB or alternatively an NWDAF comprises at least one processor 210, at least one memory 220. According to at least some example embodiments, the network entity may further comprise at least one interface 230 configured for communication with at least another apparatus. The processor (i.e. the at least one processor 210, with the at least one memory 220 and the computer program code and with and with the at least one interface 230) is configured to record network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn); retrieve recorded network conditions information associated with at least one timestamp (ti) indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; compare the retrieved network conditions information (Xi) with network conditions information (X') of a second time (tj) being different from the first time; and transmit, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time. Alternatively, according to at least some example embodiments as illustrated in Fig. 8(b), an apparatus representing the network entity 200' like e.g. an LMF or a gNB comprises means for receiving (e.g. a receiver) 250 e.g. from a device like e.g. a terminal device or UE, means for transmitting (e.g. a transmitter) 260 e.g. to a terminal device, processing means (e.g. a processor) 270 e.g. for measuring or comparing respective network conditions information and storing means (e.g. memory) 280 for storing information like e.g. recorded network conditions information, time information, etc. As used herein, the term "circuitry" may refer to one or more or all of the following: (a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry) and (b) combinations of hardware circuits and software, such as (as applicable): (i) a combination of analog and / or digital hardware circuit(s) with software / firmware and (ii) any portions of hardware processor(s) with software (comprising digital signal processor(s)), software, and memory(ies) that work together to cause a device, such as a mobile phone or server, to perform various functions) and (c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not utilized for operation. This definition of circuitry applies to all uses of this term herein, comprising in any claims. As a further example, as used herein, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in a server, a cellular network device, or other computing or network device. For the purpose of this disclosure as described herein above, it should be noted that: - method steps may be implemented as software code portions and run using at least one processor at a network server or network entity (as examples of devices, devices and / or modules thereof, or as examples of entities comprising devices and / or modules therefore), and may further be implemented as software code independent and can be specified using any known or future developed programming language as long as the functionality defined by the method steps is preserved; - generally, any method step may be implemented as software and / or by hardware without changing the example embodiments and its modification in terms of the functionality implemented; - method steps and / or devices, units or means likely to be implemented as hardware components at the above-defined devices, or any module(s) thereof, (e.g., devices carrying out the functions of the devices according to the example embodiments as described herein) are hardware independent and can be implemented using any known or future developed hardware technology or any hybrids of these, such as MOS (Metal Oxide Semiconductor), CMOS (Complementary MOS), BiMOS (Bipolar MOS), BiCMOS (Bipolar CMOS), ECL (Emitter Coupled Logic), TTL (Transistor-Transistor Logic), etc., using for example ASIC (Application Specific IC (Integrated Circuit)) components, FPGA (Field-programmable Gate Arrays) components, CPLD (Complex Programmable Logic Device) components or DSP (Digital Signal Processor) components; - a device like the database or a distributed node, local trainer, etc. may be implemented by a semiconductor chip, a chipset, or a (hardware) module comprising such chip or chipset; this, however, does not exclude the possibility that a functionality of a device or module, instead of being hardware implemented, be implemented as software in a (software) module such as a computer program or a computer program product comprising executable software code portions for execution / being run on a processor; - an entity may be regarded as a device or as an assembly of more than one device, whether functionally in cooperation with each other or functionally independently of each other but in a same device housing, for example. In general, it is to be noted that respective functional blocks or elements according to above-described aspects can be implemented by any known means, either in hardware and / or software, respectively, if it is only adapted to perform the described functions of the respective parts. The mentioned method steps can be realized in individual functional blocks or by individual devices, or one or more of the method steps can be realized in a single functional block or by a single device. Generally, any method step is suitable to be implemented as software or by hardware without changing this disclosure. Devices and means can be implemented as individual devices, but this does not exclude that they also can be implemented in a distributed fashion throughout the system, as long as the functionality of the device is preserved. Software in the sense of this description comprises software code as such comprising code means or portions ora computer program or a computer program product for performing the respective functions, as well as software (or a computer program or a computer program product) embodied on a tangible medium, such as a computer-readable (storage) medium having stored thereon a respective data structure or code means / portions or embodied in a signal or in a chip, potentially during processing thereof. This disclosure also covers any conceivable combination of method steps and operations described above, and any conceivable combination of nodes, devices, modules or elements described above, as long as the above-described concepts of methodology and structural arrangement are applicable. Even though the disclosure is describes various example embodiments with reference to the accompanying drawings, it is to be understood that the disclosure is not restricted thereto. Rather, it is apparent to those skilled in the relevant art(s) that this disclosure can be modified in various ways without departing from the scope of the various example embodiments disclosed herein. According to at least some example embodiments, there may be provided a device comprising at least one processor and at least one memory (e.g., non-transitory computer readable medium) storing instructions that, when executed by the at least one processor, cause the device to perform at least any method of this disclosure. The term "non-transitory," as used herein, is a limitation of the medium itself (e.g., tangible, not a signal) as opposed to a limitation on data storage persistency (e.g., RAM vs. ROM). According to at least some example embodiments, there may be provided a computer program comprising instructions which, when executed by a device, cause the device to perform at least any method of this disclosure. As used herein, "at least one of the following: " and "at least one of " and similar wording, where the list of two or more elements are joined by "and" or "or," mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements. As used herein, the expression "and / or" also includes any and all combinations of the listed terms, including at least any one of the elements, or at least any two or more of the elements, or at least all of the elements. As used herein, the term "or" refers to a non-exclusive "or" unless otherwise indicated (e.g., use of "or else" or "or in the alternative"). As used herein, unless stated explicitly, performing a respective feature, step, or functionality "in response to A" does not indicate that the respective feature, step, or functionality is performed immediately after "A" occurs as one or more unstated and intervening features, steps, or functionalities may be performed (at least in part) between an occurrence of the respective feature, step, or function and "A". Analogously, performing a respective feature, step, or functionality "based on A" does not indicate that the respective feature, step, or functionality is performed solely based on "A" as the respective feature, step, or functionality may be further based on one or more unstated features, steps, or functionalities in addition to "A". Throughout the present disclosure, the following abbreviations are used as summarized below. PARTIAL LIST OF ABBREVIATIONS AI Artificial Intelligence ARP Antenna Reference Point CN Core Network gNB next Generation Node B LCS Location Service LMF Location Management Function LOS Line-Of-Sight LPP LTE Positioning Procedure ML Machine-learning NLOS Non-Line-Of-Sight NW Network NWDAF NetWork Data Analytics Function PLMN Public Land Mobile Network PRS Positioning Reference Signal PRU Positioning Reference Unit QoS Quality of Service RAN Radio Access Network RAT Radio Access Technologies RS Reference Signal RSRP Reference Signal Received Power Rx Receive SINR Signal to Interference + Noise Ratio SNR Signal to Noise Ratio TDOA Time Difference of Arrival TRP Transmission Reception Point Tx Transmit UE User Equipment

Claims

1. An apparatus (200), comprising:at least one processor (210), andat least one memory (220) storing computer program codes that,when executed by the at least one processor, cause the apparatus at least to:record network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn);retrieve recorded network conditions information associated with at least one timestamp (ti) indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time;compare the retrieved network conditions information (Xi) with network conditions information (X') of a second time (tj) being different from the first time; andtransmit, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time.

2. The apparatus according to claim 1, wherein the transmitted result of comparing comprises at least one of the following:- information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X7) of the second time are substantially the same,- information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or- an indication whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction.

3. The apparatus according to claim 1 or 2, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to receive time information comprising the at least one timestamp (ti) indicative of the first time from the terminal device.

4. The apparatus according to claim 1 or 2, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to:transmit network configuration information for obtaining the data, with which the predictive model is trained, to the terminal device; andstore at least one timestamp (ti) at the time of transmitting the network configuration information as the at least one timestamp (ti) indicative of the first time.

5. The apparatus according to claim 3, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to:perform retrieving the recorded network conditions information, comparing, and transmitting the result of comparing to the terminal device in response to receiving the time information.

6. The apparatus according to claim 1 to 5, wherein the network conditions information (Xo, ..., Xn) is recorded in association with the at least one timestamp (ti) indicative of the first time.

7. The apparatus according to claim 6, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to:compare the retrieved network conditions information (Xi) with network conditions information (X') of the second time,and in response to determining that the retrieved network conditions information (X,) and the network conditions information (X') of the second time are different,transmit, to the terminal device, the result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time.

8. The apparatus according to any one of claims 1 to 7, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to:receive, from a third-party device, the data, with which the predictive model is trained, andtransmit the received data to the terminal device.

9. The apparatus according to any one of claims 1 to 8, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to:receive information about the trained predictive model, andstore the received information about the trained predictive model.

10. The apparatus according to any one of claims 1 to 9, wherein the at least one timestamp (ti) indicates a predefined time slot.

11. An apparatus (100), comprising:at least one processor (110), andat least one memory (120) storing computer program codes thatwhen executed by the at least one processor cause the apparatus at least to:receive, from a network entity, a result of comparing network conditions information (X,) associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with data obtained under the network conditions of the first time; anddecide, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction.

12. The apparatus according to claim 11, wherein the received result of comparing comprises at least one of the following:- information indicating whether the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same,- information indicative of differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time, - information indicative of an amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (X') of the second time, or- an indication whether to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction.

13. The apparatus according to claim 11 or 12, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to perform at least one of model monitoring or model retraining before feeding the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, in response to determining that at least one of the following conditions is met:- the retrieved network conditions information (Xi) and the network conditions information (Xz) of the second time are different,- the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time indicate a difference not included in a predefined set of differences, - the amount of differences between parameters of the retrieved network conditions information (Xi) and parameters of the network conditions information (Xz) of the second time is more than a predetermined threshold, or - the received result of comparing includes an indication not to feed the data obtained at the second time into the predictive model trained with the dataobtained under the network conditions of the first time for inference of a prediction.

14. The apparatus according to any one of claims 11 to 13, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to decide to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction, in response to determining that at least one of the following conditions is met:- the retrieved network conditions information (Xi) and the network conditions information (X') of the second time are substantially the same, or- the information indicative of the differences between the retrieved network conditions information (Xi) and the network conditions information (X') of the second time do not indicate a difference not included in a predefined set of differences, or- the amount of differences between parameters of the retrieved network conditions information (Xi) and the parameters of the network conditions information (X') of the second time is less than a predetermined threshold, or- the received result of comparing includes an indication to feed the data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction.

15. The apparatus according to any one of claims 11 to 14, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to transmit, to the network entity, time information comprising the at least one timestamp (ts) indicative of the first time.

16. The apparatus according to claim 15, wherein transmitting the time information to the network entity is performed in response to training the predictive model.

17. The apparatus according to any one of claims 11 to 16, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to receive the data used for training the predictive model from the network entity.

18. The apparatus according to any one of claims 11 to 17, wherein the at least one memory and the computer program codes, when executed by the at least one processor further cause the apparatus to transmit, in addition to the time information, information about the predictive model to the network entity.

19. The apparatus according to any one of claims 11 to 18, wherein the at least one timestamp (t) indicates a predefined time slot.

20. The apparatus according to any one of claims 11 to 19, wherein the data, with which predictive model is trained, obtained at the first time comprises at least one of:- a downlink reference signal configuration,- a measurement regarding a transmitted downlink reference signal,- time information indicating a time of performing the measurement regarding the transmitted downlink reference signal,- a quality indicator indicating a quality of the measurement, or- ground truth labels.

21. A method, comprising:recording network conditions information (Xo, ..., Xn) in association with a respective timestamp (to, ..., tn);retrieving recorded network conditions information associated with at least one timestamp (ti) Indicative of a first time, wherein a predictive model was trained with data obtained under the network conditions of the first time;comparing the retrieved network conditions information (Xi) with network conditions information (X') of a second time, the second time being different from the first time;transmitting, to a terminal device, a result of comparing the retrieved network conditions information (Xi) with the network conditions information (X') of the second time.

22. A method, comprising:receiving, from a network entity, a result of comparing network conditions information (Xi) associated with at least one timestamp (ti) indicative of a first time and network conditions information (X') of a second time different from the first time, wherein a predictive model was trained with the data obtained under the network conditions of the first time; anddeciding, based on the received result of comparing, whether to feed data obtained at the second time into the predictive model trained with the data obtained under the network conditions of the first time for inference of a prediction.

23. A computer program product comprising computer-executable computer program code, which when the program is run on a computer, is configured to cause the computer to perform the method according to claim 21 or 22.

24. A non-transitory computer-readable medium comprising program instructions that, when executed by an apparatus, cause the apparatus to perform the method according to any one of claims 21 or 22.

Images