A Learning Approach for User Localization and Movement Prediction

with Limited Information

Quang-Vinh Tran

and Quang-Diep Pham

Switching Technology Research Center, Viettel Hi-Technology Industries Corporation, Hanoi, Vietnam

Keywords:

5G Core Network, User Localization, Handover, Movement Prediction.

Abstract:

In the 5G network system, users continuously travel among areas managed by different User Plane Functions

(UPFs), leading to the need for efﬁcient handover between UPFs. Conventional handover relies on signal

measurements between user devices and neighboring base stations, so it is a ”re-active” scheme. Therefore,

this procedure results in long response time of the Packet Data Unit (PDU) session establishment, and affecting

data service quality. Another approach is an ”pro-active” scheme, in which the position of users are estimated,

hence the decision of UPF handover can be made earlier. We propose a solution using machine learning

techniques to model user movement behavior in the network and predict user positions in advance. The

predicted UPF managing the next location will be announced accordingly to take preparatory steps for serving

the incoming users, thereby reducing the new PDU session establishment latency, increasing processing speed,

and improving the quality of experience. We propose the model combining the K-means clustering algorithm

and the Gated Recurrent Unit deep learning network for time series data. The solution was tested with Viettel’s

5G network data and demonstrated its feasibility in real-world dataset.

1 INTRODUCTION

Viettel Group is currently the largest telecommuni-

cations group in Vietnam with tens of million cus-

tomers. With the development trend of network gen-

erations, Viettel’s 5G network is being developed and

perfected. In order to meet the low latency and high

speed requirements according to 3GPP standards, we

constantly pose problems. The 5G core network

system is being built with a complex model includ-

ing many services performing different functions, in

which the User Plane Function (UPF) is responsible

for packet routing and forwarding, packet inspection,

QoS handling, and external PDU session for intercon-

necting data network (DN), in the 5G architecture.

Currently, UPF performs connection establish-

ment operations with the User Equipment (UE) in a

reactive manner. When the UE enters a cell managed

with a certain UPF instance, this UPF instance will

initiate several connection procedures, for example,

establishing tunnels for data communication or trans-

ferring connections, to allow the UE seamlessly us-

ing the service when standing at a new location. This

https://orcid.org/0009-0008-1092-0878

https://orcid.org/0009-0002-4854-6929

causes process latency to increase, reducing the qual-

ity of service experience. As the coverage area of

a base station (gNB) becomes smaller to accommo-

date different services, especially internet of things

(IoT), ultra reliable and low latency communications

(uRLLC) and streaming services, this requirement be-

comes more important. Therefore, we propose a prob-

lem to predict which UEs will be in the area under

which UPFS’ management. By statistically learning

the user’s movement behaviors, we can predict the

following cell location that the UE would move in.

Our approach is to classify and learn subscribers’

movement behavior or patterns and then learning the

movement patterns within each cluster. The cho-

sen clustering algorithm is K-means, where the ”be-

havior” of users is measured by their deviation over

the same time period, and then users are distributed

into different clusters. Gated recurrent unit (GRU)

is employed to learn user behavior within each clus-

ter. Unlike traditional Recurrent Neural Network

(RNN), which suffer from issues like vanishing gradi-

ents during long sequence learning, GRUs introduce

gating mechanisms that regulate the ﬂow of informa-

tion and mitigate these problems. This makes GRUs

highly effective in capturing temporal dependencies

over longer periods.

Tran, Q.-V. and Pham, Q.-D.

A Learning Approach for User Localization and Movement Prediction with Limited Information.

DOI: 10.5220/0013221600003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 717-723

ISBN: 978-989-758-749-8; ISSN: 2184-4992

717

Our contributions can be summarized as follows:

• A preprocessing solution for log ﬁles by convert-

ing them into a dataset,

• A model that combines machine learning and

deep learning to solve the position prediction

problem.

The rest of the paper is organized as follows.

Sec. 2 summarizes some related work. Sec. 3 intro-

duces our proposed approach to the problem of user

localization and movement prediction, while Sec. 4

presents the performance evaluation of our proposed

solution. Finally, some conclusions and perspectives

are drawn in Sec. 5.

2 RELATED WORK

The UPFs play a crucial role as a central user plane

element, responsible for various essential functions in

managing user data trafﬁc, consist of packet routing,

quality-of-service enforcement, and trafﬁc monitor-

ing. The PDU session establishment initiated by UEs’

requests will include the participation of AMF (Ac-

cess and Mobility NF), SMF (Session Management

NF). In such procedure, the SMF selects the most ap-

propriate User Plane Function (UPFs) for the session

considering the layout of the network, the available

resources of UPFs, and quality policies for UE. SMF

might even ask the Policy Control Function (PCF)

for additional session-related guidelines. Afterwards,

the SMF sends detailed instructions and policies to

the UPFs including quality parameters, session identi-

ﬁers, and conﬁrmation of the chosen UPFs. The UPFs

then conﬁgure the data path based on these instruc-

tions.

To ensure the continuity of various PDU sessions

while an UE is moving, the procedure of UEs’ han-

dover will include AMF, SMF and UPF as well. Ses-

sion and Service Continuity (SSC) enables to ad-

dress the different continuity requirements of differ-

ent applications and services for UEs. Some appli-

cations like the IP Multimedia Subsystem (IMS) re-

quire an always-on Protocol Data Unit (PDU) session

that the User Plane resource establishes for every tran-

sition from the 5GMM-IDLE mode to the 5GMM-

CONNECTED mode. The UE requests the establish-

ment of a PDU session as an always-on PDU session

based on the request indication of the upper layers. It

is the network that decides whether to establish a PDU

session as an always-on PDU session. Therefore, an

”proactive” scheme based on the UE position predic-

tion allowing the planning of always-on PDU session

and other resources (scaling up or down UPFs) could

be an efﬁcient way to save the network resource while

the service quality commitment.

Research on user localization and movement pre-

diction has evolved signiﬁcantly, with various ap-

proaches ranging from traditional probability-based

models to advanced machine learning techniques.

These approaches can be classiﬁed into three main

categories: traditional probabilistic models, machine

learning-based approaches, and hybrid methods.

In study (Arifﬁn et al., 2013), the Markov model

parameter that inﬂuences the prediction is the tran-

sition probability matrix. Inaccuracy in the value

of the transition probability matrix will lead to in-

correct predictions and may cause failure. Further-

more, in this paper, the value of the transition prob-

ability matrix is assumed and has not yet been de-

termined by human behavior. Author in (Yan et al.,

2021) presents a weighted Markov prediction model

that incorporates mobile user classiﬁcation. It ﬁrst

extracts trajectory data from real mobile communica-

tion records, with entropy used to measure the com-

plexity of users’ movements. Users are then classi-

ﬁed based on their behavior patterns using machine

learning. The Markov model’s step thresholds and

weighting factors are adjusted for each user group to

improve mobility prediction. However, while the per-

formance improves, the beneﬁts are less pronounced

for users with high or irregular mobility. Zhang et al.

(Zhang et al., 2009) proposed a Bayesian Network-

based location prediction model using multiple fac-

tors like topology, road typology, and movement data

to improve prediction accuracy. Predictive factors are

encoded in the network nodes, and location predic-

tion is calculated using cell topology information. A

factor distribution mechanism addresses cases where

nodes lack direct prediction data.

Although the results are promising, the study re-

mains at the simulation. Similar to (Zhang et al.,

2009), Liu et al. (Liu et al., 2010) proposed a

novel approach based on Bayesian network to predict

a moving object’s future location under uncertainty.

This paper proposed several algorithms to construct

a Bayesian network from trajectory information and

suggested a method to predict a moving object’s fu-

ture location under uncertainty using this network.

The Bayesian network allows inference and calcu-

lates the probabilities of all possible states of an un-

observed node based on current data. However, its

drawback lies in modeling nonlinear relationships, as

accurately capturing such dynamics or changes in tra-

jectories may require additional assumptions, increas-

ing the model’s complexity.

In recent years, machine learning-based methods

have grown increasingly powerful. Recurrent Neu-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

718

ral Networks (RNNs), especially Long Short-Term

Memory (LSTM) and Gated Recurrent Units (GRUs),

have been effective in predicting user mobility due to

their ability to capture temporal dependencies. How-

ever, RNNs face the vanishing gradient problem when

input data sequences become long (Hochreiter et al.,

2001). Meanwhile, LSTM (Schmidhuber et al., 1997)

and GRU (Cho et al., 2020) mitigate this limitation

by introducing a cell state, with GRU offering greater

computational efﬁciency compared to LSTM. An-

other study uses RNNs to build users’ personalized

preferences and model their historical behavior. It

then incorporates an attention mechanism to dynam-

ically weight historical user behaviors based on the

user’s current message transmission (Gou and Wu,

2022). Z. Xiong et al. (Xiong et al., 2019) proposed a

Deep Reinforcement Learning (DRL) method to op-

timize paging strategies in real time by continuously

adapting to changing user movement patterns. This

dynamic approach improved performance in dense

5G environments with frequent mobility.

While standalone machine learning models such

as RNNs, LSTM, and GRU have shown potential in

user mobility prediction, they often encounter difﬁ-

culties in managing large-scale data and maintaining

real-time efﬁciency, particularly in dense 5G envi-

ronments. To address these challenges, hybrid ap-

proaches that combine clustering techniques with ma-

chine learning models have emerged as a promis-

ing solution. These methods ﬁrst cluster users based

on similar mobility patterns, enabling the prediction

models to focus on speciﬁc user groups. This ap-

proach can signiﬁcantly reduce paging overhead by

narrowing the search space for user localization, thus

improving prediction accuracy and minimizing com-

putational complexity. The authors in (Kim et al.,

2022) proposed clustering to group UE based on mo-

bility patterns, time, and day of the week, then apply

the GRU model to predict UE locations. The results

demonstrated that the GRU model achieved a paging

accuracy of over 80%. However, this method heav-

ily depends on the clustering process and appropriate

parameter conﬁguration. Choosing the suitable clus-

tering method depends on the nature of user mobility

data, with K-means being appropriate for uniformly

distributed data.

To overcome these challenges, our work focus is

on improving the clustering process and developing

more efﬁcient data handling techniques. By doing so,

the system can better reﬂect the constantly changing

user behaviors in dense 5G environments, thus im-

proving prediction accuracy.

3 PROPOSED APPROACH

Due to occupation, gender, etc., each user’s habits and

movements are different. Training the model on a

dataset of all users will result in the model not be-

ing able to converge. It is necessary to specialize the

model for each user so that the model is suitable for

each subscriber’s movement behavior.

We propose a solution consisting of three pro-

cesses: (i) preprocessing, (ii) UE clustering, and (iii)

movement behavior learning. This clustering is based

on subscriber movement, not on individual informa-

tion for clustering, because this is all conﬁdential

information. Our proposed behavioral learning and

decision-making model uses deep learning techniques

to train on time series data. The overall architecture

of our solution is illustrated in Fig. 1, where t is se-

quence length, {X

,...,X

} is input of GRU net-

work, n is number of gNB stations, [P

,...,P

] is a

probability vector representing the UE’s ability to be

at the stations in the (t + 1) timestamp.

3.1 Pre-Processing of Log Files

In this section, we will present the process of han-

dling the initial raw data (log ﬁles) for use in subse-

quent phases. Log ﬁles are recorded on a per-minute

basis, facilitating ease of reading and normalization.

However, they may miss certain values and include

stations in areas with low user density. Therefore,

an effective processing procedure is required to avoid

noise and data imbalance. The log ﬁles contain 52

data columns, of which we utilized the following:

• EVENT TIME: The time when the event oc-

curred,

• EVENT ID: The identity of the event,

• EVENT RESULT: The outcome of the event pro-

cedure,

• IMSI: International Mobile Subscriber Identity.

This information element contains data com-

monly used to identify the UE in the Core Net-

work,

• ECI: E-UTRAN Cell Identity. This is used to

identify cells within a Public Land Mobile Net-

work (PLMN).

We employ the following processes:

Data Cleaning. EVENT RESULT (The outcome of

the event procedure) is used to ﬁlter out failed events.

For example, when a ”l handover” event fails, the

ECI is not updated to reﬂect the new cell, result-

ing in redundant information. In this context, we

speciﬁcally focus on the EVENT ID values of ”l tau”,

A Learning Approach for User Localization and Movement Prediction with Limited Information

719

Behavioral Learning

GRU

softmax

GRU

Dataset

Log

ﬁles

Cluster ID

Dataset

Pre-

Processing

K-Means

Clustering

Figure 1: Overall architecture.

”l service request”, and ”l handover”, as these events

are directly related to the UE updating its location

during mobility. The cell information is converted to

gNB station information using the following formula,

gNB ID =



ECI

256



. (1)

Data Filtering. Events related to UEs exhibiting ab-

normal behavior will also be ﬁltered out to minimize

noise. We establish the following exclusion criteria:

UEs that exhibit a ping-pong phenomenon, charac-

terized by continuously switching their location be-

tween multiple base stations; UEs that transition be-

tween more than three base stations within one minute

(in which case, the three base stations may indicate

that the UE is traversing a region where the cover-

age overlaps of three different stations); or UEs that

move to fewer than three base stations within a 60-

minute period (which provides insufﬁcient informa-

tion for the machine learning process). Additionally,

gNB stations with a low number of users, particularly

in mountainous regions, will also be excluded. Fur-

thermore, UE that has a history of movement through

these stations will be discarded as well.

Data Imputation. The generated dataset will have

a time interval of 1 minute, and with log ﬁles cover-

ing a duration of 120 minutes, the dataset will contain

120 timestamps. Therefore, the following data ﬁlling

rules are necessary: if, within a 1-minute period, a

UE has movement data across a base station, prior-

ity will be given to retaining the information from the

most recent base station. In cases where no events oc-

cur during a minute (resulting in the absence of base

station information), this will be ﬁlled with data from

the subsequent base station. For the ﬁnal minutes of

the dataset, which lack information, the value will be

populated with the last updated base station value of

the UE.

After processing, we obtained movement

data for 6, 472 UEs over a period of 120 min-

utes, traversing 57 gNB stations. An exam-

ple of the dataset is illustrated in Table 1,

wherein [45204·······20,45204·······34,...]

represents the list of IMSIs, and

[900083,900389,900326,909520,...] represents

the list of gNB stations.

3.2 Cluster Selection

In our problem, we focus on clustering user behav-

ior to categorize movement trajectories into distinct

clusters. Each cluster shares similar characteristics,

enabling the deep learning model to converge more

rapidly and mitigating the issue of underﬁtting. A key

challenge, however, is the lack of additional user in-

formation beyond the IMSI identiﬁer, which prevents

us from applying rule-based grouping or actively la-

beling users. Consequently, we opted for a cluster-

ing algorithm to automatically group users based on

emerging data patterns (Faizan et al., 2020; Yin et al.,

2024).

We choose K-means model (MacQueen, 1967) for

clustering process. K-means offers the advantages

of simplicity, high speed, it is capable of supporting

complex problem solving and multidimensional data

(Rahamathunnisa et al., 2020; Kakbra, 2014; Subbiah

and Christopher, 2012), which is why we expect it to

perform well on our dataset. The primary goal of K-

means is to partition a set of data points into K clus-

ters, each deﬁned by its centroid. The algorithm oper-

ates by optimizing the positions of the centroids and

assigning data points to the nearest cluster. Euclidean

Distance used to calculate the distance between a data

point and a centroid:

d(x

) =

∑

k=1

− c

)

, (2)

where d(x

) is the distance between data point x

and centroid c

; x

, c

is he value of the k-th feature

of x

and c

After assigning data points to clusters, the cen-

troid of each cluster is updated by calculating the av-

erage of all data points in that cluster:

∑

∈C

, (3)

where C

is the set of data points belonging to cluster

j and |C

| is number of points in cluster j. The objec-

tive function to be optimized is the total squared dis-

tance between the data points and their corresponding

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

720

Table 1: Dataset from log ﬁles.

Time

IMSI

45204··· ··· ·20 45204··· ··· ·34 . . .

20240626.1400 900083 900389 . . .

20240626.1401 900326 909520 . . .

20240626.1402 900326 900387 . . .

. . . . . . . . . . . .

centroids:

J =

∑

j=1

∑

∈C

d(x

)

. (4)

With the data on the sequence of base stations along

the UE’s trajectory, we encode and concat them into

input vectors for K-means. Example for the i-th UE

over a period of T minutes is x

= x

⊕x

⊕··· ⊕ x

We propose two solutions for this encoding:

• One-hot encoding of gNB station codes, followed

by concatenating them into a single vector for

each UE. Example for the i-th UE in t-th minutes:

= [0,0,...,1,...,0],t ∈ [1,T ]. (5)

• Utilizing the geographical coordinates of the gNB

stations and the geographic displacement between

consecutive timestamps, then concatenating them

into a single vector for each UE.

= [LAT

,LON

], (6)

= [LAT

,LON

,d(x

i(t−1)

)],∀t ∈ [2, T ],

(7)

where d(x

i(t−1)

) is the distance from x

i(t−1)

and can be calculated using the Haversine for-

mula (Robusto, 1957) or the Euclidean distance

(since the distance between two stations is small

compared to the Earth’s radius).

The set of vectors as inputs to classify into 20 clus-

ters (i.e., 20 UEs movement patterns). If a UE does

not have sufﬁcient data for clustering (e.g., a newly

registered subscriber), it will be assigned to the clus-

ter with the largest number of UEs.

3.3 Behavioral Learning

The mobility data of subscribers exhibits a strong

temporal correlation and presents complexities in be-

havior modeling. Therefore, we opted for deep learn-

ing networks to address this challenge. GRU net-

works (Gated Recurrent Units) are a suitable choice

for capturing the intricate mobility patterns of users

as they move across different gNB stations.

A Gated Recurrent Unit (GRU) network (Cho

et al., 2014) is a specialized type of recurrent neu-

ral network (RNN) designed to efﬁciently model se-

quential data, making it particularly suitable for time

series forecasting and prediction tasks. GRUs are

highly effective in capturing temporal dependencies

over longer periods.

The GRU has two primary gates: the update gate

and the reset gate. These gates control how much of

the past information should be carried forward and

how much new information should be incorporated at

each time step.

1. Update Gate. The update gate z

decides the extent

to which the hidden state from the previous time step

t−1

will be carried forward. It is computed as:

= σ(W

· [h

t−1

]), (8)

where W

is the weight matrix, x

is the input at time

t, and σ is the sigmoid activation function.

2. Reset Gate. The reset gate r

determines how

much of the previous hidden state should be forgotten,

allowing the model to discard irrelevant information.

It is given by:

= σ(W

· [h

t−1

]), (9)

where W

is the weight matrix.

3. Candidate Hidden State. Using the reset gate, the

GRU computes the candidate hidden state

, which

incorporates the new input and some portion of the

previous hidden state. It is calculated as:

= tanh(W · [r

⊙ h

t−1

]), (10)

where ⊙ represents element-wise multiplication, and

tanh() is the hyperbolic tangent activation function.

4. Final Hidden State. Finally, the new hidden state

is a combination of the previous hidden state and the

candidate hidden state, weighted by the update gate:

= (1 − z

) ⊙ h

t−1

+ z

⊙

(11)

The GRU’s ability to adaptively control the informa-

tion ﬂow makes it particularly suited for time series

data, where the model must effectively balance be-

tween retaining past observations and learning new

patterns. By reducing the number of gates compared

to the Long Short-Term Memory (LSTM) network,

GRUs are computationally more efﬁcient while main-

taining robust performance on a wide range of time-

dependent tasks.

In applying this approach to our problem, each

subscriber (represented by an IMSI) is associated

A Learning Approach for User Localization and Movement Prediction with Limited Information

721

with a trajectory sequence over a speciﬁc time period

(i.e., the sequence of gNB stations that the subscriber

moves through). We encoded this gNB sequence into

one-hot vectors, combined with the encoding of clus-

ter Id, to form a complete vector set as input to the

behavior learning model:

X(Model

input

) = X

,...,X

, (12)

where X

= [encode o f gNB⊕encode o f cluster ID],

i ∈ [1,t]; t is sequence length. The output of the model

is ﬁtted with the one-hot encoding of the gNB sta-

tion for the t + 1 timestamp (Y = encode o f gNB =

[0,0,...,1,...,0]).

Outputs of the model are a probability vectors, in-

dicating the likelihood of the subscriber being at a

speciﬁc gNB station at the predicted time. Ideally, the

probability of a subscriber moving to the correct sta-

tion is 1, while the probabilities for all other stations

are 0.

4 PERFORMANCE EVALUATION

The test data was collected on June 26, 2024, from

14:00 to 15:59 GMT+7. After ﬁltering, we obtained

data from 6, 472 subscribers moving through 57 gNB

stations.

Following the Clustering and Behavioral learning

process, we set the parameter values as follows:

• Number of clusters: k = 20

• Number of GRU network layers: 3

• Learning rate: 0.001

• Number of training epoch: 13

• Training/testing ratio: 70/30

• Sequence length: 16

Our model is designed to suggest the most likely

options for the next station’s location, so it is difﬁcult

to use the top-1 or top-k parameter for evaluation. In-

stead, we use the average predicted probability across

all samples as accurary of model. The prediction ac-

curacy for timestamp t of i-th UE is:

|−1

∑

j=0

[ j],

= Y

⊙

in this context, Y is one-hot encode of gNB and

Y is a

probability vector representing the prediction. Then,

the model’s accuracy (score testing) for all predictions

deﬁned as follows:

P([UEs]) =

∑

j=t+1

i j

30% × N × (T −t)

, (13)

where i represents 30% of UEs for testing phase, N is

number of UEs, j is testing timestamp, t is sequence

length and T is length of data collected (in minutes).

The loss values on the training set (green), testing

set (blue), and testing score (red) (P([UEs])) during

training phase visualized as Figure 2 (K-means GRU),

the loss values decrease over time, and the score test

parameter saturates at the 10-th epoch (87.11%).

We compare the performance of the GRU network

with LSTM. In Table 2, the Accuracy value is calcu-

lated as formula (13), Execute time is the inference

time of the model for each UE at each time point in the

testing phase, calculated by dividing the total execu-

tion time by the sample that performed the inference.

The K-means model combined with a 3-layer GRU

network is more accurate and executes faster than the

LSTM network (about 0.2 ms/sample). We changing

the number of GRU network layers, and observing the

change in accuracy: as the number of GRU network

layers increased, accuracy also increased, and when

this value reached 3, the accuracy stabilized, while

the execution time remained acceptable.

Figure 2: Performance results of KMeans-GRU model.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

722

Table 2: Accurary of methods (* is our method), KM is K-means.

Method KM-LSTM

KM-GRU

1 layers 2 layers 3 layers(*) 4 layers

Accuracy (%) 82.45 77.12 86.08 87.11 87.23

Execute time

1.520 1.346

(ms/sample)

5 CONCLUSION

In this paper, we present a novel approach for predict-

ing user movement behavior to facilitate the seamless

handover of User Plane Function (UPF) in network

environments. Initial results validate the effective-

ness of the proposed solution in addressing the key re-

quirements of the problem. However, signiﬁcant chal-

lenges persist, particularly regarding the deployment

of the solution across the large-scale data generated

by the network.

As the number of base stations increases, the data

quality may degrade, especially in scenarios involv-

ing User Equipment (UE) with low mobility or base

stations with a limited number of active subscribers.

These factors contribute to increased variability and

complexity in the data, potentially impacting the ac-

curacy and efﬁciency of the solution. Therefore, fur-

ther experimentation and reﬁnements are essential to

optimize the approach and ensure its robustness under

diverse real-world conditions.

REFERENCES

Arifﬁn, S. H., Abd, N., Ghazali, N. E., et al. (2013). Mobil-

ity prediction via markov model in lte femtocell. Inter-

national Journal of Computer Applications, 65(18).

Cho, K., van Merri

enboer, B., Gulcehre, C., Bahdanau, D.,

Bougares, F., Schwenk, H., and Bengio, Y. (2014).

Learning phrase representations using RNN encoder–

decoder for statistical machine translation. In Pro-

ceedings of the 2014 Conference on Empirical Meth-

ods in Natural Language Processing (EMNLP), pages

1724–1734, Doha, Qatar. Association for Computa-

tional Linguistics.

Cho, K., Van Merrienboer, B., Gulcehre, C., Bahdanau, D.,

Bougares, F., Schwenk, H., and Bengio, Y. (2020).

Learning phrase representations using rnn encoder-

decoder for statistical machine translation. arxiv 2014.

arXiv preprint arXiv:1406.1078.

Faizan, M., Zuhairi, M. F., Ismail, S., and Sultan, S. (2020).

Applications of clustering techniques in data mining:

A comparative study. International Journal of Ad-

vanced Computer Science and Applications, 11(12).

Gou, F. and Wu, J. (2022). Message transmission strategy

based on recurrent neural network and attention mech-

anism in iot system. Journal of Circuits, Systems and

Computers, 31(07):2250126.

Hochreiter, S., Bengio, Y., Frasconi, P., Schmidhuber, J.,

et al. (2001). Gradient ﬂow in recurrent nets: the dif-

ﬁculty of learning long-term dependencies.

Kakbra, J. (2014). Image classiﬁcation using k-mean al-

gorithm. International Journal of Emerging Trends &

Technology in Computer Science (IJETTCS), 3:38–41.

Kim, B., Raza, S. M., Yang, H., Challa, R., Le, D.-T.,

Choi, H., Lee, D., Kim, M., and Choo, H. (2022).

ipaas: Intelligent paging as a service. IEEE Network,

37(2):238–245.

Liu, C.-L., Jou, E., and Lee, C.-H. (2010). Analysis and

prediction of trajectories using bayesian network. In

2010 Sixth International Conference on Natural Com-

putation, volume 7, pages 3808–3812. IEEE.

MacQueen, J. (1967). Some methods for classiﬁcation and

analysis of multivariate observations. In Le Cam,

L. M. and Neyman, J., editors, Proceedings of the ﬁfth

berkeley symposium on mathematical statistics and

probability, pages 281–297. University of California

Press.

Rahamathunnisa, U., Nallakaruppan, M., Anith, A., and

Kumar K.S., S. (2020). Vegetable disease detec-

tion using k-means clustering and svm. In 2020

6th International Conference on Advanced Computing

and Communication Systems (ICACCS), pages 1308–

1311.

Robusto, C. C. (1957). The cosine-haversine formula. The

American Mathematical Monthly, 64(1):38–40.

Schmidhuber, J., Hochreiter, S., et al. (1997). Long short-

term memory. Neural Comput, 9(8):1735–1780.

Subbiah, B. and Christopher, S. C. (2012). Image classi-

ﬁcation through integrated k-means algorithm. Inter-

national Journal of Computer Science Issues (IJCSI),

9(2):518.

Xiong, Z., Zhang, Y., Niyato, D., Deng, R., Wang, P., and

Wang, L.-C. (2019). Deep reinforcement learning for

mobile 5g and beyond: Fundamentals, applications,

and challenges. IEEE Vehicular Technology Maga-

zine, 14(2):44–52.

Yan, M., Li, S., Chan, C. A., Shen, Y., and Yu, Y. (2021).

Mobility prediction using a weighted markov model

based on mobile user classiﬁcation. Sensors, 21(5).

Yin, H., Aryani, A., Petrie, S., Nambissan, A., Astudillo,

A., and Cao, S. (2024). A rapid review of clustering

algorithms. ArXiv, abs/2401.07389.

Zhang, Y., Hu, J., Dong, J., Yuan, Y., Zhou, J., and Shi, J.

(2009). Location prediction model based on bayesian

network theory. In GLOBECOM 2009-2009 IEEE

Global Telecommunications Conference, pages 1–6.

IEEE.

A Learning Approach for User Localization and Movement Prediction with Limited Information

723