University of Texas Rio Grande Valley

ScholarWorks @ UTRGV
School of Mathematical and Statistical
Sciences Faculty Publications and College of Sciences
Presentations

2023

Machine Learning for Wireless Network Throughput Prediction

Gustavo A. Fernandez

Follow this and additional works at: https://scholarworks.utrgv.edu/mss_fac

Part of the Mathematics Commons

Machine Learning for Wireless Network Throughput
Prediction
Gustavo A Fernandez (  gustavo.fernandez02@utrgv.edu )
The University of Texas Rio Grande Valley

Research Article

Keywords: Wireless Network, Machine Learning, Regression, Random Forest

Posted Date: August 16th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-3267046/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations: No competing interests reported.

 Machine Learning for Wireless Network
Throughput Prediction

Gustavo A Fernandez
The University of Texas Rio Grande Valley.

Abstract
This paper analyzes a dataset containing radio frequency (RF) measurements and Key Per-
formance Indicators (KPIs) captured at 1876.6MHz with a bandwidth of 10MHz from an
operational 4G LTE network in Nigeria. The dataset includes metrics such as RSRP (Refer-
ence Signal Received Power), which measures the power level of reference signals; RSRQ
(Reference Signal Received Quality), an indicator of signal quality that provides insight
into the number of users sharing the same resources; RSSI (Received Signal Strength Indi-
cator), which gauges the total received power in a bandwidth; SINR (Signal to Interference
plus Noise Ratio), a measure of signal quality considering both interference and noise; and
other KPIs, all derived from three evolved node base stations (eNodeBs). After meticulous
data cleaning, a subset of measurements from one serving eNB, spanning a 20-minute
duration, was selected for deeper analysis. The PDCP DL Throughput, as a vital KPI met-
ric, plays a paramount role in evaluating network quality and resource allocation strategies.
Leveraging the high granularity of the data, the primary aim was to predict throughput.
For this purpose, I compared the predictive capabilities of two machine learning mod-
els: Linear Regression and Random Forest. Metrics such as Mean Absolute Error (MAE),
Root Mean Squared Error (RMSE) were used to examine the models as they offer a com-
prehensive insight into the models accuracies. The comparative analysis highlighted the
superior performance of the Random Forest model in predicting the PDCP DL Through-
put. The insights derived from this research can potentially guide network engineers and
data scientists in optimizing network performance, ensuring a seamless user experience.
Furthermore, as the telecommunication industry advances towards the integration of 5G
and beyond, the methodologies explored in this paper will be invaluable in addressing the
increasingly complex challenges of future wireless networks.

Keywords: Wireless Network, Machine Learning, Regression, Random Forest

1
1 Introduction
In today’s digital age, telecommunications stands as a cornerstone of global connectivity.
As the world becomes increasingly interconnected, cellular network operators grapple with
the relentless challenge of accommodating escalating user demands. The explosion in media
consumption, especially with the introduction of bandwidth-intensive applications, real-time
media streaming on social platforms, and the rapidly evolving realm of connected and
autonomous vehicles, has placed unprecedented pressure on network resources. To address
these challenges, operators are in a continuous quest for cutting-edge solutions. One of the
primary objectives is to refine resource allocation and load balancing mechanisms, ensuring
that networks can handle the ever-growing data traffic without compromising on performance.
The anticipatory approach to resource allocation and network management is a groundbreak-
ing paradigm that offers a potential solution to these challenges. At the heart of this approach
lies the ability to predict network connectivity fluctuations before they occur. By proactively
identifying potential changes in connectivity, operators can take preemptive actions, ensuring
that the user’s Quality of Service (QoS) remains consistent and reliable. A prime example
of this forward-thinking strategy is the concept of pre-buffering video content. By allocat-
ing additional resources in anticipation of a potential drop in future throughput values for a
user, operators can guarantee uninterrupted streaming experiences. The need for such proac-
tive measures stems from the paramount importance of delivering consistent and high-quality
network connectivity. The proliferation of bandwidth-demanding applications and the expo-
nential growth in media publishing and streaming on social platforms highlight the critical
need for innovative network management strategies.
Certain studies have made significant contributions to the field among the plethora of
research dedicated to enhancing network QoS. For instance, Yue et al. [1] embarked on a com-
prehensive correlation analysis, exploring the intricate relationships between Radio Signals
(RSs) and throughput across various scenarios, from stationary settings to dynamic highway
driving conditions. Their findings emphasized the potential of the Random Forest machine
learning model in predicting network performance based on metrics like RSRP, RSRQ, and
CQI. In a similar vein, Raca et al. [2] delved into the realm of predicting future throughput
windows, evaluating the predictive prowess of diverse machine learning models, from Ran-
dom Forest and Support Vector Machine (SVM) to Neural Networks (NN). Furthermore, a
study by A.Y. et al. [3] emphasized the significance of machine learning models in predicting
downlink throughput on 4G-LTE networks. Their research provides invaluable insights into
the practical applications of these models in real-world network scenarios.
Building on the seminal work of these researchers and addressing the requirements of
contemporary telecommunication networks, this paper presents a detailed analysis of a 4G LTE
network dataset. Concentrating on key metrics such as RSRP, and RSRQ, the research aims
to employ machine learning techniques, notably Linear Regression and Random Forest, to
predict PDCP DL Throughput – a crucial metric in assessing network quality. Recent studies,
including those by D. Minovski et al.[4], and R. Zhohov et al. [5], have further emphasized
the importance and potential of machine learning in throughput prediction, underscoring the
relevance and timeliness of the present research.
While many studies have ventured into network throughput prediction, the distinctiveness
of this research manifests in several pivotal areas:

2
Real World Data:
Grounded in data sourced from an operational 4G LTE network in Nigeria, this research offers
a pragmatic vantage point often eclipsed in predominantly theoretical pursuits.

Granularity of the Data:

The dataset vividly depicts real-time network dynamics with its intricate granularity captured
within a mere 20-minute frame. Such granularity is instrumental in discerning intricacies that
expansive datasets might inadvertently bypass.

Focused Predictors:
Singular emphasis is placed on RSRP (Reference Signal Received Power) and RSRQ (Refer-
ence Signal Received Quality) as the chief predictors, enabling a meticulous probe into these
pivotal metrics.

Temporal Feature Engineering:

In addition to RSRP and RSRQ, a feature engineered with a lag of 1 is seamlessly integrated,
infusing a temporal essence into the predictors. This strategic inclusion is geared towards
encapsulating the inherent temporal interdependencies, bolstering the predictive prowess of
the models.

Revisiting Linear Regression:

Contrary to the prevailing trend of gravitating toward intricate machine learning architectures,
this research reaffirms the merits of the foundational Linear Regression model. Enhanced with
a temporal facet, it underscores its relevance and efficacy in specific contexts.
This research heralds a pragmatic and astute approach to throughput prediction. As the
ensuing pages unravel the findings and insights, it is imperative to acknowledge the meticulous
strategies employed and fathom their ramifications for the telecommunications realm.

2 Materials and Methods

2.1 Data Collection
This research is centered on data derived from a functioning 4G LTE network in Nigeria [6].
The data, specifically the Key Performance Indicators (KPIs), were obtained from stationary
transmitters, commonly referred to as evolved node base stations (eNodeB). These eNodeBs,
with an average height of 25 meters, were equipped with commercial gear from a leading
network provider in Nigeria. Throughout the study, specialized Drive Test (DT) equipment was
employed to capture a range of metrics, including SINR, RSRP, RSRQ, RSSI, among other
vital KPIs, directly from the active sectors of the eNodeBs. My research specifically hones
in on metrics associated with the Packet Data Convergence Protocol (PDCP) Downlink (DL)
Throughput, particularly emphasizing radio measurements like RSRP and RSRQ. The data in
focus was recorded at a 4G LTE frequency of 1876.6MHz, operating within a bandwidth of
10MHz.

3
2.2 Data Preprocessing
Wireless network datasets are inherently intricate, owing to their exposure to various fluc-
tuating environmental and technical variables. Recognizing this complexity, rigorous data
preprocessing was essential to ensure the robustness and reliability of the subsequent analysis.
To achieve a consistent dataset free from site-specific anomalies, focus was narrowed to
data sourced from a single site. This approach aimed to eliminate discrepancies or inconsis-
tencies that might emerge from variations across different sites. To ensure a comprehensive
understanding of the data’s completeness, missingness heatmaps were used to represent miss-
ing values across features visually. Given the critical nature of certain columns, rows with
missing values in these columns were eliminated, and they were determined to be missing at
random. The result of this meticulous cleanse was a dataset with heightened integrity.
The importance of temporal features in the analysis became evident. The Date Time
column was converted into a date time data type, laying the foundation for time series analysis.
The data was then grouped by this temporal feature, and specific aggregations were applied
to other columns to capture the mean within each time group. To further optimize the dataset,
rows with specific abnormal values in the Serving EARFCN column were removed. A lag
feature was introduced based on the PDCP Throughput DL column to add depth to the
analysis. This temporal aspect provides a time-shifted perspective, invaluable for forecasting
and understanding patterns.

2.3 Model Selection

A combination of modeling strategies was employed to address the multifaceted challenge of
forecasting PDCP throughput. Linear Regression, a foundational pillar of statistical modeling,
was harnessed. Enhanced with a temporal feature, it adeptly discerned and accounted for
the time-based patterns and trends intrinsic to the dataset. The essence of this model was to
draw a linear relationship between the target variable, PDCP DL Throughput, and its array of
predictors, prominently spotlighting the temporal aspect.
On the other hand, the Random Forest algorithm provided a more intricate perspective.
This ensemble technique, by weaving together insights from numerous decision trees during
its training phase, offers a profound depth of analysis. For classification, it delivers the mode of
the classes, while in regression scenarios, it presents the mean prediction. Its inherent ability
to grapple with non-linear complexities and unearth subtle patterns in the dataset proved
invaluable in this forecasting endeavor.
These methodologies, each with distinctive strengths, converged to form a robust and
comprehensive forecasting framework.

2.4 Model Evaluation

The efficacy of the employed models was rigorously evaluated using an array of metrics, each
uniquely tailored to gauge different facets of prediction accuracy and model reliability:
Mean Absolute Error (MAE): A straightforward metric, the MAE computes the average of the
absolute discrepancies between forecasted and actual outcomes. Mathematically, it is defined

4
as:
𝑛
1 ∑︁
MAE = |𝑦 𝑖 − 𝑦ˆ 𝑖 |
𝑛 𝑖=1
where 𝑦 𝑖 is the actual value, 𝑦ˆ 𝑖 is the predicted value, and 𝑛 is the number of observations.
Root Mean Squared Error (RMSE): Delving deeper into error magnitudes, the RMSE cap-
tures the square root of the mean of squared deviations between predictions and actual
observations. Its formula is:
v
t 𝑛
1 ∑︁
RMSE = (𝑦 𝑖 − 𝑦ˆ 𝑖 ) 2
𝑛 𝑖=1

R-squared: Primarily associated with linear regression, the 𝑅 2 value elucidates the proportion
of variance in the dependent variable that the independent variables in the model account for.
It is computed as:
SSres
𝑅2 = 1 −
SStot
where SSres is the sum of squares of the residuals and SStot is the total sum of squares.
By harnessing these evaluation techniques, the aim was to measure the prediction precision
of the models for PDCP DL throughput and to furnish insights that can illuminate pathways
for subsequent research endeavors in this arena.

3 Results
3.1 Descriptive Analysis
An examination of the PDCP DL Throughput data over the specified 20-minute interval
revealed its inherently dynamic nature. While no discernible pattern was immediately evident,
the data vividly portrayed wireless networks’ ever-fluctuating and volatile nature. Every pass-
ing second exhibited throughput alterations, underlining the network environment’s non-static
and rapidly evolving characteristics. This continuous oscillation in throughput underscores the
challenges and intricacies of predicting such a metric, given its susceptibility to a multitude
of factors that can change from moment to moment. A visual representation of this dynamic
throughput over the interval can be seen in Figure 1.

5
Fig. 1 Dynamic PDCP DL Throughput over a 20-minute interval.

To further comprehend the data’s characteristics, summary statistics of the numerical

features were computed. These statistics offer insights into the distribution, central tendency,
and spread of the data for each feature.

Table 1 Descriptive statistics of the dataset’s numerical features.

Serving RSRP Serving RSRQ Serving RSSI PCC SINR PHY Throughput DL PDCP Throughput DL
count 1504.0 1504.0 1504.0 1504.0 1504.0 1504.0
mean -85.27 -9.00 -62.02 9.08 7186.58 6026.89
std 7.88 1.10 7.52 6.65 5201.80 4696.75
min -99.91 -14.25 -76.56 -7.62 128.0 0.0
25% -91.28 -9.62 -67.51 3.79 3461.82 2676.51
50% -86.75 -8.83 -63.95 7.95 5326.36 4454.01
75% -79.72 -8.30 -56.82 13.7 9596.29 8297.02
max -59.05 -3.7 -35.92 26.08 28890.94 28040.11

3.2 Model Performance

Central to this research was the task of gauging the forecasting aptitudes of two distinct models
for PDCP DL throughput. Their performances, detailed in Table 2, offer a lens into their
predictive strengths.
The Linear Regression model, bolstered with a temporal aspect, showcased a commendable
accuracy, as evidenced by its 𝑅 2 value. However, when benchmarked against metrics like
MAE and RMSE, the Random Forest model, with its ensemble-based approach, proved to be
slightly superior in its predictive accuracy.
Comparing the performance metrics, it becomes evident that the Random Forest model
holds a slight advantage over the Linear Regression model in terms of predictive accuracy, as
indicated by measures like MAE and RMSE. This comparison is further visualized in Figure 2,
which plots predicted values against actual values. Through this figure, areas where each
model excels or struggles are highlighted, offering a clear view of their respective prediction
strengths and limitations.

6
 Table 2 Performance metrics for model evaluation.

Model MAE RMSE 𝑅2

Linear Regression with Temporal Feature 1,188.59 1,789.64 0.8218
Random Forest 1,100.69 1,736.90 0.8321

Fig. 2 Comparison of predicted values against actual values for the Linear Regression and Random Forest models.

4 Discussion
Predicting PDCP DL Throughput in wireless networks is an intricate endeavor, laden with
both challenges and avenues for the application of advanced predictive modeling. This study
delved deep into these intricacies, utilizing both Linear Regression—enhanced with a temporal
feature—and the Random Forest model to shed light on throughput predictability.
A cornerstone in model evaluation, the Mean Squared Error (MSE) speaks volumes
about prediction accuracy. Both models displayed admirable prowess. Yet, the Random Forest
model slightly edged out its counterpart, registering an MSE of 3,016,817.89 against Linear
Regression’s 3,202,810.38. This edge can be attributed to the ensemble nature of Random
Forest, adept at discerning non-linearities and subtle data patterns.
The 𝑅 2 score, delineating the explanatory power of the models regarding the variations in
PDCP DL Throughput, painted a congruent picture. Both models posted impressive 𝑅 2 scores
exceeding 0.8. The Random Forest model, however, with an 𝑅 2 of 0.8321, slightly surpassed
the 0.8218 score of the Linear Regression model.
Further insights were gleaned from the Mean Absolute Error (MAE) and Root Mean
Squared Error (RMSE) metrics. The close MAEs of 1,188.59 for Linear Regression and
1,100.69 for Random Forest, coupled with respective RMSEs of 1,789.64 and 1,736.90,

7
reiterate the neck-to-neck performance of the two models. Yet, the slight superiority of the
Random Forest model remained consistent across all metrics.
While the empirical data leans toward Random Forest, the virtues of each model in varied
contexts cannot be understated. With its transparency, Linear Regression elucidates clear
feature-target relationships—priceless in situations where clarity supersedes sheer accuracy.
With its nuanced handling of complex feature dynamics, Random Forest becomes the go-to
when top-tier prediction accuracy is the order of the day.
However, it’s imperative to temper these findings with the understanding of the dataset’s
scope—focused on a singular site over a 20-minute span. This dataset, albeit rich, captures a
mere moment in the vast expanse of network operations. When faced with diverse conditions
or prolonged durations, the true mettle of these models beckons further exploration.
To encapsulate, this investigation accentuates the significance of judicious model selection
in the realm of throughput forecasting. While Random Forest clinched slightly superior metrics
in this endeavor, the ultimate choice hinges on the unique demands of the task—whether
it’s model interpretability, sheer accuracy, or computational nimbleness. As the tapestry of
wireless communication grows more intricate, the tools we harness must evolve in tandem,
propelling the field to new pinnacles of innovation and service par excellence.

4.1 Limitations and Future Directions

Regardless of its depth and rigor, every research endeavor has inherent limitations, and this
study is no exception. One of the perennial challenges in machine learning is overfitting,
where a model becomes too attuned to the training data, compromising its generalization to
new, unseen data. Given the high granularity of our dataset and the complexity of the Random
Forest model, there’s a potential risk of overfitting. Regularization techniques or pruning might
be needed to ensure that our models are robust and generalizable.
Additionally, while offering a detailed snapshot, the study’s focus on a singular site over
a brief 20-minute span does not encompass the myriad dynamics of wireless networks over
extended periods or across diverse geographical locales. Such constraints could influence the
model’s adaptability to broader network scenarios.
Future research could address these limitations by incorporating data from multiple sites
or by exploring longer timeframes, thus ensuring a more comprehensive representation of net-
work dynamics. Moreover, experimenting with other advanced predictive models or ensemble
techniques might further enhance the predictive accuracy and robustness against overfitting.
The integration of additional features, perhaps from external datasets or newer technologi-
cal advancements in the wireless domain, could also prove invaluable in refining throughput
predictions.
Ultimately, as the field of wireless communication continues to evolve, there will be an
ever-growing imperative to adapt, innovate, and refine the methodologies employed, ensuring
that research remains abreast of technological progress.

5 Conclusion
Wireless networks are the bedrock of our increasingly digitalized world. Ensuring their optimal
performance is more than just a technical imperative; it’s pivotal to the seamless integration
of technology into our daily lives. In this study, the endeavor to predict PDCP DL Throughput

8
via Linear Regression and Random Forest models cast light on the multifaceted nature of
such a task. While the Random Forest model slightly edged ahead, showcasing the prowess
of ensemble methodologies in deciphering complex data patterns, the Linear Regression’s
performance was not to be overshadowed. Its robustness, especially when bolstered with a
temporal dimension, reiterated the lasting relevance of traditional statistical approaches.
The scope of the research, limited to a dataset from a singular location within a concise
time window, serves as a snapshot—a vignette of the grander tableau of challenges in wireless
network predictions. A key takeaway is the absence of a one-size-fits-all solution. The choice
of predictive model hinges on the nuanced requirements of the task at hand, be it sheer
predictive accuracy, model transparency, or computational pragmatism.
Looking ahead, as we stand at the cusp of a 5G-dominated world with whispers of
6G innovations, the imperative for refined, accurate, and adaptable forecasting tools grows
exponentially. This study underscores the necessity for an adaptive research ethos—one that
is receptive to the swift currents of technological progress. By championing such a spirit of
relentless innovation and introspection, we pave the way for wireless networks that are not
just technically superior but also deeply resonant with the dynamic needs of their users.

References
[1] Yue, C., Jin, R., Suh, K., Qin, Y., Wang, B., Wei, W.: Linkforecast: Cellular link bandwidth
prediction in lte networks. IEEE Transactions on Mobile Computing 17(7), 1582–1594
(2017)

[2] Raca, D., Zahran, A.H., Sreenan, C.J., Sinha, R.K., Halepovic, E., Jana, R., Gopalakrish-
nan, V.: On leveraging machine and deep learning for throughput prediction in cellular
networks: Design, performance, and challenges. IEEE Communications Magazine 58(3),
11–17 (2020)

[3] Al-Thaedan, A., Shakir, Z., Mjhool, A.Y., Alsabah, R., Al-Sabbagh, A., Salah, M., Zec,
J.: Downlink throughput prediction using machine learning models on 4g-lte networks.
International Journal of Information Technology, 1–7 (2023)

[4] Minovski, D., Ogren, N., Ahlund, C., Mitra, K.: Throughput prediction using machine
learning in lte and 5g networks. IEEE Transactions on Mobile Computing (2021)

[5] Zhohov, R., Palaios, A., Geuer, P.: One step further: Tunable and explainable throughput
prediction based on large-scale commercial networks. In: 2021 IEEE 4th 5G World Forum
(5GWF), pp. 430–435 (2021). IEEE

[6] Imoize, A.L., Orolu, K., Atayero, A.A.-A.: Analysis of key performance indicators of a
4g lte network based on experimental data obtained from a densely populated smart city.
Data in brief 29, 105304 (2020)