NS-CUK Seminar: S.T.Nguyen, Review on "Weather-Aware Fiber-Wireless Traffic Prediction Using Graph Convolutional Networks", IEEE 2022
1. LAB SEMINAR
Nguyen Thanh Sang
Network Science Lab
Dept. of Artificial Intelligence
The Catholic University of Korea
E-mail: sang.ngt99@gmail.com
Weather-Aware Fiber-Wireless Traffic Prediction
Using Graph Convolutional Networks
--- Abdullah, Mariam, Jiayuan He, and Ke Wang ---
2023-04-28
3. 2
Introduction
+ Fiber-wireless technology: a flexible and efficient solution for high-speed data transmission, enabling users to access the internet
and communicate seamlessly across multiple devices and locations.
+ Fiber-wireless networks are commonly used in a range of applications, including internet access, cellular networks, and
enterprise networks. These networks can support a wide range of devices, including smartphones, laptops, tablets, and other
connected devices.
+ The fiber-wireless integrated network has been widely investigated, taking advantage of the large capacity and low transmission
loss properties of the optical fiber to extend the coverage of wireless networks and the cellular networks in particular.
4. 3
Problems
❖ Machine learning (ML) strategies have been put forth to optimize resource allocation in
these networks. These techniques seek to forecast cellular traffic in advance to enable
proactive resource allocation.
❖ The spatial correlation of adjacent based stations and the temporal dynamics noted in
historical records are the two key aspects taken into account by existing works when
predicting traffic.
❖ The unexplored aspect, i.e., meteorological factors, such as rain, wind, or temperature,
have not yet been considered.
5. 4
Contributions
• A wireless traffic prediction model that seamlessly unifies the three aspects: temporal, spatial,
and meteorological factors that learns the traffic patterns through a novel GCN-GRU cell, which
has a hierarchical structure to learn the impact of all three factors.
• Considering an important and yet unexplored factor: the meteorological context, such as rain,
wind, and temperature.
6. 5
Weather impact on cellular traffic
• The weather is sunny, more users will go to the outdoor areas such as parks and beaches, which result in
high loads in the two corresponding base stations
• The weather is rainy, more users will be indoor, and more traffic loads will shift to base stations near those
indoor regions.
7. 6
Weather-traffic correlation
• Performing statistical analysis with telecommunication and weather information.
• Consider four features of weather context, namely temperature, wind speed, wind direction, and
humidity.
• The kernel density estimation (KDE) of the four weather features in relation to telecommunication
activity.
• Spearman’s rank correlation coefficient for each weather context:
8. 7
Network
• Communication network:
• Traffic records: a record of the network traffic is a 3-tuple <t, v, x>, where t is the
timestamp associated with the traffic record, v is a node in G, and x is the observed traffic
volume.
• Weather context: a 5-tuple <t, v, temp, hum, wind>, where t the timestamps associated
with the traffic record, v is a node in G, and temp, hum, and wind are the observed
temperatures, humidity and wind speed.
9. 8
Architecture
• Modeing spartial features: GCN model works by constructing filters in
the Fourier domain, which are then applied to all nodes and their first-
order neighbours:
• Modeling temporal features:
• Two child GRUs to capture the temporal dynamics in historical traffic
records and weather contexts, respectively
10. 9
Training
• Treating the prediction of traffic volume as a sequence to sequence learning task:
• Mean Squared Error loss:
11. 10
Dataset
• The multi-source dataset of urban life in the city of Milan and the Province of Trentino.
• The traffic and weather data were aggregated by hour for a total of 1,463 records over
the two month period.
13. 12
Baseline experiments
• Only the spacial correlation carried by the historical traffic features is crucial to the
prediction of future traffic.
• This is proved by the worse performance of Model 2, which extracts spatial features from
the combined hidden states of the traffic and weather data, compared with Model 3,
which extracts spatial features form the traffic data directly.
14. 13
Influence of hyperparameters
• when the number of the hidden layers increases, the performances of all models first improve up to a
certain point before decreasing.
because when the complexity of model increases, a model generally has better learning capability.
• However, when the complexity increases to a certain point, the model starts to overfit due to the limited
dataset size and hence, the model performance decreases.
15. 14
Model interpretation
• Model 3 is able to predict traffic patterns with a higher precision than the baseline model.
• Model 3 seems to be less accurate when there are sudden changes in traffic.
This could be due to the smooth filter in the Fourier domain which is defined by the GCN, which makes
the model less accurate prediction with sudden changes.
16. 15
Conclusions
• Proposed a novel GCN-GRU model for the accurate prediction of wireless traffic which shown its
capability in jointly learning the spatial, temporal, and weather dynamics to predict cellular traffic.
• This paper demonstrated the importance of the highly-efficient incorporation of the weather context (e.g.,
humidity, wind speed, and temperature) in wireless traffic prediction.
• The experiments have confirmed the effectiveness of the proposed model, with up to 20.2%, 24.8%, and
3.3% of improvement in RMSE, MAE, and accuracy, respectively.
• The model can be extrapolated to other seasons in principle, whilst accuracy will decrease (e.g. in
summer months where the temperature is higher, the traffic trend typically indicates more indoor usage).