🤖 machine learning

Multimodal Spatiotemporal-Frequency Fusion with Peak Enhancement for Cellular Traffic Forecasting

This paper proposes MSPF-Net, a multimodal framework that integrates spatiotemporal-frequency analysis, peak enhancement for burst detection, and news context encoding to significantly improve the accuracy of cellular traffic forecasting by jointly modeling intrinsic dynamics and exogenous urban events.

Original authors: Qingzhong Li, Yue Hu, Hui Ma, Yajun Zhang, Xinjun Pei, Ming Yan, Fei Xing

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Qingzhong Li, Yue Hu, Hui Ma, Yajun Zhang, Xinjun Pei, Ming Yan, Fei Xing

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how busy a city's phone network will be in the next hour. It's like trying to guess how many people will be at a concert, but the "concert" is happening in every neighborhood simultaneously, and the crowd size changes every minute.

Most current prediction tools are like weather forecasters who only look at the thermometer. They see the temperature rising and falling in a predictable pattern (like rush hour traffic) and guess the future based on that. But they often miss the sudden, chaotic spikes—like when a famous singer shows up unexpectedly, or a major news story breaks, causing everyone to grab their phones at once.

This paper introduces a new tool called MSPF-Net. Think of it as a super-smart traffic forecaster that doesn't just look at the phone data; it also reads the local newspaper, watches the news, and pays extra attention to sudden surges.

Here is how it works, broken down into four simple parts:

1. The "Big Picture" Scanner (Spatiotemporal-Frequency Traffic Encoder)

First, the system looks at the phone traffic history. It doesn't just see a line going up and down; it analyzes the rhythm.

Time: It sees the daily patterns (morning commute, lunch break).
Space: It knows that if one neighborhood gets busy, the neighbors usually do too.
Frequency: It listens to the "beat" of the data, spotting repeating cycles like a song's chorus.
The Metaphor: This is like a musician listening to a song and identifying the steady drumbeat and the melody.

2. The "Spike Detector" (Peak Enhancement Module)

Here is where the new system shines. Regular patterns are easy to predict, but sudden, loud spikes are hard. Standard tools often smooth these spikes out, treating them like background noise.

The Innovation: MSPF-Net has a special "spotlight" that zooms in on sudden jumps. It calculates how fast the traffic is changing and looks for the difference between the highest point and the average point in a short window.
The Metaphor: If the regular traffic is a calm river, this module is the sensor that screams, "Look out! A waterfall just appeared!" It ensures the system doesn't ignore the sudden rush of water.

3. The "News Reader" (News Context Representation Module)

Phone usage isn't just about the phone; it's about what's happening in the real world. A big sports game, a traffic accident, or a breaking news story causes people to call and text more.

The Innovation: The system reads news headlines and event logs. It turns these stories into a simple list of numbers (e.g., "5 new events," "3 cities mentioned"). It then feeds this "context" into the prediction model.
The Metaphor: This is like a weather forecaster who doesn't just look at the clouds but also checks the local event calendar. If they see "Parade at 2 PM," they know the traffic will be weird, even if the clouds look normal.

4. The "Smart Mixer" (Dynamic Fusion Prediction Module)

Now the system has three different types of information: the phone history, the sudden spikes, and the news context.

The Innovation: Instead of just averaging these three things together (which is like making a smoothie where you can't taste the individual fruits), this module acts like a DJ. It listens to the current situation and decides, "Right now, the news is most important," or "Right now, the sudden spike matters most." It blends them together dynamically.
The Metaphor: Imagine a chef tasting a soup. Sometimes they add more salt (news), sometimes more pepper (spikes), and sometimes just let the broth (history) shine. This module decides the perfect recipe for the exact moment.

The Results

The researchers tested this system on real data from cities like Milano and Trento, as well as general mobile network data.

The Outcome: MSPF-Net was much better at predicting traffic than older methods. It made fewer mistakes, especially when things got chaotic or when a big event happened.
Why it matters: By combining the "rhythm" of the network, the "shouts" of sudden spikes, and the "stories" from the news, the system can see the future more clearly than tools that only look at the past.

In short, this paper presents a smarter way to predict phone network traffic by teaching the computer to pay attention to the news and the sudden surprises, not just the usual daily routine.

Technical Summary: MSPF-Net for Cellular Traffic Forecasting

Problem Statement
Accurate forecasting of cellular network traffic is critical for network planning, resource allocation, and quality-of-service assurance. However, real-world traffic is characterized by complex spatiotemporal dynamics, including bursty endogenous fluctuations and disturbances triggered by external urban events. Existing methods face two primary limitations:

Inadequate Burst Modeling: Most approaches rely on unimodal traffic representations or shallow fusion strategies, failing to capture sudden spikes, anomalies, and short-term fluctuations caused by event-driven demand surges.
Insufficient Exogenous Context Utilization: Current models predominantly focus on intrinsic temporal or spatiotemporal dependencies derived from historical traffic sequences. They often underutilize heterogeneous external signals (e.g., urban events, news streams) that substantially influence traffic dynamics but are not explicitly reflected in historical observations alone.

Methodology: MSPF-Net
To address these challenges, the authors propose MSPF-Net (Multimodal Spatiotemporal-Frequency Fusion with Peak Enhancement Network), a unified framework that jointly models endogenous traffic dynamics and exogenous contextual signals. The architecture comprises four cooperative components:

Spatiotemporal-Frequency Traffic Encoder: This module captures intrinsic traffic patterns by integrating three dimensions:
- Temporal Encoding: Uses temporal self-attention to model long-range dependencies and regular evolution.
- Frequency Encoding: Applies Fast Fourier Transform (FFT) to extract amplitude spectra, preserving periodic and oscillatory patterns.
- Spatial Encoding: Utilizes graph convolution over a normalized adjacency matrix to model interactions among neighboring cellular nodes.
  These representations are aggregated via residual fusion to form the foundational traffic embedding ( $H_{traf}$ ).
Peak Enhancement Module: Designed to explicitly preserve high-impact local variations that global encoders might smooth out. It computes temporal differences and constructs a peak descriptor using short-window statistics (combining raw data, differences, and max-minus-average pooling). A short-window 1D convolutional network then encodes these features to produce a burst-aware representation ( $H_{peak}$ ).
News Context Representation Module: This module encodes exogenous information derived from urban events and news streams. It constructs a five-dimensional context vector per time step (active users, mentioned cities, entities, event types, and news items). These vectors are mapped to a latent space and processed by a Transformer encoder to capture the temporal dynamics of external signals ( $H_{news}$ ), which are then broadcast across the spatial dimension.
Dynamic Fusion Prediction Module: Instead of static fusion, this module performs adaptive multimodal integration. It projects the three representations ( $H_{traf}$ , $H_{peak}$ , $H_{news}$ ) into a shared space and uses cross-modal attention to enrich the traffic representation with burst cues and news context. A gating network computes dynamic weights ( $\alpha, \beta, \gamma$ ) for each modality at every spatiotemporal position, allowing the model to adaptively determine the contribution of each signal source before generating the final forecast via a multi-layer perceptron (MLP).

Key Contributions
The paper outlines three main contributions:

Multimodal Framework: A unified framework that integrates endogenous traffic dynamics with exogenous contextual signals, enabling joint modeling of intrinsic spatiotemporal-frequency patterns and external event influences.
Peak Enhancement: A dedicated module designed to learn burst-aware representations, capturing sudden traffic spikes and variation patterns often smoothed by conventional sequence encoders.
Dynamic Fusion: A prediction module that performs adaptive multimodal fusion of traffic, burst, and news signals, outperforming static fusion methods by adjusting to changing conditions.

Experimental Results
The framework was evaluated on three datasets: Milano, Trento, and LTE traffic.

Performance: MSPF-Net achieved the lowest prediction errors (MAE and RMSE) across all datasets compared to baselines including LSTM, Transformer, FEDformer, TimeMixer, and various graph-based methods (e.g., DDGCRN, FISTGCN). For instance, on the Milano dataset, MSPF-Net achieved an MAE of 2.2850, significantly outperforming the next best method (MSCR at 3.9326).
Ablation Study: Removing the Peak Enhancement Module caused a clear accuracy drop, particularly during burst-heavy intervals. Similarly, removing the News Context module degraded performance, confirming that intrinsic signals alone are insufficient for handling exogenous disturbances. Replacing the dynamic fusion with static aggregation further reduced accuracy, validating the necessity of adaptive weighting.
Visualization: Comparisons of ground-truth versus predicted values demonstrated that MSPF-Net more closely follows real traffic trends and better captures sudden fluctuations than second-best baselines.

Significance and Claims
The authors claim that treating cellular traffic forecasting as a multimodal representation learning problem allows for a more comprehensive capture of spatiotemporal dependencies, spectral periodic patterns, and the influence of external events. The study demonstrates that jointly modeling traffic dynamics, burst patterns, and news contextual signals effectively improves forecasting performance, particularly in burst-heavy and event-driven scenarios. The work suggests that adaptive multimodal fusion is critical for robust forecasting under real-world disturbances. Future work is suggested to explore richer graph-based spatial modeling and improved alignment between traffic signals and real-world event streams.