This project investigates multi-task learning (MTL) architectures for real-time interference detection and Key Performance Indicator (KPI) prediction in 5G Radio Access Networks (RANs). The objective is to jointly solve two heterogeneous tasks:

Interference detection (binary classification)

KPI prediction (continuous regression)

while mitigating negative transfer between tasks and minimizing computational overhead, which is critical for real-time deployment in wireless networks.

We conduct a systematic comparison between Single-Task Learning (STL) and several state-of-the-art MTL architectures, analyzing trade-offs among prediction accuracy, regression error, model complexity, and inference latency.

The following models are implemented and evaluated:

• STL (Single-Task Learning) Separate GCN models for each task

• Hard Parameter Sharing Shared GCN backbone with task-specific output heads

• MMoE (Multi-gate Mixture-of-Experts) Shared expert layers with task-specific gating networks

• Cross-Stitch Networks Learnable feature-sharing layers between tasks

• PLE (Progressive Layered Extraction) Multi-level shared and task-specific experts with gating

• Attention-based MTL GCN backbone with task-specific attention mechanisms

Model Outputs:

regression_out → [RSRP, RSRQ, SINR]

cls_out → interference label

All architectures are fully compatible with the same preprocessing pipeline and training loop.

• Training loop implementation
• Validation and testing pipeline
• Multi-task loss computation
  • Classification loss
  • Per-task regression losses
• Model statistics computation
  • Number of parameters
  • FLOPs
  • Model size

We also defined the network architecture and training hyperparameters for the models, which are presented in the following:

regression_out & cls_out predictions (saved + visualized)

Comparison of models by accuracy, loss, training time, inference time, and size

• numpy
• pandas
• torch
• scikit-learn
• matplotlib
• seaborn
• thop

For questions or collaboration:

Email: [email protected]

Alternative: [email protected]