Time Series Prediction in Large Scale Data

Overview

This project is a Python-based framework designed for efficient and flexible time series forecasting. It employs a hybrid architecture combining Transformers and Echo State Networks (ESN) to process large-scale datasets effectively. The project is structured for easy customization and experimentation with various hyperparameters.

The work is contribution on below PoC and Demo:

1. F. Rezazadeh, S. Barrachina-Muñoz, H. Chergui, J. Mangues, M. Bennis, D. Niyato, H. Song, and L. Liu, “Toward Explainable Reasoning in 6G: A Proof of Concept Study on Radio Resource Allocation”, IEEE Open Journal of the Communications Society, 2024. [IEEE Xplore] [arxiv]

2. F. Rezazadeh, S. Barrachina-Muñoz, E. Zeydan, H. Song, K.P. Subbalakshmi, and J. Mangues-Bafalluy, “X-GRL: An Empirical Assessment of Explainable GNN-DRL in B5G/6G Networks”, IEEE NFV-SDN, 2023. [IEEE Xplore] [arxiv]

---

Requirements

• Python Version: 3.11.11

Ensure you are using Python 3.11 for compatibility. Install the necessary dependencies using the requirements.txt file:

pip install -r requirements.txt

---

Running the Project

Main Script

The project is run using the main.py script, which is fully configurable through command-line arguments.

Example Usage

python main.py \
    --dataset_name "kimdaegyeom/5g-traffic-datasets" \
    --file_relative_path "5G_Traffic_Datasets/Video_Conferencing/Zoom/Zoom_1.csv" \
    --sequence_length 65 \
    --batch_size 128 \
    --d_model 64 \
    --n_heads 8 \
    --num_encoder_layers 3 \
    --ff_dim 128 \
    --reservoir_size 100 \
    --spectral_radius 0.85 \
    --sparsity 0.1 \
    --lr 0.001 \
    --weight_decay 1e-5 \
    --epochs 50 \
    --patience 10 \
    --max_data 150000

Argument Details

• --dataset_name: (str) Kaggle dataset identifier (default: "kimdaegyeom/5g-traffic-datasets").
• --file_relative_path: (str) Relative file path inside the dataset (default: "5G_Traffic_Datasets/Video_Conferencing/Zoom/Zoom_1.csv").
• --sequence_length: (int) Sequence length for time series data (default: 65).
• --batch_size: (int) Batch size for training (default: 128).
• --d_model: (int) Transformer model dimension (default: 64).
• --n_heads: (int) Number of attention heads (default: 8).
• --num_encoder_layers: (int) Number of Transformer encoder layers (default: 3).
• --ff_dim: (int) Feedforward network dimension (default: 128).
• --reservoir_size: (int) ESN reservoir size (default: 100).
• --spectral_radius: (float) ESN spectral radius (default: 0.85).
• --sparsity: (float) Reservoir sparsity (default: 0.1).
• --lr: (float) Learning rate (default: 0.001).
• --weight_decay: (float) Weight decay for optimizer (default: 1e-5).
• --epochs: (int) Number of training epochs (default: 50).
• --patience: (int) Early stopping patience (default: 10).
• --max_data: (int) Maximum number of data points to use (default: 150000).

---

Example Workflow

1. Download Dataset: The dataset is downloaded using the kagglehub library. Specify the --dataset_name and --file_relative_path to locate the desired dataset file.

2. Preprocessing: The script automatically scales the dataset and creates sequences for training.

3. Model Training: The hybrid Transformer-ESN model is trained on the processed dataset with specified hyperparameters.

---

Results

Below is a visualization of the model's performance (MSE = 0.005):

---

Contributions

Feel free to fork this repository, report issues, or submit pull requests to improve the framework.

---

License

This project is licensed under the MIT License.