SliceVision-F2I: Synthetic Visual Dataset for Network Slicing

SlicVision-F2I is a novel multimodal dataset that transforms network slice Key Performance Indicators (KPIs) into multiple visual representation patterns, designed to bridge telecommunications and computer vision for next-generation network management systems.

Dataset Description

SlicVision-F2I contains 30,000 samples of network slice KPIs converted into four distinct visual representation patterns. Each sample represents one of three primary 5G network slice types:

1. eMBB (Enhanced Mobile Broadband): High-throughput applications
2. URLLC (Ultra-Reliable Low-Latency Communications): Mission-critical services
3. mIoT (Massive Internet of Things): Large-scale sensor networks

The dataset enables novel approaches to network management by providing multiple visual representations of identical underlying KPI data, facilitating multimodal learning and cross-pattern analysis.

Key Features

• Multi-pattern representations: Four distinct visual encodings per sample
• Comprehensive KPI coverage: 10 key network performance metrics with realistic correlations
• Real-world characteristics:
  • Simulated measurement noise (5-15%)
  • Missing values (~5% random missingness)
  • Natural class imbalance (eMBB:URLLC:mIoT = 2:1:7 ratio)
• Production-ready:
  • Pre-normalized values (0-1 range)
  • Missing value imputation (median-based)
  • Consistent 16×16 RGB image format

Potential Use Cases

Network Management Applications

• Anomaly Detection: Train vision-based models to identify slice performance degradation
• Slice Classification: Multi-class recognition of slice types from KPI patterns
• Quality Prediction: Regression models for QoE metrics from visual representations

Machine Learning Research

• Multimodal Learning: Study cross-pattern relationships and ensemble methods
• Data Augmentation: Test augmentation strategies across different representations
• Explainable AI: Visual interpretability of network performance decisions

Telecommunications Education

• Teaching Resource: Visual demonstrations of network performance concepts
• Benchmarking: Standard dataset for comparing network AI approaches
• Prototyping: Rapid development of visual network analytics tools

Dataset Structure

SlicVision-F2I/
├── numeric_data.csv            # Raw KPI measurements and slice labels
├── guided_patterns.npy         # Physics-inspired representations
├── perlin_patterns.npy         # Procedural noise-based patterns
├── wallpaper_patterns.npy      # Structural/geometric patterns
└── fractal_patterns.npy        # Recursive branching patterns

Data Fields Specification

Field	Description	Normalized Range	Physical Range
slice_type	Slice category	-	{eMBB, URLLC, mIoT}
delay	End-to-end latency	[0,1]	0-100ms
jitter	Latency variation	[0,1]	0-50ms
loss	Packet loss rate	[0,1]	0-10%
throughput	Data rate	[0,1]	0-300Mbps
retransmissions	Retry rate	[0,1]	0-10%
packet_discard_rate	Drop rate	[0,1]	0-10%
rssi	Signal strength	[0,1]	-100dBm to -30dBm
snr	Signal quality	[0,1]	0-40dB
cpu_util	Processor usage	[0,1]	0-100%
mem_util	Memory usage	[0,1]	0-100%
label	Class encoding	{0,1,2}	-

Visual Representation Patterns

1. Physically-Guided Patterns

Design Philosophy: Embeds physical network relationships into spatial patterns

Key Characteristics:

• Gaussian blobs represent concentrated performance metrics
• Wave patterns show periodic behaviors
• Color channels encode related KPI groups:
  • 🔴 Red: Latency metrics (delay/jitter/loss)
  • 🟢 Green: Throughput metrics
  • 🔵 Blue: System health (RSSI/SNR/resource usage)

2. Perlin Noise Patterns

Design Philosophy: Organic patterns reflecting natural network variability

Key Characteristics:

• Noise parameters dynamically adjust to KPI values
• Each channel has unique generation parameters:
  • Red: Octaves scaled by loss rate
  • Green: Persistence set by throughput
  • Blue: Lacunarity adjusted by SNR

3. Wallpaper Patterns

Design Philosophy: Structural representations of network periodicities

Key Characteristics:

• Combines multiple geometric primitives:
  • Stripes → Throughput levels
  • Grids → Packet loss patterns
  • Radial gradients → Signal strength
• Parameters adapt to slice type characteristics

4. Fractal Branching Patterns

Design Philosophy: Tree structures modeling network paths

Key Characteristics:

• Branching complexity scales with traffic volume
• Leaf density reflects packet success rates
• Trunk stability indicates connection reliability
• Color gradients show resource utilization

Dataset Statistics

Category	Metric	Value
Samples	Total	30,000
	eMBB	6,000 (20%)
	URLLC	3,000 (10%)
	mIoT	21,000 (70%)
Patterns	Resolution	16×16px
	Color Channels	3 (RGB)
	Size on Disk	~550MB
Quality	Missing Values	4.8%
	Noise Level	5-15%
	Normalization	Min-Max [0,1]

Dataset Testing Suite

This dataset-testing folder contains the core analysis scripts for evaluating the SlicVision-F2I dataset using both traditional machine learning and deep learning approaches.

Files Overview

dataset-testing/
├── traditional_ml_analysis.py  # Classical ML analysis script
├── cnn_analysis.py            # CNN-based evaluation script
└── requirements.txt           # Python dependencies

1. Traditional ML Analysis

traditional_ml_analysis.py - Evaluates the raw KPI data using classical machine learning algorithms.

Features:

• Tests three classifier types:
  • Random Forest
  • Support Vector Machine (SVM)
  • XGBoost
• Generates:
  • Feature importance plots
  • Correlation matrices
  • Classification report

Usage:

python traditional_ml_analysis.py

2. CNN Analysis

cnn_analysis.py - Evaluates all four visual pattern types using convolutional neural networks.

Features:

• Tests four pattern representations:
  • Guided patterns
  • Perlin noise patterns
  • Wallpaper patterns
  • Fractal patterns
• Generates:
  • Training history plots
  • Model accuracy comparisons
  • Classification reports

Usage:

python cnn_analysis.py

Requirements

Install dependencies with:

pip install -r requirements.txt

requirements.txt contents:

numpy
pandas
scikit-learn
xgboost
tensorflow
matplotlib
seaborn

Running the Full Test Suite

1. First run traditional ML analysis:

python traditional_ml_analysis.py

2. Then run CNN evaluation:

python cnn_analysis.py

Adding New Models

To extend this testing suite:

1. For traditional ML:

   • Add new classifiers to traditional_ml_analysis.py
   • Append evaluation metrics to existing JSON output

2. For CNN models:

   • Create new architecture in cnn_analysis.py
   • Follow the existing pattern comparison framework

Troubleshooting

Common issues:

• Shape mismatches: Ensure numpy files match CSV metadata length
• Memory errors: Reduce batch size in CNN training
• NaN values: Confirm all KPIs are properly imputed in the CSV

Citation

@misc{rafi2025slicevision,
  author       = {Rafi, Abid Hasan and Johora, Mst. Fatematuj and Bhowmik, Pankaj},
  title        = {{SliceVision-F2I: Synthetic Visual Dataset for Network Slicing}},
  year         = {2025},
  publisher    = {Mendeley Data},
  version      = {V1},
  doi          = {10.17632/68xp3vszsz.1},
  url          = {https://doi.org/10.17632/68xp3vszsz.1}
}

Core Contributors

• Abid Hasan Rafi
  Dataset creation, pattern design, KPI modeling
• Mst. Fatematuj Johora
  Dataset management, Data documentation
• Pankaj Bhowmik
  Dataset evaluation, Process supervision

Disclaimer

Any sort of use, reproduction, or distribution of the content in this repository without proper citation or reference to the author is strictly prohibited. Please respect the intellectual property and give appropriate credit when using or referencing any part of this work.