Skip to main content

KYNET Real-Time Mobile Network Performance Prediction

Overview#

KYNET (Know Your Network) is a full-stack system I designed and developed to predict mobile network performance in real time.
It provides users with predictions of:

  • Download/Upload Speeds
  • Network Latency (Ping)
  • Signal Strength (RSRP – Reference Signal Received Power)
info

Unlike conventional speed-test apps that only show current results, KYNET provides future-oriented, real-time predictions using incremental ML models that continuously adapt to new data.

Problem Statement#

Mobile network performance is non-stationary — it changes dynamically with time, location, and environment. This creates several challenges:

  • Spatial Variability: Signal strength fluctuates between indoor/outdoor locations, across different floors, and even within the same building.
  • Temporal Variability: Networks behave differently during peak hours vs. late nights.
  • ISP Differences: Service quality varies between providers like Jio, Airtel, and BSNL.
  • Environmental Factors: Weather, temperature, and user mobility (walking vs. stationary) affect performance.

Gaps in Existing Solutions#

  • Traditional apps provide snapshot measurements but no predictive insights.
  • No public dataset existed for network speed and signal strength prediction.
  • Most ML approaches in this domain use batch learning, which fails to adapt to continuous data streams.
warning

The lack of public datasets was a major bottleneck — without data, predictive ML models for mobile networks cannot be trained effectively.

Solution#

To overcome these challenges, I created KYNET, an integrated platform consisting of a data collection app, prediction models, and visualization tools.

1. Data Collection Module#

I built a custom Android app in Flutter to crowdsource real-world mobile network data.
Over 21,000+ samples were collected across Anna University campus.

info

This dataset is one of the first of its kind — designed specifically for real-time mobile network prediction.

Each record includes a rich set of network, spatial, temporal, and environmental parameters, such as:

  • Network Quality Metrics

    • RSSI, RSRP, RSRQ, SINR
    • RF-NC, p-a, Band, Num Carriers
    • RS Path Loss, Timing Advance (TA), Shannon, Cell Load
    • RF-RX0, RF-RX1, RF-RX2, RF-RX3
    • RI Sum, CQI, CRI

  • Temporal Information

    • Timestamp
    • Day of Week (Mon–Sun)
    • Type of Day (Weekday/Weekend)
    • Time Period (Morning/Afternoon/Evening/Night)

  • Network & ISP Details

    • ISP (Jio, Airtel, BSNL)
    • Network Type (2G/3G/4G/5G)
    • Download Speed, Upload Speed, Latency
    • General Signal Strength

  • Spatial & Environmental Context

    • Location Name, Latitude, Longitude
    • Indoor/Outdoor environment
    • Environment Type (Crowded/Free)
    • Floor (if applicable)
    • Weather & Temperature

2. Prediction Engine#

  • Models Used:
    • Adaptive Random Forest → location-based predictions
    • Aggregated Mondrian Forest Regressor → speed & latency prediction
  • Feature Engineering:
    • Split Latitude/Longitude → improved spatial accuracy
    • Temporal encoding → captured peak/off-peak patterns
tip

Using incremental learning allowed KYNET to continuously update its models without retraining from scratch — saving compute and enabling scalability.

3. Location Prediction & Visualization#

  • Forecasted best areas on campus for connectivity.
  • Produced heatmaps and overlays for ISP-specific performance.
  • Offered ranked location recommendations for users.

4. Real-Time User App#

  • Cross-platform Flutter app (Android/iOS ready).
  • Provides instant predictions of speed, latency, and signal strength.
  • Dashboards with charts, time-series insights, and map visualizations.

Technology Stack#

  • Frontend (Mobile App): Flutter
  • Backend: Python (Flask REST APIs)
  • Machine Learning: Scikit-learn + River (incremental ML)
  • Database: SQLite
  • Hosting: Microsoft Azure VM
  • Visualization: Google Maps API, Flutter charts

Results#

  • Model Accuracy:

    • Download speed → 91.97% (R²)
    • Upload speed → 85.83% (R²)

  • Real-time Predictions:

    • Latency, RSRP, speeds predicted instantly
    • Incremental updates with new data

  • Network Insights:

    • Best times/locations identified
    • ISP quality comparisons (Jio vs Airtel vs BSNL)
tip

KYNET is not just a research prototype — it was deployed campus-wide and validated with 21,000+ real samples, making it practical and field-tested.

Novelty & Impact#

  • First-of-its-kind crowdsourced dataset.
  • Incremental ML models rarely applied in this domain.
  • Spatial feature engineering improved accuracy.
  • Real-time recommender system for end-users.
  • Multi-ISP support (Jio, Airtel, BSNL).

Screenshots#

  • Data Collection Interface
  • Real-Time Prediction Dashboard
  • Heatmaps (best zones for connectivity)
  • ISP Analysis

(Screenshots go in static/img/kynet/.)

Key Learnings#

  • Designing incremental ML pipelines for streaming data.
  • Building datasets from scratch when none exist.
  • Bridging research and practical application in a mobile app.
note

KYNET was both a research project and a working system, combining machine learning, mobile development, and cloud engineering in one project.

Links#

Thesis Report

You can download the full KYNET Project Thesis here:
📥 Download PDF