Machine Learning-Based Cellular Network Handover Prediction: From Signal Quality to Intelligent Decision-Making

This article introduces the GitHub open-source project network-handover-prediction (released on June 8, 2026) maintained by chynaenye. The project uses machine learning to predict cellular network handover events. By analyzing signal quality metrics (RSRP, RSRQ, SINR), device movement data (speed, acceleration), and geographic location information, combined with classic algorithms such as logistic regression, decision trees, random forests, and XGBoost, it achieves intelligent handover prediction, providing new ideas for mobile network optimization.

Traditional handover algorithms based on fixed thresholds struggle to meet the low-latency and high-reliability requirements of the 5G era. For example, in high-speed rail scenarios, switching too early wastes resources, while switching too late causes call drops or video stuttering. This project aims to solve this classic communication problem using machine learning.

The project uses four classic algorithms for comparison: 1. Logistic regression (baseline model with strong interpretability); 2. Decision tree (captures non-linear relationships between features and generates intuitive rules); 3. Random forest (integrates multiple trees to reduce overfitting and adapt to noisy data); 4. XGBoost (gradient boosting framework, expected to perform best on structured data).

Feature design includes: signal quality metrics (RSRP signal strength, RSRQ signal purity, SINR signal-to-interference-plus-noise ratio), device movement data (speed, acceleration), and geographic location information (identifies cell boundaries and coverage blind spots). The combination allows understanding the signal environment leading to handover failures.

For operators: Optimize resource allocation and reduce signaling overhead from ping-pong handovers; For users: Improve call and video experience in high-speed mobile scenarios (high-speed rail, subway); For researchers: Provide an extensible benchmark framework to facilitate testing new features or deep learning models.

The project proves that classic supervised learning plus reasonable feature engineering can solve complex wireless network problems; For future 5G/6G scenarios (fast millimeter-wave attenuation, frequent handovers in ultra-dense networks, millisecond-level latency for V2X), time-series models (such as LSTM, Transformer) or reinforcement learning need to be introduced. Data-driven optimization has become an industry consensus.