Reproduce “Bridging the Sampling Distribution Shift in Radio Map Estimation: A Trajectory-Aware Paradigm” on OpenResearch

About this paper

Feng Qiu, Zheng Fang, Shuhang Zhang, Kangjun Liu, Longkun Zou, Jing Liu, Ke Chen

This work characterizes the "sampling distribution shift" in Radio Map Estimation (RME), where models trained on idealized random samples fail on real-world trajectory-based observations. Researchers from Xidian University and Pengcheng Laboratory developed a trajectory-aware training paradigm, Stochastic-Triggered Trajectory-Based Sampling (ST-TBS), which consistently improved model robustness, reducing RMSE by up to 78.31% on the SpectrumNet dataset under trajectory-constrained testing.

Key insights

• Trajectory-based sampling (TBS) inherently leads to reduced spatial diversity and increased information redundancy compared to random sampling (RS), lowering the expected mutual information and increasing the unavoidable reconstruction error.
• RME models trained on RS suffer catastrophic performance degradation (up to nearly seven-fold RMSE increase) when evaluated with TBS, indicating that architectural improvements alone do not resolve this fundamental sampling distribution mismatch.
• Aligning the training distribution with realistic trajectory-constrained sampling via ST-TBS significantly improves model robustness and generalization to real-world deployment scenarios without compromising performance under idealized random sampling conditions.

Summary by alphaXiv.