Reliable Wireless Indoor Localization via Cross-Validated Prediction-Powered Calibration

TL;DR

This work tackles the challenge of uncertainty quantification in RSSI fingerprint–based indoor localization under scarce labeled calibration data. It introduces RCPS-CPPI, a cross-validated, semi-supervised calibration framework that leverages synthetic labels from an auxiliary predictor while correcting bias via cross-validation, yielding prediction sets with rigorous $\alpha$-level risk guarantees and improved efficiency. The method provides unbiased risk estimators and per-fold confidence bounds, ensuring $(\alpha,\delta)$-reliability and tighter coverage than existing RCPS approaches. Experiments on a WiFi fingerprinting dataset demonstrate that RCPS-CPPI achieves target coverage with notably smaller prediction sets, highlighting practical gains in calibration data efficiency for real-world localization systems.

Abstract

Wireless indoor localization using predictive models with received signal strength information (RSSI) requires proper calibration for reliable position estimates. One remedy is to employ synthetic labels produced by a (generally different) predictive model. But fine-tuning an additional predictor, as well as estimating residual bias of the synthetic labels, demands additional data, aggravating calibration data scarcity in wireless environments. This letter proposes an approach that efficiently uses limited calibration data to simultaneously fine-tune a predictor and estimate the bias of synthetic labels, yielding prediction sets with rigorous coverage guarantees. Experiments on a fingerprinting dataset validate the effectiveness of the proposed method.

Figures (4)

• Figure 1: Comparison of risk-controlling prediction sets for the task of wireless indoor localization of mobile devices under three calibration strategies: RCPS Bates2021RCPS, which uses only real-world labeled data; RCPS-PPI Romano2024RCPS_PPI, which splits the real data into two subsets—one for fine-tuning the label-generating predictive model and the other for estimating the model-induced bias on synthetic labels; and the proposed RCPS-CPPI, which uses the entire labeled dataset for both predictor fine-tuning and bias estimation via cross-validation, yielding more efficient prediction sets without compromising coverage.
• Figure 2: Empirical coverage and inefficiency of SS, RCPS, RCPS-PPI, and RCPS-CPPI versus the number of labeled calibration samples $n$ for target risk $\alpha=0.1$ and confidence $\delta=0.1$ ($N=15650$, $K=5$).
• Figure 3: Empirical coverage and inefficiency of RCPS-CPPI as a function of the number of folds $K$, for $\alpha=0.1$ and $\delta=0.1$ ($N=15650$, $K=5$).
• Figure 4: Empirical coverage and inefficiency of SS, RCPS, RCPS-PPI, and RCPS-CPPI versus the validation MSE of the labeling predictor $(N=15650, n=200, K=5)$.

Theorems & Definitions (2)

• Theorem 1
• proof