Conformal Prediction for Indoor Positioning with Correctness Coverage Guarantees

Zhiyi Zhou, Hexin Peng, Hongyu Long · 2025

Read Conformal Prediction for Indoor Positioning with Correctness Coverage Guarantees by Zhiyi Zhou, Hexin Peng, Hongyu Long (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-01810.txt.

Introduction

Finding your place inside a large building, where satellite signals fail, often relies on matching Wi-Fi fingerprints to a map. Zhiyi Zhou and two co-authors add a safety layer to that task in 2025 using conformal prediction. The method turns a model's guess into a set of places with a stated chance of holding the true one.

They also extend the idea to route guidance, keeping error rates under control along a path. On a public indoor data set, their models reached about eighty-five percent accuracy on unseen data and hit their coverage targets closely. Set sizes varied by model, showing the price of certainty. The authors say such guarantees matter for services people trust with their location.

Catalog record

CREATORZhiyi Zhou, Hexin Peng, Hongyu Long
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-01810

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Conformal Prediction for Indoor Positioning with Correctness Coverage Guarantees Zhiyi Zhou, Hexin Peng, Hongyu Long Abstract With the advancement of Internet of Things (IoT) technologies, high-precision indoor positioning has become essential for Location-Based Services (LBS) in complex indoor environments. Fingerprint-based localization is popular, but traditional algorithms and deep learning-based methods face challenges such as poor generalization, overfitting, and lack of interpretability. This paper applies conformal prediction (CP) to deep learning-based indoor positioning. CP transforms the uncertainty of the model into a non-conformity score, constructs prediction sets to ensure correctness coverage, and provides statistical guarantees. We also introduce conformal risk control for path navigation tasks to manage the false discovery rate (FDR) and the false negative rate (FNR).The model achieved an accuracy of approximately 100% on the training dataset and 85% on the testing dataset, effectively demonstrating its performance and generalization capability. Furthermore, we also develop a conformal p-value framework to control the proportion of position-error points. Experiments on the UJIIndoLoc dataset using lightweight models such as MobileNetV1, VGG19, MobileNetV2, ResNet50, and EfficientNet show that the conformal prediction technique can effectively approximate the target coverage, and different models have different performance in terms of prediction set size and uncertainty quantification. 1 Introduction The development of IoT technologies has led to the widespread use of Location-Based Services (LBS) in numerous applications, including smart cities. However, traditional Global Positioning System (GPS) performance is severely affected in environments with signal blockages, such as tunnels, large petrochemical facilities, underground mines, airports, and shopping centers, due to signal weakening and multipath interference Farahsari et al. (2022) ; Wahab et al. (2022) . As a result, high-precision indoor positioning technologies have been developed and extensively studied. Despite considerable progress, the inherent complexity and variability of indoor environments present substantial challenges in achieving accurate, cost-effective, real-time, and robust indoor positioning systems. Fingerprint-based positioning has become a leading approach to indoor positioning due to its low hardware demands and ability to adapt to challenging environments Zhu et al. (2020) . By collecting spatial signal characteristics (e.g., WiFi, geomagnetic, Bluetooth) and constructing a fingerprint database, this technique enables single-point positioning and path matching with high accuracy in challenging environments. However, traditional fingerprint-matching algorithms exhibit significant limitations in high-dimensional feature extraction, complex pattern analysis, and adaptability to dynamic environmental changes, which hinder their performance in real-world applications. Despite advancements in fingerprint resolution and data scale, improvements in indoor positioning accuracy remain constrained Alhomayani and Mahoor (2020) . When applying deep learning models for indoor positioning Sun et al. (2024) , although they demonstrate strong performance in certain scenarios, several critical challenges remain that significantly impact real-world deployments. Deep learning models generally rely on extensive and high-quality datasets for training. Insufficient data may impair the model’s generalization ability during testing, leading to inferior accuracy compared to less complex indoor positioning algorithms Alzubaidi et al. (2023) . Moreover, these models are particularly vulnerable to environmental dynamics - even minor changes in indoor layouts, signal propagation patterns due to furniture rearrangements, or the introduction of new electronic devices can cause dramatic performance degradation. For instance, in areas with signal fluctuations or at boundaries between different signal coverage zones, deep learning models often produce high-confidence but inaccurate predictions, creating a dangerous false sense of reliability Corbière et al. (2019) . Additionally, when deployed on new devices with different sensors or antenna characteristics, these models may produce unpredictable position estimates without providing any indication of increased uncertainty Yang et al. (2024) . The overfitting issue in deep learning models, particularly in complex architectures like Convolutional Neural Networks (CNNs), leads to excellent performance on training data while resulting in substantial errors in novel environments, which can cause misleading conclusions or positioning errors with no warning mechanism Nessa et al. (2020) ; Khan et al. (2024) . Furthermore, the inherent "black-box" nature of these models severely limits error diagnosis and accountability Guidotti et al. (2018) , creating significant barriers for safety-critical applications where positioning reliability must be guaranteed. These practical limitations demand a systematic approach to uncertainty quantification that can provide statistical guarantees regardless of the underlying model architecture or data distribution. To address these fundamental challenges, we employ conformal prediction (CP) Wang et al. (2024) ; Wang et al. (2025a) as our solution framework, a choice driven by its unique properties that align perfectly with the requirements of robust indoor positioning. Unlike other uncertainty quantification methods, such as Bayesian neural networks (which require specific model architectures) or ensemble methods (which impose significant computational overhead), CP offers model-agnostic and distribution-free guarantees under the mild assumption of data exchangeability. This property is particularly valuable in indoor positioning where signal distributions are complex and difficult to model parametrically. CP converts the intuitive concept of uncertainty into a statistically rigorous measure called the non-conformity score, providing mathematically guaranteed coverage rates regardless of the underlying positioning model or signal distribution characteristics Wang et al. (2025b) . In this work, we focus primarily on single-point positioning as our core research problem while extending our framework to path navigation as a practical application. For single-point positioning, we convert the average positioning error into a non-conformity score, set a quantile threshold by defining a desired error rate, and construct a prediction set based on this threshold. This approach provides a statistically guaranteed solution where the prediction set contains the true position with a pre-specified probability (e.g., 95%), allowing system designers to make informed reliability-accuracy tradeoffs. The size of the prediction set becomes an interpretable measure of positioning difficulty and uncertainty in different environments. For the extended application in path navigation, we introduce conformal risk control to provide strict statistical guarantees on FDR and FNR, offering systematic control over navigation errors. This addresses the critical challenge of maintaining reliable guidance in areas with varying signal quality or environmental complexity. Finally, we propose a conformal p-value framework to control the proportion of position-error points in path navigation, ensuring that the probability of misleading guidance remains bounded by a user-specified significance level. Our contributions can be summarized as follows: • We propose a novel uncertainty-aware indoor positioning framework that integrates conformal prediction with deep learning models, achieving up to 15% improvement in positioning reliability compared to traditional uncertainty quantification methods while maintaining the same accuracy. Our framework is the first to provide statistically guaranteed coverage for indoor positioning tasks, enabling system designers to make precise reliability-accuracy trade-offs that were previously impossible. • We propose a new conformal risk control mechanism specifically designed for path navigation scenarios that provides explicit control over false discovery and false negative rates. Unlike existing navigation confidence methods that offer only heuristic uncertainty estimates, our approach offers mathematically rigorous guarantees, reducing misleading navigation instructions by 23% in our experiments while maintaining path completeness. • We propose an innovative conformal p-value framework for position-error control that allows for dynamic reliability assessment of individual positioning points. This framework outperforms traditional confidence scoring methods by providing calibrated p-values that directly correspond to error probabilities, enabling applications to maintain a consistent error rate across varying environmental conditions. 2 Related Work 2.1 Indoor Positioning Technology Currently, indoor positioning technologies are primarily classified into two categories: signal measurement-based methods and fingerprint matching-based methods ( Alsamhi et al., 2023 ) .

Read the complete text