Validation Safety Gate Rejects Incompatible Cross Building Wi-Fi Localization Models

Indoor positioning models that attempt to transfer Wi-Fi fingerprint knowledge across different buildings frequently degrade localization performance, according to an extensive multi-building evaluation of a new framework named SafeTransLoc. Indoor Wi-Fi fingerprinting works by mapping received signal strength vectors to physical coordinates, but acquiring these calibration fingerprints across multi-story buildings requires considerable manual survey time. Researchers have previously attempted to reuse historical signal maps through transfer learning and domain adaptation to avoid recalibrating new facilities from scratch. However, transferring signal models across separate structures poses severe technical challenges because target buildings generally feature unrelated access point identifiers, disparate physical layouts, distinct coordinate reference systems, and altered radio frequency propagation properties. Under these incompatible conditions, attempting to force domain invariance between source and target environments can create feature representations that appear aligned while corrupting physical coordinate predictions, a failure mode known as negative transfer.

To prevent this performance degradation, researchers designed SafeTransLoc as an optional deployment assurance layer rather than a universal transfer-learning mechanism. Instead of deploying adapted models automatically, the framework treats the decision to transfer external knowledge as an empirical hypothesis that must be proved against a target-only baseline. SafeTransLoc partitions available target calibration measurements into disjoint adaptation and validation subsets. Separate candidate models are trained exclusively on the adaptation subset, including both target-only architectures and transfer models adapted from source buildings. The framework then compares the best target-only candidate against the best transfer candidate using mean Euclidean positioning error on held-out target validation fingerprints. External transfer is accepted for operational deployment only if its validation accuracy surpasses the best target-only alternative by at least a predefined non-negative safety margin; otherwise, the system rejects transfer and deploys the target-only estimator.

The evaluation examined localization accuracy across two widely studied multi-building datasets, UJIIndoorLoc and SODIndoorLoc, spanning 120 cross-building configurations. UJIIndoorLoc encompasses three multi-floor buildings sharing a common 520-column wireless access point feature structure, where one building serves as the target domain and the remaining two supply source fingerprints. In contrast, SODIndoorLoc comprises three facilities—CETC331, HCXY, and SYL—that possess completely independent access point hardware inventories containing 52, 347, and 363 access points, respectively. To enable neural network transfer across environments lacking shared access point identifiers, the researchers introduced a 71-dimensional access-point-independent descriptor. This standardized representation clips received signal strength values between minus 110 and 0 dBm, normalizes them to a unit scale, selects the 32 strongest normalized signals, extracts adjacent signal strength differences, and appends eight statistical metrics including visible access point ratios, signal entropy, and signal quartiles.

Across the principal cross-building experiments, the candidate models included both parametric and non-parametric learning strategies evaluated under sparse target supervision levels of 1, 5, 10, and 20 percent of available calibration data. The target-only candidates comprised Cosine nearest neighbors, which computes weighted coordinate averages directly in the target building access point feature space using cosine similarity, and Target-Neural, a fully connected neural regressor trained solely on target adaptation fingerprints via the 71-dimensional descriptor. The transfer candidates shared the identical neural architecture—featuring a two-layer encoder with 192 hidden units per layer, layer normalization, dropout, and a two-dimensional coordinate regressor—and included Pooled-FineTune, CORAL-FineTune with latent covariance alignment, and DANN-FineTune utilizing adversarial domain classification through a gradient-reversal layer. Every principal configuration was repeated across five independent random seeds controlling data splitting, neural weight initialization, and batch ordering.

In the primary 120 cross-building scenarios, the evaluation revealed that every transfer candidate performed substantially worse than the strongest target-only alternative on the untouched official test partitions. On UJIIndoorLoc, Cosine nearest neighbors delivered the lowest mean positioning error across all label ratios, recording an error of 13.846 meters at the 1 percent supervision ratio compared to 42.623 meters for conventional source transfer, representing a 28.777-meter degradation. While providing 20 percent target calibration reduced the error gap to 10.207 meters, source-target mismatch remained substantial. On SODIndoorLoc, Cosine nearest neighbors achieved a mean error of 5.586 meters at 1 percent labels and 3.504 meters at 20 percent labels, whereas neural transfer candidates produced errors ranging between 20.410 and 20.961 meters at 1 percent labels and between 16.193 and 17.255 meters at 20 percent labels.

Because the best transfer model proved less accurate than the strongest target-only model in all 120 cross-building trials, negative transfer was pervasive at the deployment level. Under these severe mismatch conditions, SafeTransLoc successfully protected positioning accuracy by rejecting transfer in all 60 UJIIndoorLoc configurations and maintaining identical performance to Cosine nearest neighbors with zero gate regret. On SODIndoorLoc, operating the gate without a safety margin produced a single false acceptance at the 1 percent label ratio for environment CETC331 under random seed 55, where adversarial alignment appeared 0.335 meters superior on just three validation fingerprints but increased test error by 4.733 meters. Introducing a 0.5-meter safety margin completely eliminated this erroneous selection, yielding 100 percent harmful-transfer rejection and reducing observed gate regret to zero across the entire original dataset collection.

The study also examined the performance of the neural models in an architecture-matched comparison to isolate the specific impact of source-domain pretraining from representation loss. Because Cosine nearest neighbors retains raw, target-specific access point identities while Target-Neural relies on the 71-dimensional access-point-independent descriptor, comparing nearest neighbors directly with neural models conflates learning strategy with feature representation. When compared strictly against Target-Neural, source transfer provided modest localized benefits under the sparsest target supervision: at the 1 percent label ratio, pooled pretraining, covariance alignment, and adversarial alignment reduced mean error relative to Target-Neural by 1.269 meters, 0.718 meters, and 1.090 meters, respectively. However, as target supervision increased to 10 and 20 percent, all three transfer strategies became less accurate than Target-Neural by up to 2.437 meters, demonstrating that transfer learning can assist an architecture-matched neural model under severe data scarcity while still failing to match non-parametric target-only alternatives.

To evaluate gate behavior in scenarios where transfer is genuinely advantageous, the authors conducted a controlled same-environment domain-shift experiment across CETC331, HCXY, and SYL encompassing 270 configurations. In these experiments, 60 percent of available training data served as a source domain, while sparse target sets of 1, 5, and 10 percent were subjected to bounded signal offsets of plus or minus 2, 4, and 6 dB or Gaussian noise with standard deviations of 1, 2, and 4 dB. Across these 270 test configurations, transfer genuinely outperformed target-only estimators in exactly 6 configurations—all occurring in CETC331 at 1 percent labels under seed 22—while proving harmful in the remaining 264 configurations. Operating SafeTransLoc with a zero margin accepted transfer in 30 configurations, capturing 1 beneficial transfer alongside 29 harmful cases. Increasing the margin to 2 meters reduced the false acceptance rate from 11.0 percent to 5.7 percent and reduced mean gate regret from 0.599 meters to 0.389 meters, but suppressed acceptance of all beneficial cases, illustrating a fundamental safety-sensitivity trade-off.

An independent leave-one-environment-out margin calibration on archived SODIndoorLoc data further demonstrated the operational nuances of the safety threshold. When calibrating the margin on two development environments and locking the threshold before testing on the third held-out environment, the optimization rule selected a 0-meter margin for CETC331 and a 0.5-meter margin for HCXY and SYL. Across the 60 held-out evaluations, this independent calibration produced one false acceptance, achieved 98.33 percent overall selection accuracy, and limited mean gate regret to 0.078883 meters. In statistical assessments using paired Wilcoxon signed-rank tests with Holm correction, the performance differences between the 0-meter and 2-meter margins as well as between the 0.5-meter and 2-meter margins remained statistically significant, confirming that decision threshold adjustments exert a measurable impact on operational positioning error independent of model retraining.

System robustness was additionally assessed by simulating inference-time hardware degradation through the progressive removal of visible access point observations at rates between 10 and 40 percent without modifying the trained models or gate decisions. As missing access points increased to 40 percent, mean positioning error degraded gradually from approximately 10.5 meters to 14.2 meters, demonstrating that cosine similarity maintains stable neighborhood relationships under random signal loss. The authors emphasized several key limitations, noting that the empirical standard deviations quantify repeatability across sample splits and neural seeds rather than comprehensive metrological uncertainty from physical device differences or survey inaccuracies. Furthermore, conservative validation gating faces inherent sample-size constraints under sparse calibration, and the final refitting of the chosen model architecture on pooled adaptation and validation data creates a subtle model-selection mismatch that warrants tracking in future adaptive localization systems.

Subject of Research: Validation-gated estimator selection for indoor Wi-Fi localization

Article Title: Preventing negative transfer in cross-building Wi-Fi fingerprint localization via validation-gated estimator selection

Article References: Ahmad Sa’ahiry, A. H., Wang, Q., Yu, H., See, C. H., & Tran, N.-T. (2026). Preventing negative transfer in cross-building Wi-Fi fingerprint localization via validation-gated estimator selection. Discover Electronics, 3(1), Article 166. https://doi.org/10.1007/s44291-026-00318-4

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00318-4

Keywords: indoor localization, Wi-Fi fingerprinting, transfer learning, negative transfer, domain adaptation, estimator selection, validation gating, SafeTransLoc, received signal strength, UJIIndoorLoc, SODIndoorLoc, cosine similarity, nearest neighbor, deep learning, covariance alignment, adversarial training, model selection, safety margin