Deep Attention-based Sequential Ensemble Learning for BLE-Based Indoor Localization in Care Facilities

cs.LG cs.HC Minh Triet Pham, Quynh Chi Dang, Le Nhat Tan · Mar 22, 2026
Local to this browser
What it does
This paper addresses BLE-based indoor localization in care facilities by shifting from independent-window classification to sequential learning. The proposed DASEL framework combines frequency-based feature engineering, bidirectional GRUs...
Why it matters
Achieving a 53. 1% improvement over traditional baselines on the ABC 2026 challenge dataset, the work demonstrates that capturing temporal dependencies is critical for accurate indoor localization in complex real-world environments.
Main concern
The paper presents a technically sound and well-motivated approach that achieves substantial performance gains through sequential modeling. The DASEL framework effectively addresses RSSI instability via frequency-based features and handles...
Community signal
0
0 up · 0 down
Sign in to vote with arrows
AI Review AI reviewed
Plain-language introduction

This paper addresses BLE-based indoor localization in care facilities by shifting from independent-window classification to sequential learning. The proposed DASEL framework combines frequency-based feature engineering, bidirectional GRUs with attention mechanisms, and a two-level hierarchical ensemble to model temporal movement trajectories. Achieving a 53.1% improvement over traditional baselines on the ABC 2026 challenge dataset, the work demonstrates that capturing temporal dependencies is critical for accurate indoor localization in complex real-world environments.

Critical review
Verdict
Bottom line

The paper presents a technically sound and well-motivated approach that achieves substantial performance gains through sequential modeling. The DASEL framework effectively addresses RSSI instability via frequency-based features and handles unknown sequence boundaries during inference through multi-directional ensembles. The rigorous 4-fold temporal cross-validation demonstrates robustness across varying data conditions, including challenging train/test ratios. However, the evaluation is limited to a single caregiver over four days, and the computational cost of the ensemble approach (35 forward passes per timestamp) presents practical deployment challenges that the authors acknowledge but do not fully resolve.

“DASEL achieves a macro F1 score of 0.4438, representing a 53.1% improvement over the best traditional baseline (0.2898)”
paper · Abstract
“the two-level ensemble requires multiple forward passes per timestamp, which may challenge real-time deployment on resource-constrained mobile devices”
paper · Discussion
What holds up

The core insight—that traditional methods plateau due to the independence assumption—is convincingly demonstrated by the narrow performance range of baseline methods (0.2805–0.2898 macro F1) across different optimization strategies. The frequency-based feature representation provides device-agnostic robustness to RSSI fluctuations without requiring calibration, while the attention mechanism effectively weights informative timesteps during stable room occupancy versus noisy transitions. The two-level ensemble elegantly addresses the fundamental inference challenge of unknown sequence boundaries by combining multi-seed variance reduction with confidence-weighted aggregation across seven directional windows.

“all traditional methods remained within a narrow 0.2805 to 0.2898 range”
paper · Results
“DASEL reconceptualizes indoor localization as a sequential learning problem through four integrated phases”
paper · Methodology
Main concerns

The primary limitation is dataset scale and diversity: the evaluation relies on a single caregiver (User ID 90) over four days with severe class imbalance, raising significant questions about generalization to multi-user scenarios with varying movement patterns and device heterogeneity. The computational cost of the inference pipeline—requiring 5 models × 7 directional windows with multiple bidirectional GRU passes—may preclude real-time deployment on resource-constrained mobile devices, despite the paper's claim of practical deployability. Additionally, the baseline comparison omits modern sequential architectures (e.g., temporal CNNs, Transformers) that would better isolate the specific contributions of DASEL's design choices from the general benefits of sequential modeling.

“This step is essential to address data quality issues”
paper · Dataset
“Future work could explore model compression techniques such as knowledge distillation to maintain accuracy while reducing computational demands”
paper · Discussion
Evidence and comparison

The evidence strongly supports the superiority of sequential over independent-window modeling, with consistent improvements across all four folds (ranging from 22.7% to 70.0% relative gain). The temporal cross-validation protocol is methodologically sound, testing generalization across days with varying room distributions (12–18 rooms) and train/test ratios (0.79× to 31.4×). However, the baseline comparison is incomplete: while the authors compare against optimized XGBoost variants (including Garcia and Inoue's relabeling approach), they do not establish performance against comparable deep learning baselines such as unidirectional RNNs or CNNs without attention, making it difficult to attribute gains specifically to the bidirectional GRU + attention combination versus sequential modeling generally.

“DASEL achieves a 53.1% relative improvement over the best traditional baseline (Variation 1: 0.2898 → 0.4438)”
paper · Results
“Variation 2 applied signal pattern-based relabeling following Garcia and Inoue”
paper · Traditional Baseline
Reproducibility

The methodology is documented with sufficient architectural detail to permit replication, including specific hyperparameters (Bidirectional GRU layers with 128 and 64 units, dropout rates of 0.3), exact window configurations (10s and 15s lengths), and seed values [42, 1042, 2042, 3042, 4042]. The dataset originates from the ABC 2026 challenge and spans four days with approximately 1.1 million labeled samples. However, the paper lacks a code availability statement or links to repositories, and does not specify the deep learning framework used (PyTorch/TensorFlow). While preprocessing steps are clearly described, exact reproduction of the frequency-based features and temporal alignment procedures would benefit from open-source implementation, particularly given the complexity of the multi-level ensemble inference algorithm.

“We train 5 independent models with different random initialization seeds [42, 1042, 2042, 3042, 4042]”
paper · Methodology
“The final preprocessed dataset contains approximately 1.1 million labeled samples”
paper · Dataset
Abstract

Indoor localization systems in care facilities enable optimization of staff allocation, workload management, and quality of care delivery. Traditional machine learning approaches to Bluetooth Low Energy (BLE)-based localization treat each temporal measurement as an independent observation, fundamentally limiting their performance. To address this limitation, this paper introduces Deep Attention-based Sequential Ensemble Learning (DASEL), a novel framework that reconceptualizes indoor localization as a sequential learning problem. The framework integrates frequency-based feature engineering, bidirectional GRU networks with attention mechanisms, multi-directional sliding windows, and confidence-weighted temporal smoothing to capture human movement trajectories. Evaluated on real-world data from a care facility using 4-fold temporal cross-validation, DASEL achieves a macro F1 score of 0.4438, representing a 53.1% improvement over the best traditional baseline (0.2898).

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.