Few-Shot Learning in Wi-Fi-Based Indoor Positioning.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Indoor positioning systems often struggle with limited labeled data.
• Convolutional Neural Networks (CNNs) are effective but data-hungry.
• Meta-learning offers a solution for data-scarce scenarios.

Purpose of the Study:

• To investigate the efficacy of few-shot learning combined with meta-learning for Wi-Fi-based indoor positioning.
• To enhance the accuracy and efficiency of positioning systems in data-limited environments.
• To compare the performance of base CNN models against meta-learning approaches.

Main Methods:

• Utilized few-shot learning tasks (e.g., N-shot) within a meta-learning framework.
• Applied CNNs for Wi-Fi signal processing and indoor localization.
• Evaluated model performance across various scenarios with differing data availability and novel classes.
• Employed filtering by cosine similarity (FCS) for data preprocessing and meta-learning stages.

Main Results:

• Base CNN accuracy varied significantly based on data samples per class (K) post-FCS.
• Meta-learning demonstrated acceptable performance in scenarios with limited samples, particularly for novel classes.
• With 20 samples/class, base CNN reached 0.80 pre-training accuracy; meta-learning (3-way 1-shot) achieved 0.78 on novel classes.

Conclusions:

• Few-shot learning with meta-learning effectively addresses data limitations in Wi-Fi indoor positioning.
• The proposed approach enhances model generalization for new, unseen indoor environments.
• Meta-learning provides a viable strategy for improving positioning accuracy when labeled data is scarce.