Distributed Representations of Wi-Fi Fingerprints from Non-Contextual Text-Embedding Techniques with Applications in Crowdsourcing Zone-Level Localization

Abstract

Over the past decade, indoor localization systems have gained increasing attention and found widespread applications in commercial and research environments. Specifically, a Wi-Fi fingerprint-based system offers a low-cost solution over its counterparts such as Bluetooth, ultra-wideband (UWB), and radio frequency identification (RFID) technologies due to the ubiquity of Wi-Fi access points (WAPs) in most buildings. However, the main disadvantage of the fingerprint-based system is intensive survey effort required during system initialization and maintenance. This work explores a solution to alleviate this limitation by considering a crowdsourcing approach for zone-level localization.  Instead of relying only on the labelled fingerprint data from trained surveyors, this approach uses the more-attainable unlabelled fingerprint data collected by participating volunteers. This unlabelled data is then used to augment the survey data in a process called pseudo labelling, forming a more comprehensive training dataset for subsequent localization tasks; this semi-supervised approach allows for minimal survey effort during system initialization and maintenance.  To enable such solution, this work introduces a novel approach of employing non-contextual word embedding techniques to construct distributed vector representations of fingerprint data to overcome 3 challenges; (a) high memory requirement in the downstream tasks due to high-dimensional non-distributed vector representations from the “standard” vector transformation, (b) inclusion of an arbitrary value that represents missing WAPs which can affect the performance of the downstream localization tasks in a non-transparent manner, and most importantly, (c) poor pseudo-labelling and semi-supervised zone-prediction performances due to poor data separability in a feature space. The choice of the non-contextual text-embedding techniques, as opposed to the contextual counterparts, leads to less computational requirement in model training and distributed-representation generation due to simpler model architectures (no deep learning) and no requirement for pre-trained model during distributed-representation generation. To this end, we considered non-contextual word embedding techniques commonly used in natural language processing such as Word2Vec, GloVe, and Doc2Vec in the distributed-representation transformation, and compared the resulting downstream performances with those from well-recognized dimensionality reduction techniques such as PCA, Isomap, and UMAP.  The results show that Word2Vec and GloVe transformations outperform other types of transformations in terms of separability in fingerprint representations, pseudo-labelling performance, and semi-supervised zone-prediction accuracy. Together with the promising robustness property against potential data inhomogeneity, Word2Vec and GloVe transformations are the recommended transformation processes for constructing vector representations of fingerprints in crowdsourcing zone-level localization.

HIGHLIGHTS

• This work introduces a novel approach of employing non-contexual word-embedding techniques to construct distributed vector representations of Wi-Fi fingerprint data to facilitate pseudo-labelling and semi-supervised zone-prediction tasks in crowdsourcing zone-level localization
• The benefits of employing word-embedding techniques are (a) lower memory requirement in the downstream tasks due to distributed vector representations (b) no inclusion of an arbitrary value that represents missing WAPs which can affect the performance of the downstream localization tasks in a non-transparent manner (c) improved pseudo-labelling and semi-supervised zone-prediction performances due to improved data separability in a feature space
• The benefit of employing non-contextual techniques, as opposed to the contextual counterparts, is less computational requirement in model training and distributed-representation generation due to simpler model architectures (no deep learning) and no requirement for pre-trained model during distributed-representation generation
• The results show that Word2Vec and GloVe transformations outperform other types of transformations in terms of separability in fingerprint representations, pseudo-labelling performance, and semi-supervised zone-prediction accuracy

GRAPHICAL ABSTRACT