[Paper Review] Transfer Learning-Based Outdoor Position Recovery with Telco Data

Yige Zhang, Aaron Yi Ding+5 more|arXiv (Cornell University)|Dec 10, 2019

Indoor and Outdoor Localization Technologies35 references4 citations

TL;DR

This paper proposes TLoc, a transfer learning-based framework for outdoor mobile device localization using telecommunication (Telco) measurement report (MR) data. By leveraging a relative coordinate space, domain similarity measurement, and structured transfer learning, TLoc transfers accurate localization models from data-rich source domains to data-scarce target domains, reducing median localization error by up to 49.22% compared to state-of-the-art methods on 2G and 4G datasets in Shanghai.

ABSTRACT

Telecommunication (Telco) outdoor position recovery aims to localize outdoor mobile devices by leveraging measurement report (MR) data. Unfortunately, Telco position recovery requires sufficient amount of MR samples across different areas and suffers from high data collection cost. For an area with scarce MR samples, it is hard to achieve good accuracy. In this paper, by leveraging the recently developed transfer learning techniques, we design a novel Telco position recovery framework, called TLoc, to transfer good models in the carefully selected source domains (those fine-grained small subareas) to a target one which originally suffers from poor localization accuracy. Specifically, TLoc introduces three dedicated components: 1) a new coordinate space to divide an area of interest into smaller domains, 2) a similarity measurement to select best source domains, and 3) an adaptation of an existing transfer learning approach. To the best of our knowledge, TLoc is the first framework that demonstrates the efficacy of applying transfer learning in the Telco outdoor position recovery. To exemplify, on the 2G GSM and 4G LTE MR datasets in Shanghai, TLoc outperforms a nontransfer approach by 27.58% and 26.12% less median errors, and further leads to 47.77% and 49.22% less median errors than a recent fingerprinting approach NBL.

Motivation & Objective

• Address the challenge of poor localization accuracy in outdoor Telco networks due to insufficient geo-tagged measurement report (MR) samples in certain urban areas.
• Overcome the high data collection cost associated with acquiring sufficient MR samples for training accurate localization models.
• Enable knowledge transfer from data-rich source domains to data-scarce target domains to improve localization performance without retraining from scratch.
• Design a novel coordinate space and similarity metric to support effective transfer learning in outdoor Telco localization.

Proposed method

• Introduce a relative coordinate space centered on a serving base station to enable consistent representation across domains, facilitating knowledge transfer.
• Propose a domain similarity metric based on signal strength patterns and geographic proximity to identify the most relevant source domains for transfer.
• Adapt structured transfer learning (STL) to fine-tune pre-trained Random Forest regression models by re-estimating decision tree node thresholds using target domain data.
• Divide the area of interest into fine-grained subareas (domains) to enable localized modeling and targeted transfer learning.
• Use a grid-based representation (parameterized by g) to discretize domains, balancing spatial resolution and model generalization.
• Apply a RaF-based regression model to learn the mapping from MR features (e.g., signal strength) to geographic positions within each domain.

Experimental results

Research questions

• RQ1 Can transfer learning effectively improve outdoor Telco position recovery in areas with limited geo-tagged MR samples?
• RQ2 How can a consistent coordinate space be designed to enable knowledge transfer across different geographic domains in outdoor Telco localization?
• RQ3 What criteria can be used to select the most suitable source domains for transfer learning in a heterogeneous urban environment?
• RQ4 To what extent does transfer learning reduce localization error compared to non-transfer and fingerprinting-based baselines?

Key findings

• TLoc reduces median localization error by 27.58% and 26.12% compared to a non-transfer learning baseline on 2G GSM and 4G LTE datasets in Shanghai, respectively.
• TLoc achieves 47.77% and 49.22% lower median error than the recent fingerprinting approach NBL on the 2G and 4G datasets, respectively.
• The proposed domain similarity metric effectively identifies relevant source domains, improving transfer performance across diverse urban subareas.
• TLoc demonstrates robustness to base station relocations and hardware upgrades by enabling model adaptation through transfer learning.
• The framework maintains strong performance even in rural or low-signal-density areas where baseline methods fail due to insufficient training data.
• Visualization results show that TLoc-generated trajectories closely follow road segments, while non-transfer methods exhibit significant positional drift.