A Simplistic Machine Learning Approach to Contact Tracing

TL;DR

This paper presents a simple machine learning method using handcrafted features and models like GBM and MLP to estimate phone-to-phone distance, outperforming previous challenge results.

Contribution

It introduces a novel approach that excludes time-based criteria and demonstrates superior accuracy in contact distance estimation.

Findings

Outperforms HKUST challenge results significantly

Uses handcrafted features from phone data

Employs GBM and MLP models effectively

Abstract

This report is based on the modified NIST challenge, Too Close For Too Long, provided by the SFI Centre for Machine Learning (ML-Labs). The modified challenge excludes the time calculation (too long) aspect. By handcrafting features from phone instrumental data we develop two machine learning models, a GBM and an MLP, to estimate distance between two phones. Our method is able to outperform the leading NIST challenge result by the Hong Kong University of Science and Technology (HKUST) by a significant margin.