From Encounters to Guidance: Privacy-Preserving Risk Modeling on Temporal Contact Networks for Proactive Digital Contact Tracing

Abstract

COVID-19 highlighted the promise and limits of digital contact tracing. Decentralized deployments protected privacy, but alerts were reactive, binary, and often triggered broad quarantines. Network-based targeting can reduce and delay transmission by focusing on high-risk individuals, yet typical models require full visibility of the population contact network, which conflicts with privacy. This thesis investigates which benefits a network-based digital contact tracing policy can retain under a strictly privacy-preserving setup. It designs and evaluates a proactive scheme that combines realistic temporal contact networks with local risk signals and addresses three questions: how to synthesize realistic temporal networks at scale, what can be inferred about individual spreading potential under local visibility, and how effective a proactive local policy can be. <br/>

First, it introduces a pipeline that synthesizes temporal contact networks by calibrating human mobility models to empirical contact datasets with Bayesian optimization, and then embeds the calibrated models in an agent-based framework to generate multi-venue temporal networks at scale. The resulting networks reproduce structural and epidemic signatures and provide high-resolution testbeds for evaluation. <br/>

Second, the thesis quantifies what can be inferred about an individual's spreading potential from locally visible contacts. Across diverse networks, multiple vital node identification methods are benchmarked under <em>k</em>-hop visibility. Simple degree, which uses only minimal local information, already yields solid estimates of spreading potential. Modestly extending visibility beyond the immediate neighborhood lets more expressive methods approach the accuracy of full-network information. With modest information sharing, lightweight machine learning models that aggregate local neighbor features add gains without sharing raw topology. <br/>

Finally, these insights are operationalized as <em>Network-based Proactive Contact Tracing</em>. In simulations on the generated and empirical networks, individuals map recent encounter intensity to a spreading potential-based risk score and compare it to a single epidemic-aware threshold. Exceeding it triggers a graded intervention that reduces future contacts. Evaluation shows infection peak reductions up to 40% while suppressing about 20% of contacts. It outperforms non-adaptive baselines and is robust to parameter choices. <br/>

The contributions are: 1) a calibrated generator for realistic multi-venue temporal contact networks; 2) a quantified privacy-accuracy trade-off for ego-bounded risk estimation; and 3) a privacy-preserving, proactive scheme with measured cost-benefit, robustness, and distributional properties. Together, these results provide methods, metrics, and empirical evidence to guide the design and evaluation of alternative digital contact tracing policies based on realistic temporal contact structures and local, privacy-preserving risk estimation.