Computer Science

The Privacy Sandbox aims "to create technologies that both protect people's privacy online and give companies and developers tools to build thriving digital businesses." This talk will describe some of the design, implementation and practical challenges in evolving measurement solutions away from persistent cross-site identifiers.

Abstract This work connects models for virus spread on networks with their equivalent neural network representations. Based on this connection, we propose a new neural network architecture, called Transmission Neural Networks (TransNNs) where activation functions are primarily associated with links and are allowed to have different activation levels. This connection also leads to the discovery and the derivation of three new activation functions with tunable or trainable parameters. We show that TransNNs with a single hidden layer and a fixed non-zero bias term are universal function approximators. Moreover, we establish threshold conditions for virus spread on networks where the dynamics are characterized by TransNNs. Finally, we present new derivations of continuous time epidemic models on networks based on TransNNs. *This is joint work with Peter E. Caines.

Abstract Dengue virus is the most significant viral mosquito-borne infection in terms of its human impact. Mathematical modeling has contributed to our understanding of its transmission and control strategies aimed at halting its spread. We consider the spread of dengue at the level of a city. Because the Aedes aegypti mosquito that transmits dengue has relatively low dispersal over its lifetime, human movement plays a major role in its spread and the household is a key spatial scale on which transmission occurs. Simple multi-patch deterministic models---metapopulation models, which consider the population to be described as a network of well-mixed patches---have been used to model city-level spatial spread and can provide expressions for key epidemiological quantities such as the basic reproduction number, $R_0$. We compare dynamics predicted by such models with results from individual-based network models and illustrate several discrepancies. We argue that the small size of households and local depletion of susceptibles are key features of the dynamics that are not captured in the standard $R_0$ analysis of the ODE model. In order to gain analytic understanding, we propose the use of household-level models, which can be analyzed using branching process theory. Our work, which echoes results previously found for directly-transmitted infections, highlights the importance of correctly accounting for the relevant spatial scales on which transmission occurs

Abstract The ongoing COVID-19 pandemic has had devastating impacts on global public health and socioeconomic stability. Although highly efficacious COVID-19 vaccines were developed at an unprecedented rate, the ongoing evolution of SARS-CoV-2 and consequential changes in infectivity and immunological resistance of new variants continues to present challenges. Computing the growth rates of emerging variants is complicated by many issues, including vaccine uptake, regional levels of prior infection, viral resistance to protective antibodies, and the relative infectivity of new variants in complex populations. While epidemic forecasting has played an important role in decision-making, forecast accuracy has been limited, especially at key tipping points in the pandemic, by the inability to incorporate important factors, such as the emergence of phenotypically novel variants. In this talk, I will describe a flexible strategy to characterize variant transition dynamics through three simple summaries, the speed, the relative timing, and the magnitude of the variant transition. This foundational research is intended to better understand the implication of SARS-CoV-2 evolution to ultimately inform regional epidemiological forecasting.  

Abstract Online social networks often mirror community formation in real-world networks (based on common demographics, interests, or affinities). Such patterns are often picked up and used by algorithms that leverage social data for the purpose of providing recommendations, diffusing information, or forming groups. In this talk, we'll discuss the influence maximization problem where multiple communities exist, showing that common centrality metrics may exclude minority communities from adopting the information diffused. Using the preferential attachment model with unequal communities, we'll characterize the relationship between homophily, network centrality, and bias through the power-law degree distributions of the nodes, and study the conditions in which diversity interventions can actually yield more efficient and equitable outcomes. We find a theoretical condition on the seedset size that explains the potential trade-off between outreach and diversity in information diffusion. To wrap up, we’ll discuss a novel set of algorithms that leverage the network structure to maximize the diffusion of a message while not creating disparate impact among participants based on community affiliation.

Abstract We study how communication platforms can improve social learning without censoring or fact-checking messages, when they have members who deliberately and/or inadvertently distort information. Message fidelity depends on social network depth (how many times information can be relayed) and breadth (the number of others with whom a typical user shares information). We characterize how the expected number of true minus false messages depends on breadth and depth of the network and the noise structure. Message fidelity can be improved by capping depth or, if that is not possible, limiting breadth, e.g., by capping the number of people to whom someone can forward a given message. Although caps reduce total communication, they increase the fraction of received messages that have traveled shorter distances and have had less opportunity to be altered, thereby increasing the signal-to-noise ratio.  

Abstract Identifying the optimal set of individuals to first receive information (‘seeds’) in a social network is a widely-studied question in many settings, such as diffusion of information, spread of microfinance programs, and adoption of new technologies. Numerous studies have proposed various network-centrality based heuristics to choose seeds in a way that is likely to boost diffusion. Here we show that, for the classic SIR model of diffusion and some of its generalizations, randomly seeding s + x individuals can prompt a larger diffusion than optimally targeting the best s individuals, for a small x. We prove our results for large classes of random networks, and verify them in several small, real-world networks. Our results identify practically relevant settings under which collecting and analyzing network data to boost diffusion is not cost-effective.

Abstract   This talk will revisit a 2011 Review of Economic Studies paper written with Daron Acemoglu, Munther Dahleh and Asuman Ozdaglar. We consider the canonical social learning model but where observations of past actions are constrained by a social network. The network is generated stochastically and neighborhoods can have arbitrary distributions. We are interested in what kinds of networks and signal structures lead to asymptotic learning (convergence in probability to the correct action). We prove a necessary and sufficient condition for asymptotic learning if signals are of unbounded strength, as well as network properties that allow learning irrespective of the signal structure.  

Abstract Due to noisy data and nonlinear dynamics, even simple stochastic epidemic models such as the Susceptible-Infectious-Removed (SIR) present significant challenges to inference. In particular, computing the marginal likelihood of such stochastic processes conditioned on observed endpoints a notoriously difficult task. As a result, likelihood-based inference is typically considered intractable in missing data settings typical of observational data, and practitioners often resort to intensive simulation methods or approximations. We discuss recent contributions that enable "exact" inference, focusing on a perspective that makes use of latent variables to explore configurations of the missing data within a Markov chain Monte Carlo framework. Motivated both by count data from large outbreaks and high-resolution contact data from mobile health studies, we show how our data-augmented approach successfully learns the interpretable epidemic parameters and scales to handle large realistic data settings efficiently.

Abstract In this talk, we present a modeling framework to study the effects of testing policy on voluntary social distancing and the spread of an infection. Agents decide their social activity level, which determines the social network over which the virus spreads. Testing enables the isolation of infected individuals, slowing down the infection. But greater testing also reduces voluntary social distancing or increases social activity, exacerbating the spread of the virus. We show that the effect of testing on infections is non-monotone. This non-monotonicity also implies that the optimal testing policy may leave some of the testing capacity of society unused. This also implies that testing should be combined with mandatory social distancing measures to avoid these adverse behavioral effects.

A complex contagion is an infectious process in which individuals may require multiple transmissions before changing state. These are used to model behaviours if an individual only adopts a particular behaviour after perceiving a consensus among others. We may think of individuals as beginning inactive and becoming active once they are contacted by a sufficient number of active partners.  Here we study the dynamics of the Watts threshold model (WTM).  We adapt techniques developed for infectious disease modelling to develop an analyse analytic models for the dynamics of the WTM in configuration model networks and a class of random clustered (triangle-based) networks.  We derive conditions under which cascades happen with an arbitrarily small initial proportion active. We also observe hybrid phase transitions when cascades are not possible for small initial conditions, but occur for large enough initial conditions.

Abstract In the talk I will briefly outline the idea of the so-called dynamical survival analysis (DSA) which uses survival analysis methods to build approximate models of individual level epidemic dynamics by utilizing some well known mean-field approximations. I will show the DSA connection with classical agent based models for epidemics and also some frailty models that have been successfully applied to recent COVID-19 epidemic.