Empirical networks for localized COVID-19 interventions using WiFi infrastructure at university campuses

Infectious diseases, like COVID-19, pose serious challenges to university campuses, which typically adopt closure as a non-pharmaceutical intervention to control spread and ensure a gradual return to normalcy. Intervention policies, such as remote instruction (RI) where large classes are offered online, reduce potential contact but also have broad side-effects on campus by hampering the local economy, students’ learning outcomes, and community wellbeing. In this paper, we demonstrate that university policymakers can mitigate these tradeoffs by leveraging anonymized data from their WiFi infrastructure to learn community mobility—a methodology we refer to as WiFi mobility models (WiMob). This approach enables policymakers to explore more granular policies like localized closures (LC). WiMob can construct contact networks that capture behavior in various spaces, highlighting new potential transmission pathways and temporal variation in contact behavior. Additionally, WiMob enables us to design LC policies that close super-spreader locations on campus. By simulating disease spread with contact networks from WiMob, we find that LC maintains the same reduction in cumulative infections as RI while showing greater reduction in peak infections and internal transmission. Moreover, LC reduces campus burden by closing fewer locations, forcing fewer students into completely online schedules, and requiring no additional isolation. WiMob can empower universities to conceive and assess a variety of closure policies to prevent future outbreaks.

WIMOB is our approach to describe contact between people and movement of people between locations. The first step requires using WiFi network logs to infer when individuals were at specific locations on campus by determining when devices were connected to the corresponding APs. Our system mines the WiFi network logs that are populated via the Simple Network Management Protocol (SNMP) -a standard and widely used monitoring protocol to organize device association behavior to a WiFi network. Periodic SNMP updates can be caused either by poll requests to the APs that log which devices are associated with it at that time. However, devices can appear invisible to detached from an AP for multiple reasons, for example, when devices are idle. Otherwise. SNMP updates can occur whenever a new device connects, which is typical when individuals move between APs. Our approach exploits this factor to first mine periods when individuals are moving, then identify periods of dwelling between movements, and finally determine collocation when two or more individuals are dwelling near the same AP. This system follows from other studies that mine WiFi logs [12,36] and the detailed processing pipeline and evaluation is presented in [8]. This system to infer collocations has been tested against lecture attendance and reports a high precision of 0.89, but a relatively lower specificity of 0.79 [8]. While it is not likely to show false-positives, it has a possibility to erroneously mark people absent from a location even though they were there. However, for the purposes of our study, a contact network is made over an entire day and it only needs a single collocation instance for us to consider contact. And therefore we believe this limitation would not significantly affect our models.

Characterizing Logs as Contact and Movement Networks
After inferring where an individual is located on campus, we represent the entire community behavior as graphs. We describe a bipartite graph, K, that shows when a user is at a given location on campus ( Figure Figure S1). This bipartite graph has edges connecting a set of m people, P , to a set of n locations, L. An individual can have multiple edges connecting to the location if they visited that location multiple times (e.g., t 1 , t 2 ). The edge data contains the start and end times of these dwelling periods. For these bipartite graphs, we make a projection on set P to describe collocation. This projection graph, G, contains an edge between users if they were visiting the same location during overlapping times. Since we do 2

Modeling Policies and Scenarios
not use RTLS, our approach can only identify if people were in the vicinity of the same AP, but does not describe the distance between them. However, it can reasonably determine collocation in the same room [8]. Since our study is limited to localizing people indoors, we adapt the definition of proximate contact [17] where people might be "more than 6 feet but in the same room for an extended period". In our work, we use a lower bound threshold of 40 minutes to determine proximate contact. Therefore, individuals are only considered in contact when they are collocated in a room for 40 minutes or more. This threshold was set up to account for typical lecture duration on campus (for standard 3-credit hour courses taught 3 times a week). Additionally, we compared the clustering coefficient of the contact networks for different days by varying contact thresholds as 30 and 40 minutes. The Pearson's r correlation of these was very high 0.97. Thus, we chose to use the 40 minute threshold as it produced structurally similar graphs while requiring lower space constraints. Every edge between two individuals contains a list of locations where they were possibly in contact. G forms the basis of the contact-network that we use an agent-based model to simulate. Alternatively, we also make a projection on the set L. This projection is a directed graph, H, where an edge from L i to L j represents movement from the first location to the next within a span of 60 minutes. GT's large urban campus with pedestrian pathways and motorized transit services enables direct movement between any two places on campus within the threshold. The 60 minutes threshold helps discount erroneously labeling returning from outside campus (e.g., non-residential students visiting two different locations between 2 days). H effectively describes how locations are connected and which locations could be more conducive to attracting and disseminating the virus. As a consequence, the H helps inform policy design. We compute the bipartite graph and its projections for each day of the semester.

RI: Offering Large Classes Online
As a response to COVID-19, prior work has recommended using EN to enforce a form of RImoving classes large to an online remote instruction setup while other classes are offered in-person [16,5,38]. While we have access to aggregate insights on EN contact networks, our study protocol prohibits us from accessing course-specific information at an individual level. Therefore to infer individual enrollment, we analyze the edges of the bipartite graph K. For this, we first scrape the GT's course roster for Fall 2019 (filtered to only represent the Atlanta campus). This process provides us with a location and weekly schedule for every lecture conducted on campus, including its various sections. With this information, we are able to identify which edges represent visits to lectures, and subsequently, we can account for unique visitors to a lecture. Thus, we can first identify the number of unique individuals on campus who are enrolled in classes. The aggregate data from course enrollment reports that 21, 299 students were enrolled in Fall 2019. In comparison, our inference identifies 22, 248 students. The excess number can be explained by the fact that our method does not distinguish between instructors, TAs, and students. Next, we study the unique visitors to every lecture in the scraped course schedule which gives us an estimate for the size of every class. Given the limitations of our data processing, actual enrollment sizes could be larger, but our process is less likely to count false positives [8]. Finally, to model RI, for the contact network G t , we create a counterfactual network G ′ t for each day t. These exclude collocations that took place at lecture locations during lecture times. If two people were connected solely by proximity during lectures -in a class with large enrollment -they will appear disconnected in the counterfactual network.

LC: Closing Important Locations
This article demonstrates the effectiveness of localized closures,LC, which are targeted interventions to seize mobility at different spaces on campus. For this, we identify important locations on campus by analyzing H. In the main paper, LC uses PageRank [29] as an illustrative algorithm to identify important location nodes. For robustness, we apply various additional algorithms to identify highly authoritative nodes in H -betweenness centrality [13], eigenvector centrality [4], and load centrality [25]. In the SI Appendix, we distinguish these different policies as LC PRank , LC BCen , LC ECen , LC LCen . Since RI captures a weekly schedule to determine enrollment, LC is implemented to find locations based on behavior from the past 7 days of mobility. We apply the weighted version of the algorithms mentioned earlier on the directed graph representing movement, H. The edge weight is based on the number of instances of movement between any L i and L j . After sorting the locations by importance, we determine the number of locations to shut down based on different budgets induced by RImobility and risk of exposure. For this purpose, we take the approach of a greedy algorithm which successively removes highly-ranked locations till the constraint is met (within 1% margin of error). Similar to RI, LC also render counterfactual collocation networks, G" t for each day t. In these networks, we remove instances of collocations that occurred at the shutdown locations. Figure S18d and Figure S19 shows the categories of buildings where different spaces are closed by LC policies.

Inducing Budgets and Characterizing Behavioral Scenarios
We now describe how we compare the RI and LC policies. First, we consider the effects of these policies under three behavioral scenarios. These scenarios express the spillover effects of closure that lead to students avoiding campus entirely because their entire schedule is forced online. This analysis assumes that the motivation to be present on campus is determined primarily by enrollment. We consider that, if a student has a full course load (enrolled in a minimum of 3 classes) and all their classes are offered online, that student might have less incentive to visit campus at all (for any engagement) and thus practice Avoidance. Since LC could end up closing classrooms, it can also lead to academic schedules being affected and elicit Avoidance behavior. As a result, we describe three behavioral scenarios. Persistence, is the preliminary, or null scenario, which represents no Avoidance. This counterfactual collocation graph only removes edges directly affected by RI or LC. The second scenario we model is Non-Residential Avoidance where only non-residential students with full online schedules stop visiting campus entirely. Here the counterfactual graph will remove all edges of non-residential students with fully online schedules. Lastly, the third scenario we model is Complete Avoidance where any student with fully online schedules stops activity on campus entirely (including residential students). Here the counterfactual graph will remove all edges from any student with fully online schedules. Since our study protocol prohibits us from mapping our data to other sources, we heuristically infer which individuals are likely to be residential and which are not. We label individuals as residential when they dwell an average of at least 15 minutes at residential locations between 6pm and 10am, on workdays (Monday-Thursday).
Under each behavioral scenario, we limit the number of locations that can be closed under the LC policy to ensure the level of restriction is constrained to be similar to the RI policy. We limit the number of locations under two types of restrictive budgets. The first budget is based on mobility, which is the percentage of edges remaining in the bipartite graph if a policy were to be implemented. The second budget is based on exposure risk, which is the number of unique individuals who would be in the 1-hop collocation neighborhood of positive individuals. We compute this budget by randomly sampling 2.5% of the population as positive, based on the highest 7-day average positivity rate reported by GT [15] in Fall 2019. Note, however, the effect of RI on campus can vary in different behavioral scenarios, thereby changing the budget available to design a comparable LC policy. For instance, the number of people at exposure risk is much lower in Complete Avoidance. As a result, we build multiple alternate networks representing the effect of policies under counterfactual behavioral scenarios.
The infection reduction outcomes and burdens of different policy interventions (under various behavioral scenarios and budgets) is described in Table S4-Table S7 presents boxplots that compares the distribution of disease control outcomes. Figure S10- Figure S13 show cumulative plots of disease control outcomes

Agent-based Model
We constructed an agent-based model (ABM) that captures the spread of COVID-19 between individuals active within the GT community. The model is used to evaluate the effectiveness of different policy interventions. We consider a modified version of the SEIR framework for simulating the spread of COVID-19 [34,7] by using an underlying contact network given by WIMOB. Figure S2 shows the compartments of the framework. The susceptible state (S) represents individuals who have not been infected and can contract the disease by having contact with an infectious individual. The exposed state (E) is canonically equivalent to the "incubation period" and is similar to the pre-symptomatic state found in related work [39,18]. Individuals are considered infectious when they are in either the asymptomatic state (Asym) or symptomatic state (Sym). Individuals in the asymptomatic state are assumed to be the major "spreaders" [18] and transmit the infections to susceptible individuals before they are recovered (R) [23] -after 7 days [18]. Since asymptomatic is considered a state of mild severity [32], individuals in this state do not have a risk of fatality. By contrast, for individuals in the symptomatic state, will be eventually isolated (Iso) (e.g. self-quarantine, or hospitalization on campus). Once in the isolated state, they cannot transmit the disease to individuals in the susceptible state. Unlike the asymptomatic track, the symptomatic state is considered critical severity. Therefore, after moving to the isolated state, individuals have risk of fatality and entering the death state (D). If the isolated individual survives, they enter the recovered state. We assume immunity is preserved and therefore after recovery the individual is no longer susceptible.

Definitions
Let t = {0, 1, 2, 3, ..., T } be the index of days in simulations. We denote the sequence of dynamic collocation networks indexed by day t, as {G t (A t , B t )} T t=0 . A t is the set of vertices, i.e. individuals on campus, and B t is the set of edges. The universe set of the population throughout the simulation time period is given by M = T i=1 A t . For convenience, we use a i ∈ M to index every person in the universe population set.
The SEIR model consists of seven compartments. Each of these corresponds to a function of population subsets with respect to day t: susceptible S(t), exposed E(t), asymptomatic Asym(t), symptomatic Sym(t), isolation I(t), recovered R(t), and dead D(t). For example, a i ∈ I(t) means a i is in the isolation state at day t. We use N t S→E , N t E→Asym , N t E→Sym , N t Asym→R , N t Sym→I , N t I→R , and N t I→D to denote the transitions between states between day t and day t + 1.

Model Initialization
The entire population M is fixed where M = S(t) + E(t) + Asym(t) + Sym(t) + I(t) + R(t) + D(t) for all t. To capture the positivity out of the students coming back to campus at the start of the semester, we initialize the system by setting a subset of M into Asym(0) and the reminder into S(0). The initial Frontiers 5

SUPPLEMENTARY METHODS
percentage of asymptomatic is described by: where I 0 is a parameter defined as the initial percentage of Asymptomatic at day t = 0.

New exposures
We consider two ways that an individual in the ABM could be exposed: (i) exposures that occur due to contacts among individuals captured by the mobility network (internal transmission) and (ii) exposures that occur due to contacts that occur outside of the mobility network (external transmission).
Internal transmissions happen exclusively among individuals in the model. On any given day, an edge becomes effective, when one of the susceptible individual comes in contact with the other which is infectious, i.e. asymptomatic or symptomatic, individual. Therefore, for every effective edge between two such people, the probability of the susceptible individual getting exposed is described by the transmission probability p, which is another model parameter. The probability for an susceptible individual a i entering exposed at the end of day t is given by the following function: Here, e(t, a i ) is the number of effective edges of individual a i at time t. Since (1 − p) e(t,a i ) is the probability that a i does not contracted the disease at time t under e(t, a i ) Bernoulli trials, 1 − (1 − p) e(t,a i ) is the probability that at least one effective edge leading a i to exposed.
In addition to exposure due to internal transmission, we also consider new exposure due to external transmission. We consider external transmission to be exposure resulting from the physical collocations outside the scope of mobility network. For instance, the WIMOB does not capture the connections between individuals without access to the campus WiFi or someone contacting infectious persons outside the campus. To reflect this risk in our model, for any day t, I out (t) describes the probability of infection on day t from a collocation that is external to the mobility network. We assume that the probability an individual is infected due to an external source is proportional to the number of cases in the broader community. Therefore, we model the probability of external infection as a function of confirmed cases in Fulton county, where GT is located [27]. C t represents the confirmed cases reported by Fulton County where C max is the maximum number of the cases over the whole period, I out (t) is given by where α is a parameter scaling the normalized confirm cases in the surrounding county. The resulting number of external infections on day t is then modeled to be are Binomial with |S(t)| trials with probability of success I out (t). 6 In summary, for every day t > 0, the overall number of individuals that become newly exposed is represented as N t S→E which is the result of both external and internal transmissions.
internal transmissions

Model dynamics after exposure
After exposure, individuals in the model will progress through other disease states in our model. We update the number of individuals in each state daily to reflect transitions between them. The transitions between the states on day t are summarized according to the following equations: After an individual has been exposed, they will spend ∆ S days in an incubation period. At day ∆ S after their exposure, individuals will become a symptomatic infection with probability p S . Otherwise the agent will become an asymptomatic infection This process is given by the following two equations: Individuals who enter the asymptomatic state will recover after ∆ Asym→R days since they were first exposed. Thus, we represent the number of transitions from asymptomatic to recovered on day t as: On the other hand, individuals who enter the symptomatic will eventually enter the isolation state [18]. The time that individuals spend in the symptomatic state before entering the isolated state is normally distributed δ t I ∼ Normal(∆ I , σ 2 I ). We simulate each individual's transition between symptomatic and Frontiers 7

SUPPLEMENTARY METHODS
isolated by using a sampling function Γ(a i , t, ∆ t ) and a function τ (a i , t) that returns the days since exposed respectively: first day of a i entering exposed, a i ∈ Sym(t) +∞, otherwise The aggregated transitions N t Sym→I between symptomatic and isolated is the sum of the distribution above on each day t.
Individuals who enter the isolated state may end up with one of two states: dead or recovered. We defined N t I→D as following another binomial distribution with parameter p D : The transitions between isolation and recovered is quite similar to the transitions between symptomatic and isolation except δ t R ∼ Normal(∆ R , σ 2 R ) where ∆ R and σ R are the two parameters standing for the mean and standard deviation of days for an individual in the isolation state entering recovered since the first day of infection. This leads to:  p D Death rate under isolation 0.0006 - [20] The variables p, α, and I 0 are estimated by calibrating the simulation model on the first 5 weeks of positivity rates provided by GT surveillance for Fall 2020, while incorporating external cases from Fulton County. These parameters were found by validating the ABM on the remaining weeks of Fall 2020. Figure  S3 shows model estimate during the calibration and validation period.

Model calibration
Most of our model parameters can be estimated from previous studies (see Table S1). However, three parameters in our study are not easily estimated from previous studies: (i) the proportion of the agents that begin the semester asymptotically infected, I 0 , (ii) the probability of transmission between a given infectious individual and susceptible individual given a contact in the mobility network, p, and (iii) the scaling factor α used to determine probability of transmission due to contact outside of WIMOB network on day t, I out (t) (see (S1)). We fit these three parameters to the published weekly positivity rate (percentage of asymptomatic cases) as reported by GT's asymptomatic surveillance testing program [28]. To fit the parameters, we performed calibration to minimize the root mean square of error(r.m.s.e) between the simulation estimates of the weekly positivity rate and the observed weekly positivity rate on GT's campus of the Fall 2020 semester as reported by the surveillance testing program.
To perform the calibration, we used two sets of public data pertaining to 2020 Fall semester at GT: (i) the confirmed cases in Fulton County [27], and (ii) the aggregated surveillance test positivity rate for each week [28]. The former helps estimate the daily external infection percentage. The latter is the ground truth trajectory we fit our model on. We consider the data aggregated by week because each individual on campus can only get tested once per week. The positivity rate provided by the surveillance testing data can be interpreted as the estimated percentage of new asymptomatic cases out of the total testable population which includes susceptible, exposed, and asymptomatic -with an assumption that every testable population get tested at the same rate.

Frontiers 9
To formalize the calibration problem, let R w be the surveillance-testing aggregated result at week w. Let S(I 0 , α, p, w) be the function of the simulation model which returns the percentage of new asymptomatic in week w out of the total testable population. For every combination of parameters, the predicted result for each week w is estimated by taking the average of N simulation outputs. The objective function is: The optimization problem is: min We fit our model to the first 5 weeks of Fall 2020 and validate the results on the remaining weeks. After obtaining the optimal set of parameters, for robust comparison of policies with different viral variants, we generate a range of parameters by compromising the r.m.s.e within 40% of the minima [7]. First, we implement the Nelder Mead method [22] to discover the optimal set of parameters that minimizes the r.m.s.e. Next, we sample 40 different combinations of parameters within 40% of the minimum r.m.s.e to estimate the means and standard deviations of these parameters ( Table S1). Throughout this paper, we pool together all simulation results across those parameters over multiple runs (N = 15) and report the 2.5th and 97.5th percentiles of the simulation outputs for every policy experiment.

Sensitivity Analyses
In this section, we design complementary experiments to inspect the robustness LC policies under different setups and calibration approaches. These variations are defined as follows: Calibration periods (V1) : For the results in the main paper, we discuss results with our ABM calibrated on the first 5 weeks of surveillance testing data. For additional analyses, the model parameters are re-estimated based on the surveillance data from week 5 − 9 and 10 − 14 in Fall 2020 at GT. The calibration is validated on the remaining weeks in the semester. Figure S3 shows the calibration and validation. The results of policy comparison with these variations can be found in Table S8 and  Table S9, for weeks 5 − 9 and 10 − 14 respectively. Additionally, Figure S8 shows boxplots to compare the distributions of different policies, while Figure S14 and Figure S15 show cumulative plots of the disease control outcomes, for weeks 5 − 9 and 10 − 14 respectively.
Campuses and counties (V2) : For the results in the main paper, the calibration of our ABM reflects certain latent factors inherent to GT that could affect both mobility behavior as well as testing results. To complement this we consider calibrating our data under different settings informed by surveillance testing from other similar large universities. This analysis is intended to represent the GT community in a different geographic setting, which is influenced by a different surrounding community, policies and resources. The new parameters are estimated based on the first 5 weeks of surveillance testing from the University of Illinois at Urbana-Champaign (UIUC) and the University of California, Berkeley (Berkeley) [26,33], and the corresponding county data [10,9] The calibration is validated on the remaining weeks in the semester. Figure S4 and Figure ?? show the calibration and validation for UIUC and Berkeley respectively. The results of policy comparison with these variations can be found in Table S10 and Table S11. Additionally, Figure S9 shows boxplots to compare the distributions of different policies, while Figure S16 and Figure S17 show cumulative plots of the disease control outcomes.
The estimated parameters with these calibration variations are described in Table S3. Both RI and LC are evaluated in the same infection reduction metrics and burden metrics again under behavioral scenarios S1, S2, and S3. Since the budgets are structural (mobility, and exposure risk) the LC policies are unchanged among the variants. Moreover, since the burden metrics are structural, those results are invariant.

Implications for Policy Design
To evaluate the efficacy of policies, we inspect infection reduction by simulating the disease with contact networks from Fall 2019. Since managed WiFi networks accumulate logs for long periods of time, policymakers can use WIMOB to model data from previous semesters and experiment with closure policies like LC. We show that WIMOB can provide retrospective disease-mitigating insight into multiple counterfactual behavioral scenarios. For instance, policymakers can consider studying seasonal behaviors over multiple semesters for more robustness. Since the underlying data is longitudinal, it provides the flexibility to realistically assess policy interventions at different time points and also study updating policies. Restricting movement on campus at different time-points is known to exert varying degrees of control on disease spread [7]. Our data also shows that mobility on campus varies across the semester and therefore, allows policymakers to consider loosening shutdowns depending on the phase of the semester.
Policy design is determined by practical budgets. We model two kinds of budgets, mobility reduction and risk of exposure. The former represents disruptions in space utilization, availing services, and social life. The latter translates to the testing burden on campus. Our analysis determines the budget in different behavioral scenarios by observing the changes to the graph when large classes are moved online. This is to ensure an equitable comparison with targeted policies. However, in real situations, these budgets can be relaxed or restricted based on that campus' preparedness to tackle a pandemic. For instance, a hypothetical campus that can test everyone every day might not be constrained by risk of exposure. Alternatively, policymakers can model other tangible budgets such as the capacity in isolation wards or available hospital beds. This can be informed by practical limitations of the campus. Similarly, this paper only assesses limited forms of cost, e.g., students avoiding campus or closing locations. From a financial perspective, university campuses can digitize their core service-education-but still realize losses from other curtailed services [21,3,37]. When students avoid campus it can lead to direct losses from meal passes and parking and also quantifiable losses to learning outcomes [1,11] Policymakers can compute actual costs by complementing this data with information from other sources (e.g., revenue generated by cafes and stores on campus). This can help qualifying WIMOB to reflect different costs and in turn help design policies that optimize for financial losses. Different campuses have different priorities and challenges in implementing policies.

Privacy, Ethics and Legal Considerations
We purposefully compare our prototype targeted policies against moving classes online because of practical budgets within the university. Both the WIMOB and EN based contact networks are derived from archival data accumulated by universities. This does not require instrumenting campus or its community with any new form of surveillance infrastructure. However, its use for a different purpose demands approval

Frontiers 11
by an IRB. Moreover, acquiring these kinds of data would require collaborating with data-stewards (e.g., the IT department) to establish a data-use agreement. This document must clarify how the data will be de-identified, transferred, and stored.
For this form of data, the critical privacy challenge might not be localization itself, but rather the aggregation of data over a period of time [35]. Data spanning a longer period are more susceptible to cross-analyzing and identifying. To mitigate over-accumulation of data, we suggest an adherence to principles of data minimization [31]. Instead of storing entire mobility graphs, the campus can compute and preserve only high-level insights, such as the importance of locations. This redacts any underlying individual behavior and corresponding identifiable information. Actually, for future purposes campuses can consider a form of differential privacy that authorizes limited forms of data querying depending on the privileges of the stakeholder [2].
An operational application would require the university to update the terms of use for its managed network. Particularly, the university should disclose how this data can be used in critical circumstances that invoke shared vulnerabilities [6]. On notifying the campus community of this change it offers individuals the choice to refrain from using the university network. Prior work on a sample within the same university campus shows that 90% of students are connected to the network on any given day [8]. Therefore, proposing such an opt-out condition can be viewed as an unfair choice. As a result, the campus needs to develop a contingency plan to accommodate network access to users who do not want their mobility behavior to constitute the aggregated insights.

Limitations and Future Work
This work presents evidence that university campuses can repurpose existing data sources to inform the design of LC policies that can control COVID-19. We evaluate these policies as alternatives to other data-driven, but, broad impact policies that universities consider implementing, such as moving large classes online. One of the drawbacks of this analysis, however, is that it assumes all edges to be the same. For example, when constraining by mobility, in real scenarios losing certain visits might be more valuable than others. Decline in mobility around profit-making services, such as shops and cafeterias, versus losing mobility at common rooms have a different tangible effects on campus. Currently, we take an agnostic stance towards the mobility behavior, where all visits at all locations are the same. In reality, implementing policies could have inequitable qualitative impacts despite appearing to have a similar network configuration. This can be improved by embedding more qualitative information into the network and conceiving ingenious ways to associate costs to edges.
Similar to the assumption that all visits and locations, the current work also assumes all people to be equal. However, different people have different underlying conditions that can make their vulnerabilities more concerning [30]. The privacy safeguards of this study restricted the research team from acquiring any additional demographic or historical information. Further work can attempt to characterize the nodes by randomly seeding the network to reflect the approximate demographic break up of the community. Alternatively, researchers could try to estimate some demographic based on behavior as well. However, to leverage accurate individual information, even for operational use during a public health emergency, policymakers and researchers need to develop new privacy protocols [24].
Lastly, this paper only studies three rudimentary behavioral scenarios, persistence, non-residential avoidance, complete avoidance. Yet, other substitution behaviors are possible and the richness of networks leveraged with WIMOB enables the exploration of various new scenarios that can be triggered by policy interventions on campus. For instance, individuals might not even visit transitory spaces, such as lobbies or cafes between classes. Certain collocations could be the consequence of social ties which might never be developed because of a shutdown (e.g., project teams meeting outside of class). Further research can illuminate the effects of policies in more specific scenarios by modeling post-intervention behavior more accurately. We create a contact network of only students with WIMOB and compare it with insights from contact networks created with EN. On average, we find the contact network constructed with WIMOB shows fewer average contacts, lower density and higher average shortest path (between reachable paths). Moreover, within WIMOB itself, characterizing all spaces reveals more contacts and shorter paths than only focusing on contacts in lectures. While the proportion of the largest component appears similar, note that with WIMOB, on average about only 70% of the students visit campus on a given week. We further inspect the disease-mitigating structural changes of the RI policy on the network. We observe that the changes across all metrics with EN appear to be more drastic than compared to WIMOB. The results in the main paper use variables p, α, and I 0 as estimated by calibrating the simulation model on the first 5 weeks of positivity rates provided by GT surveillance for Fall 2020, while incorporating external cases from Fulton County. For sensitivity analyses, we perform calibrations on GT data for weeks 5 − 9 and 10 − 14. Additionally, we perform calibrations on first five weeks of UIUC and Berkeley positivity rate (along with data from their respective county). These parameters were found by validating the ABM on the remaining weeks of Fall 2020. To assess the basic reproductive number (R 0 ) of our ABM we study the first 4 weeks of the disease. We find the effective R 0 to be higher for Fall 2019 than Fall 2020 as the mobility behaviors between the 2 semesters was vastly different. Note, Fall 2020 exhibits only 39% of the mobility we observe in Fall 2019. In fact, the ABM is calibrated on Fall 2020, where behavior was subject to pandemic related closures, but in Fall 2019 the mobility was not hindered by any interventions. Thus, Fall 2019 reflects a counterfactual of Fall 2020 without any closures. Note that this table is the same as ??. We repeat the results here for easier comparison of LC PRank to other algorithms shown in Table S5, Table S6 and  Table S7. Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC PRank with RI. We find that LC PRank leads to significantly improved peak infection reduction and internal transmission. In terms of reduction in total infections, the outcomes are comparable in general but can vary by specific scenarios. In addition, every policy also exerts some burden on campus, either in terms of locations affected, students avoiding campus or isolation. We observe that LC PRank policies focus on fewer locations (except in S3 ). Moreover, these policies affect fewer student's schedules and therefore fewer people avoid campus due to completely remote schedules. Finally, LC PRank does not increase the percentage of people completely isolated on campus (p-value: < 0.01: * , < 0.001: * * ). Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC BCen with RI. We find that LC BCen leads to significantly improved peak infection reduction and internal transmission, when designed with the exposure risk budget, but can be worse with the mobility budget. In terms of reduction in total infections, the outcomes are typically worse. In addition, every policy also exerts some burden on campus, either in terms of locations affected, students avoiding campus or isolation. We observe that LC BCen policies focus on fewer locations (except in S3 ). Moreover, these policies affect fewer student's schedules and therefore fewer people avoid campus due to completely remote schedules. Finally, LC LCen does not increase the percentage of people completely isolated on campus (p-value: < 0.01: * , < 0.001: * * ). Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC ECen with RI. We find that LC ECen leads to significantly improved peak infection reduction and internal transmission. In terms of reduction in total infections, the outcomes vary by specific scenarios. In addition, every policy also exerts some burden on campus, either in terms of locations affected, students avoiding campus or isolation. We observe that LC ECen policies focus on fewer locations (except in S3 ). Moreover, these policies affect fewer student's schedules and therefore fewer people avoid campus due to completely remote schedules. Finally, LC ECen does not increase the percentage of people completely isolated on campus (p-value: < 0.01: * , < 0.001: * * ). Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC LCen with RI. We find that LC LCen leads to significantly improved peak infection reduction and internal transmission. In terms of reduction in total infections, the outcomes are comparable in some scenarios but can vary in specific scenarios. In addition, every policy also exerts some burden on campus, either in terms of locations affected, students avoiding campus or isolation. We observe that LC LCen policies focus on fewer locations (except in S3 ). Moreover, these policies affect fewer student's schedules and therefore fewer people avoid campus due to completely remote schedules. Finally, LC LCen does not increase the percentage of people completely isolated on campus (p-value: < 0.01: * , < 0.001: * * ).  Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC PRank with RI. We find that LC PRank leads to significantly improved peak infection reduction and internal transmission. In terms of reduction in total infections, the outcomes are better in general but can be comparable in specific scenarios. The burden exerted on campus is the same as structural impacts of LC PRank (Table S4). (p-value: < 0.01: * , < 0.001: * * ).  Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC PRank with RI. We find that LC PRank leads to significantly improved peak infection reduction and internal transmission. In terms of reduction in total infections, the outcomes are better in general but can be comparable in specific scenarios. The burden exerted on campus is the same as structural impacts of LC PRank (Table S4). (p-value: < 0.01: * , < 0.001: * * ).  Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC PRank with RI. We find that LC PRank leads to significantly improved peak infection reduction, internal transmission and total infections. The burden exerted on campus is the same as structural impacts of LC PRank (Table S4). (p-value: < 0.01: * , < 0.001: * * ). Within each behavioral scenario, we perform the Kruskal-Wallis H-Test [19] to compare outcomes of LC PRank with RI. We find that LC PRank leads to significantly improved peak infection reduction, internal transmission and total infections. The burden exerted on campus is the same as structural impacts of LC PRank (Table S4) Collocation Projection Figure S1: In a managed network, SNMP updates the logs by describing device association to an AP at a certain timestamp. WIMOB mines these logs to characterize mobility as a bipartite graph. The nodes are partitioned to describe people nodes (e.g., P 1, P 2) connected to locations nodes (e.g., L1, L2). Every edge across the partition describes people visiting locations on campus during different times (e.g., t 1 , t 2 ). Projecting the bipartite on people nodes helps construct a contact network (e.g., P 1 and P 2 were collocated at L1 at t 1 ), while projecting it on locations helps construct a directed movement graph (P 2 dwelled at L1 and then at L2).   [28,15], the total testable population is defined as the summation of susceptible, exposed, and asymptomatic. Infectious persons are in either symptomatic or symptomatic. For every effective edge in the mobility network, a susceptible individual that is exposed to an infectious person becomes infected with probability p. Individuals may also get infected due to an exposure not captured by the WIMOB network which occurs with probability Iout(t) on day t. account for new infected cases. (b) The mobility behavior represented by WIMOB changes every day of the semester (shown weekly here). The contact network constructed from WIMOB forms the underlying contact structure of the ABM.  Figure S3: We calibrate ABM on positivity rates from Fall 2020 at GT. The objective function of the calibration is to minimize the r.m.s.e. with the weekly average of positivity rate obtained from surveillance testing results at GT [15]. (a) The parameter that determines external transmission of infections on a given day, Iout(t), is a function of cases in Fulton county (where GT is located). (b) The models discussed in the main paper are calibrated using the first 5 weeks of data. We illustrate the output for a range of parameters that incorporate quantitative uncertainty, i.e., within 40% of the r.m.s.e. (c, d) illustrate calibration on the second period of 5 weeks and third period of 5 weeks respectively. These only show the optimal parameter output. The shaded region around the lines show the 2.5 th and 97.5 th percentile.  Comparison of RI with LC ECen . Under all behavioral scenarios, for peak infection reduction (b) and internal transmission reduction (c), LC ECen shows better disease control outcomes than RI. For total infection reduction (b), LC ECen is better in S1 and worse in S3 when designed within an exposure risk budget.

Frontiers 19
(d − f ) Comparison of RI with LC ECen . Under all behavioral scenarios, for peak infection reduction (d) and internal transmission reduction (f ), LC ECen shows better disease control outcomes than RI. For total infection reduction (e), LC ECen is better in S1 and worse in S3 when designed within an exposure risk budget.          Figure S18d: The locations shutdown by each policy are grouped into the the general building category. The distribution of locations is different between policies, for example, in S1 (a) and S2 (b), LC closes fewer locations that RI. Even when targeting spaces in similar buildings, the locations are qualitatively different -RI only affects classrooms, whereas LC also closes smaller spaces like breakout rooms, reading areas and cafes. LC In S3 (c) we find LC to target locations in a greater variety of buildings, but it also targets more locations to utilize the budget.  Figure S19: The locations shutdown by each policy are grouped into the the general building category. The distribution of locations is different between policies, for example, in S1 (a) and S2 (b), LC closes fewer locations that RI. Even when targeting spaces in similar buildings, the locations are qualitatively different -RI only affects classrooms, whereas LC also closes smaller spaces like breakout rooms, reading areas and cafes. LC In S3 (c) we find LC to target locations in a greater variety of buildings, but it also targets more locations to utilize the budget.