COVID-Forecast-Graph: An Open Knowledge Graph for Consolidating COVID-19 Forecasts
Summary
Peer-reviewed paper by Rui Zhu, Krzysztof Janowicz, Gengchen Mai, Ling Cai and Meilin Shi, COVID-Forecast-Graph: An Open Knowledge Graph for Consolidating COVID-19 Forecasts and Economic Indicators via Place and Time, published in AGILE: GIScience Series, 3, 21, 2022, DOI 10.5194/agile-giss-3-21-2022, from the 25th AGILE Conference on Geographic Information Science. The authors, from the University of California, Santa Barbara, the University of Vienna and Stanford University, argue that thousands of heterogeneous COVID-19 data sources have left scientists, decision makers, journalists and the public facing a medical and a social pandemic of distrust and misinformation. Noting that no existing knowledge graph covers COVID-19 forecasts, they motivate exposing forecasts as a knowledge graph, showcase queries, and interlink forecasts with economic indicators.
Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used
Cite this document
“Zhu Et Al 2022 Agile Giss 3 21 Covid Forecast Graph.”
Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/zhu-et-al-2022-agile-giss-3-21-covid-forecast-graph-f90a317e74/
, SHA-256 f5d69f696004ce306ae4a83eacf4d982cafbc16136600c0bf2e7833b544ef1fe.Full text
AGILE: GIScience Series, 3, 21, 2022. https://doi.org/10.5194/agile-giss-3-21-2022
Proceedings of the 25th AGILE Conference on Geographic Information Science, 2022.
Editors: E. Parseliunas, A. Mansourian, P. Partsinevelos, and J. Suziedelyte-Visockiene.
This contribution underwent peer review based on a full paper submission.
© Author(s) 2022. This work is distributed under the Creative Commons Attribution 4.0 License.
COVID-Forecast-Graph: An Open Knowledge Graph for
Consolidating COVID-19 Forecasts and Economic Indicators via
Place and Time
Rui Zhu 1,2 , Krzysztof Janowicz1,2,3 , Gengchen Mai 1,2,4 , Ling Cai 1,2 , and Meilin Shi 1,2
1
STKO Lab, Department of Geography, University of California, Santa Barbara, USA
2
Center for Spatial Studies, University of California, Santa Barbara, USA
3
Department of Geography and Regional Research, University of Vienna, AT
4
Department of Computer Science, Stanford University, USA
Correspondence: Rui Zhu (ruizhu@ucsb.edu)
Abstract. The longer the COVID-19 pandemic lasts, the itative data sources that report data occasionally, we are
more apparent it becomes that understanding its social bombarded with thousands of heterogeneous sources from
drivers may be as important as understanding the virus it- governments, research labs, the industry, media, and indi-
self. One such social driver is misinformation and distrust viduals, with an update frequency of minutes, not days.
in institutions. This is particularly interesting as the scien- Naively, one may assume that this makes for more in-
tific process is more transparent than ever before. Numer- formed citizens and decision making as the transparent
ous scientific teams have published datasets that cover al- process in which data are reported by many of these orga-
most any imaginable aspects of COVID-19 during the last nizations enables the comparison of reports and forecasts.
two years. However, consistently and efficiently integrat- Instead, we are facing two pandemics, a medical and a so-
ing and making sense of these separate data “silos” to sci- cial one at the same time. There may be different reasons
entists, decision makers, journalists, and more importantly for this such as distrust in authorities, fake news, misinfor-
the general public remain a key challenge with important mation campaigns, social inequality, the type of rewards
implications for transparency. Several types of knowledge systems powering social media more broadly, and so on
graphs have been published to tackle this issue and to en- (Roozenbeek et al., 2020; Cuan-Baltazar et al., 2020; Tas-
able data crosswalks by providing rich contextual infor- nim et al., 2020). For example, one current work warns
mation. Interestingly, none of these graphs has focused on us that with richer countries are preparing to lift public-
COVID-19 forecasts despite them acting as the underpin- health interventions, widely open the economy, and report
ning for decision making. In this work we motivate the data less frequently thanks to the success of vaccination
need for exposing forecasts as a knowledge graph, show- campaign, there are still billions of people that are vul-
case queries that run against the graph, and geographically nerable to the pandemic (Mathieu, 2022). Providing open,
interlink forecasts with indicators of economic impacts. clean, and reliable data to the general public regarding the
COVID-19 pandemic will be a long-term mission for the
Keywords. COVID-19, place and time, interoperability, whole society.
knowledge graphs, knowledge representation
We believe that one way to support citizens and also de-
cision makers in the industry, NGOs, and even local gov-
ernments, is to provide access to pre-integrated data that
visualize relationships across multiple COVID-19 facets
1 Introduction such as case loads and economic impacts. For instance, a
key argument brought up by those that questioned lock-
A pandemic such as the Spanish flu that started in 1918 down at the early stage of the pandemic was the types
is not merely an entirely different event than the ongo- of places that are being forced to close while others re-
ing COVID-19 pandemic for medical and biological rea- main open. Put differently, it is difficult to put together
sons, but also due to changes in social structures, the medical, economic, governmental, and other factors in a
densely interlinked supply chains of our global economy, coherent picture. One such example is the aggregation of
and most notably in the ways in which information is com- California’s Central Coast with its relatively lower pop-
municated. Instead of a limited number of largely author- ulation density and lower case loads with the very heav-
1 of 12
ily affected and very densely populated Southern Califor- over the graph, we made it available as a public query end-
nia region around Los Angeles. Understandably, the com- point3 .
munities that form the Central Coast rallied for the cre-
ation of a separate region as it would have allowed them
to keep their economies open (Brune, 2020). Another ex- 2 Related Work
ample is the dozens of different forecasts with their un-
derlying assumptions such as the introduction or easing of Over the past two years, a great number of data sets has
restrictions. To citizens and even the press, it often remains been introduced to help understand and mitigate the global
unclear where numbers that are used to justify new mea- COVID-19 pandemic from various aspects. Particularly in
sures or relax old one are taken from. One early example the Scientific Data community4 , more than 32 data de-
comes from Germany. In the fall 2020, former Chancellor scriptors, analysis, or comments have been published, in-
Angela Merkel urged the public to return to stricter mea- cluding topics from government intervention and policy
sures as “coronavirus infection rate could hit 19,200 per (Desvars-Larrive et al., 2020; Zheng et al., 2020; Porcher,
day in Germany if the current trend continues” (Michelle, 2020; Shiraef et al., 2021), human mobility (Zheng et al.,
2020). This statement was met with wide-spread criticism 2020; Kang et al., 2020), biomedical studies, (Ellinger
even from the press and was called alarmist as Germans et al., 2021; Liu et al., 2021; He et al., 2020; Desai et al.,
could not imagine that numbers would climb that high. As 2020), and those on psychology impacts (Yamada et al.,
it turned out just weeks later, the daily cases reported in 2021; Mondino et al., 2020; Sugaya et al., 2020; Bailon
Germany easily surpassed this estimate. A major concern et al., 2020). However, despite its importance in decision-
raised by the press was a lack of transparency. It took days making, few data sets are available that describe forecast
to find the particular forecast Chancellor Angela Merkel of COVID-19 spread, not to mention its cross-walks with
was referring to in her press conference. With an ever in- economic indicators, census data, and so on. Meanwhile,
creasing number of research teams attempting to study the several teams have developed knowledge graphs (Hogan
spread and impacts of the COVID-19 pandemic, a compar- et al., 2020) to give access to pre-integrated data related to
ison of the models underlying these forecasts and their un- COVID-19, such as case loads, impacts on transportation,
derlying, often hidden, assumptions, e.g., the widespread scientific literature, and drug repurposing (Michel et al.,
usage of vaccines, will be key to understanding the trend 2020; Wang et al., 2020; Domingo-Fernández et al., 2020).
of the pandemic for months or even years to come. For These graphs aim at establishing an ecosystem of data that
the United States alone, there are 114 models published involves multiple disciplines so that a single piece of in-
by about 93 different teams (as of February 2022) to pre- formation can be timely enriched by data from other disci-
dict cumulative and incident deaths, as well as the number plines following FAIR principle (i.e., findable, accessible,
of hospitalization1 . Each of these models comes with own interoperable, and reusable). Interestingly, none of these
assumptions, different spatial and temporal scales, and is knowledge graphs targets forecasts and their underlying
updated with new measures being introduced, variants be- assumptions despite their key roles in decision-making.
ing detected (e.g., Omicron BA2), and with additional vac-
cines becoming available. These models also range widely
in their overall fatality rates, at times even by a full or-
3 Methodology
der of magnitude and as will be discussed below. Finally,
their accuracy varies substantially over time and with geo-
graphic space, i.e., across regions. In this work, we introduce the COVID-Forecast-Graph
focusing on providing a holistic view of Web-available
In this work, we develop a geographically-enabled knowl- forecast models about COVID-19. While some platforms
edge graph for COVID-19 forecasts and an ontology to and news organizations provide a visual representation of
semantically represent the ingested data. We enrich this many of these models, they differ substantially from our
graph with data collected about the various assumptions work in several regards. First, we publish the actual model
underlying forecast models and the type of modeling they outputs, not simply visuals. Second, while these mod-
use, e.g., Machine Learning versus Susceptible, Exposed, els originate from a common API that makes them Web-
Infectious, and Recovered (SEIR) models. Next, we in- available, we integrate data about the underlying assump-
tegrate this graph with economic indicators. The entire tions behind these models from several resources, thereby
graph is indexed both geographically and temporally. Fi- enriching the source data. Third, we pre-integrate the mod-
nally, to demonstrate the value of our work, we will els with other types of data such as reported “ground
showcase several exemplary queries against the graph, truth”, credit card usage in business sectors, human mobil-
some with surprising results. The graph, ontology, testing ity, and employment rate through the nexus of place and
queries, visualizations, as well as the lifting source code time at various scales. Based on the place and time com-
are available online2 . To support full SPARQL queries ponent, we also provide means to easily navigate between
regions. For instance, we showcase that models perform
1 3
https://zoltardata.com/project/44 https://stko-roy.geog.ucsb.edu/covid
2 4
https://github.com/zhurui0509/COVID-Forecast-Graph https://www.nature.com/collections/ebaiehhfhg
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 2 of 12
best in selected states and no model outperforms all other 3.1.2 Daily Reported Cases and Deaths
models geographically. Finally, by modeling forecasts and
related data using an ontology on top of the international To validate forecasts made by different teams using dif-
SOSA/SSN standards (Janowicz et al., 2019; Haller et al., ferent methods under various assumptions, we also in-
2019) for the semantic representation of observation data, clude daily cases and deaths reported from Johns Hop-
we aim to support interoperability, machine reasoning, and kins University’s Center for Systems Science and Engi-
the prediction of new links on top of our graph. In this neering (JHU CSSE)8 as the “ground truth” (Dong et al.,
section, we first elaborate major data sources for COVID- 2020). In contrast to COVID-19 forecasts, the cases and
Forecast-Graph, based on which design of the underlying deaths are represented only as point-based values rather
ontology - COVID-SO - will be discussed. than a quantile distribution. These data sets have a daily
temporal resolution and cover spatial units including na-
3.1 Data Source tion, state, and county. In COVID-Forecast-Graph, we re-
gard JHU CSSE’s repository as the sensor and the reported
cases and deaths as observations.
We collect data directly from the listed repositories below.
Each of these repositories might have their own source of
collecting the raw data (e.g., through modeling, manually, 3.1.3 Economic Data
or from private companies). To be able to integrate dif-
ferent model forecasts, “ground truth” observations, and COVID-19 has a substantial impact on both the global and
economic indicators, we represent them all within a sen- domestic economy. This impact is expected to last for sev-
sor and observation framework (i.e., SOSA/SSN). eral years. In order to build connections between COVID-
19 variables, such as incident cases, with the local eco-
nomic status, we collect a range of economy-related obser-
3.1.1 COVID-19 Forecast
vations from the public repository of Economic Tracker9 .
These data include economic indicators such as unem-
The COVID Forecast Hub Team5 establish an open-source ployment rate, credit card use, job post, human activity,
repository for teams to upload their forecasts (Ray et al., and small business revenue, which are originally collected
2020; Cramer et al., 2021), and meanwhile provide ser- by individual private companies (Chetty et al., 2020). The
vices such as interactive visualizations and ensemble mod- Opportunity Insights team made these data available to the
els to facilitate teams and the general public to explore the public in tabular forms10 . Our work, in contrast, regards
data6 . Thanks to them requiring consistent specifications the Economic Tracker as a virtual sensor and the economic
for all uploaded data as well as their official use by the indicators of interest as observations so as to lift the data
US Centers for Disease Control and Prevention (CDC) to into a knowledge graph that semantically interlinks across
inform decision makers and the public7 , we use this repos- various data sources. These observations include nations,
itory as the main seed source to build the proposed graph. states, as well as counties spatial units as well as either
This also enables us to keep the graph up-to-date. daily or weekly temporal resolution.
Each forecast has multiple targets that aim at predicting a
specific variable (e.g., cases, deaths, or hospitalizations) at 3.2 Ontology Design
a specific time (i.e., prediction time) for a range of spatial
units (e.g., states and/or counties). The forecast is made for To improve the interoperability and reusability of these
1 through 20 weeks ahead of the prediction time for inci- separate data “silos", we design a COVID-19 Sensor
dent and accumulative death, 0 through 130 days ahead for and Observation ontology, COVID-SO, on top of the
incident hospitalization, and 1 through 8 weeks ahead for W3C recommended Semantic Sensor Network ontol-
incident cases. In addition to a single point prediction, the ogy (SSN/SOSA) and its extensions. In this section, we
variables are represented as a distribution with 23 quan- first discuss key concepts and their relations defined in
tiles including 19 ones from 0.05 to 0.95 with an interval SSN/SOSA, based on which the proposed COVID-SO will
of 0.05 plus 4 extra - 0.025, 0.01, 0.90, and 0.975. Our pro- be introduced next.
posed ontology captures all these various pieces of infor-
mation and represents them as a knowledge graph instead
3.2.1 Semantic Sensor Network Ontology and its
of a traditional plain tabular forms. More specifically, each
Extensions
forecast model is regarded as the virtual sensor while its
prediction results, including both single point prediction
and corresponding quantiles, are represented as observa- The Semantic Sensor Network (SSN) ontology (Haller
tions in the graph. et al., 2019) is a W3C11 and OGC12 recommenda-
8
https://coronavirus.jhu.edu
5 9
https://COVID19forecasthub.org/ https://tracktherecovery.org/
6 10
https://github.com/reichlab/COVID19-forecast-hub https://github.com/OpportunityInsights/EconomicTracker
7 11
https://www.cdc.gov/coronavirus/2019-ncov/COVID-data https://www.w3.org/
12
/forecasting-us.html https://www.ogc.org/
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 3 of 12
tion/standard to represent sensors, their observations, sam-
ples, associated stimulus, and so on. The Sensor, Obser-
vation, Sample, and Actuator (SOSA) ontology (Janow-
icz et al., 2019) is the lightweight core of SSN, in which
only the core classes and their corresponding properties
are specified. SOSA has been widely applied to domains
such as remotely sensed images (Kostovska et al., 2020),
smart city (Espinoza-Arias et al., 2019), humanitarian re-
lief (Zhu et al., 2021b), the Internet of Things (Honti and
Abonyi, 2019), and environmental intelligence (Zhu et al.,
2021a). To enable modeling homogeneous collections of
observations, an extension to SSN has been introduced13 .
It uses a new collection constructor to efficiently repre- Figure 1. Upper-level ontology of COVID-SO. The prefix of
covid-so for introduced classes (red boxes) and properties is re-
sent observations that share common characteristics, such
moved for brevity. Hollow arrow indicates the rdfs:subClassOf
as the feature-of-interest, phenomenon time, sensor, and property.
so on.
3.2.2 COVID-19 Sensor and Observation Ontology 19 forecast that will be discussed below, as (according
to the specifications) the sosa:phenomenonTime is the
SOSA and its extension facilitate us to integrate and reuse time of the targeted prediction, which is later than the
the cross-disciplinary and heterogeneous repositories re- sosa:resultTime when the prediction is executed.
lated to the COVID-19 pandemic. Concretely, we design
a three-tier ontology, with the upper-level establishing the 3.2.4 Middle-level COVID-SO.
relation between SOSA and COVID-19 related core con-
cepts; the middle-level specifying the relation among these SOSA, together with its extension, are designed with both
core concepts (some are reused from SOSA and some physical and virtual sensors in mind to support simula-
are newly specified); and the lower-level detailing various tions as well as forecasting (Lefrançois et al., 2016; Janow-
subclasses of the core concepts described in the middle- icz et al., 2019). As outlined above, forecasts are obser-
level. In the following, we discuss details for each level. vations whose sosa:phenomenonTime is later in time
than the corresponding sosa:resultTime. While applying
3.2.3 Upper-level COVID-SO. SOSA specifically to COVID-19 related virtual sensors
(i.e., forecast models), we have to extend it in order to fit
As illustrated in Figure 1, core components of COVID- our need of efficiently describing all aspects of the source
SO utilize elements of SOSA (orange boxes), and the data. More concretely, we focus on modeling the forecast,
COVID-19 related concepts (red boxes) are linked to their underlying assumptions, reported “ground truth”, as
them via rdfs:subClassOf relation so as to reuse the well as relevant economic data with the goal of intercon-
inherited properties. While this level mainly supports necting them, and potentially many other data sources as
interoperability and provides us with the proper con- well, through the consistent use of terminologies from the
cepts to describe regions (sosa:FeatureOfInterest), COVID-SO ontology. Hence, middle-level concepts ex-
time (time:TemporalEntity), observable properties tracted from each data repository (Section 3.1) are dis-
(sosa:ObservableProperty), etc., the specification of cussed as follows.
each individual data source is discussed in the middle-
level ontology. • COVID-19 Forecast
It is worth highlighting that the two properties - Figure 2 depicts the ontology to represent sen-
sosa:resultTime and sosa:phenomenonTime - differ sors and observations collected from the COVID-
in their semantics despite both describing temporal in- 19 Forecast repository (Section 3.1.1). To illustrate
formation. First, sosa:resultTime is a data type prop- it, we will use the research led by Georgia Tech
erty, which links a subject to a literal data type (e.g., University’s DeepOubreak team (GT-DeepCOVID)14
xsd:dateTime), while sosa:phenomenonTime is an ob- as an example. GT-DeepCOVID (covid-so:Project)
ject property, which links the subject with an object made a forecast (covid-so:Forecast) on 2022-02-14
(e.g., an instance of the time:TemporalEntity class). Se- (xsd:dataTime), in which there is a specific target
mantically, the sosa:resultTime is used to describe the (covid-so:Target) that forecasts the "8 days ahead
time when a sensor makes an observation while the incident hospitalization" (covid-so:TargetType) for
sosa:phenomenonTime records the time of the observed all US states. More specifically, this target predicts
phenomenon. This difference becomes particularly es- that the average daily new hospitalization (covid-
sential when modeling forecasts, such as the COVID- so:COVIDObservableProperty) in the state of South
13 14
https://www.w3.org/TR/vocab-ssn-ext/ https://deepcovid.github.io/
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 4 of 12
Dakota (covid-so:Place) on 2022-02-22 (covid-
so:TemporalEntity) will reach 38.11 (xsd:double).
GT-DeepCOVID (covid-so:Project) makes such a
forecast (covid-so:Forecast) using a model (covid-
so:Model) built on Deep Learning method (covid-
so:Method) with an assumption that the current in-
tervention policy will remain in the place (covid-
so:Assumption). The method belongs to the fam- Figure 3. Reported “ground truth” in COVID-SO. The prefix
ily of general Machine Learning techniques (covid- covid-so is removed for brevity.
so:MethodType). Moreover, the applied model is the
primary one (covid-so:ModelDesignation - used in
case a team has multiple models) and is reported basic classes: covid-so:GroundTruth, covid-
by Georgia Tech University (covid-so:Team) under so:COVIDObservableProperty, covid-so:Place, and
the Creative Common licence: CC BY 4.0 (covid- time:TemporalEntity (see Figure 3). For instance,
so:License). There are multiple reported funding it can represent that the reported “ground truth”
agents (covid-so:Organization) for this team on this (covid-so:GroundTruth) of the cumulative number
specific project (e.g., the US National Science Foun- of death (covid-so:COVIDObservableProperty)
dation). in Alaska (covid-so:Place) on 2021-01-15
Other projects and their forecasts are modeled in a (time:TemporalEntity) was 987.00 (xsd:double).
similar way. Nevertheless, it is worth noting that dif-
• COVID-19 Relevant Economic Data
ferent projects might target different observable prop-
erties (e.g., daily incident cases, cumulative deaths, Likewise, we organize the semantics of economic
and daily incident hospitalization) at different spatial data using middle-level components of COVID-
resolutions (e.g., county, state, and national level). SO. Specifically, covid-so:EconomicTracker
Our ontology also represents the uncertainty inherent and covid-so:EconomicStatisticsCollection are
in the forecasts since most predictive models will out- two types of sosa:ObservationCollection, and
put a quantile distribution of the estimated variable covid-so:EconomicStatistics is an instance of
rather than a single value (see Section 3.1.1). This sosa:Observation. The reason to use observation
is achieved by introducing a list of sub-properties collection here is that those statistical observations
of the original sosa:hasSimpleResult property in often share the same characteristics, such as being
SOSA. For example, most predictions made from the collected at the same time and same region. Using a
COVID-19 Forecast repository are represented as ei- collection can thus help reduce redundancy. Again,
ther one point value or a list of 23 quantiles. There- taking observations from one of the economic virtual
fore, we designed 24 sub-properties: covid-so:point sensors - Affinity15 - as an example, we can repre-
and covid-so:quantile-N, where N is a place holder sent its sensors and observations as: the company
for the value of the quantile (e.g., 0.01, 0.25, 0.4, and Affinity (covid-so:EconomicSensor) made a col-
so on). lection of observations (covid-so:EconomicTracker)
on 2021-11-24 (xsd:dateTime), one of which
is about a set of economic statistics (covid-
so:EconomicStatisticsCollection) in the state of
Texas (covid-so:Place) on the date of 2021-11-07
(time:TemporalEntity). In the economic statistics
collection (covid-so:EconomicStatisticsCollection),
one is about the seasonally adjusted credit/debit
card spending in arts, entertainment, and recre-
ation relative to January 4-31, 2020 (covid-
so:EconomicObservableProperty) and its observed
value is 0.402 (xsd:double).
Figure 2. COVID-19 forecast in COVID-SO. Boxes in blue
are designed to describe middle-level concepts about covid-
3.2.5 Lower-level COVID-SO.
so:Model; Box in green is for covid-so:Method; Box in purple
is for covid-so:Target. Yellow boxes are for literal information.
The prefix covid-so is removed for brevity. On the lower-level COVID-SO, we introduce in-
stances that have type of the middle-level classes
• COVID-19 Reported “Ground Truth” (represented using the property rdf:type). This level
depends on specific data sources. Figure 5 depicts
The ontology for reported “ground truth” is
15
designed in a similar fashion. It involves four www.affinity.solutions
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 5 of 12
and time (right). For a place, we do not only preserve its
key identifiers such as the Federal Information Process-
ing Standards (FIPS) code and name, but also match each
place to its corresponding identifier (URL) in the Wiki-
data knowledge graph by leveraging the OWL:sameAs re-
lation. Wikidata18 is an open knowledge graph that is built
as the underlying data provider for Wikipedia. By link-
Figure 4. Relevant economic data in COVID-SO. The prefix ing to Wikidata, places in our knowledge graph are en-
covid-so is removed for brevity. riched with more detailed information, such as place type
(e.g., county), geographic coordinate, population, mayor,
ruling political party, etc. So far, the graph only includes
five examples. First of all, both covid-so:Target and regional administrative identifiers in the US. However, the
covid-so:GroundTruth share the same set of covid- place concept here can be extended to other identifiers,
so:COVIDObservableProperty, which includes covid- such as Point of Interest (POI) and hierarchical grid cells
so:incidentDeath, covid-so:cumulativeDeath, covid- (Shimizu et al., 2021). Place identifiers from other coun-
so:incidentCase, and covid-so:incidentHospitalization. tries can also be incorporated into the graph to enhance
Secondly, data about covid-so:AssumptionType and the cross-walk of COVID-19 forecast and its associated
covid-so:MethodType used in different projects are from global socioeconomic indicators. Finally, places of inter-
the categorization organized by CDC16 and we manually est in COVID-Forecast-Graph can be mainly categorized
match the project and method names to the ones that are into three hierarchical scales: county-level, state-level, as
used in the COVID-19 Forecast repository. Regarding eco- well as national-level. Note that some forecasts only work
nomic data, classes in both covid-so:EconomicSensors on a subset of spatial scales while others observe all.
and covid-so:EconomicObservableProperty are obtained With respect to temporal information, we utilize Time On-
from the Economic Tracker Data Dictionary17 . It is worth tology in OWL – OWL-Time19 – in order to directly take
noting that each of the middle-level classes (e.g., covid- advantage of its reasoning capability. Specifically, time in
so:AssumptionType and covid-so:EconomicSensor)
our knowledge graph can be modeled as an instance of ei-
provides the entry points for including future lower-level ther time:Instant or time:Interval, both of which are sub-
instances and properties from new data sources. class of time:TemporalEntity. In contrast to time:Interval,
which represents a duration or extent of a temporal pe-
riod, instances of time:Instant have no extent or dura-
tion. An interval’s beginning and end points are defined
as time:Instant. In COVID-Forecast-KG, if the temporal
scale of an observation is daily, we regard it as a time in-
stant; while if the data is an aggregation of observations
across a week, the time is modeled as an interval. By
reusing OWL-Time, we can easily retrieve and compare
observations with time as an index using simple queries,
which will be illustrated in the next section.
Figure 5. Ontology of lower level concepts in COVID-SO (only
five examples are shown). The prefix covid-so is removed for
brevity.
3.2.6 Place and Time as Nexuses for Cross-Walks
Figure 6. Ontology fragments for place (left) and time (right) in
One core strength of the proposed COVID-SO is its ca-
COVID-SO.The prefix covid-so is removed for brevity.
pability to connect different data sources that are related
to COVID-19. Places and time periods act as nexuses for
cross-walks between and within these data sources. Fig-
ure 6 illustrates ontology fragments for both place (left)
16
https://github.com/cdcepi/COVID-19-Forecasts
17 18
https://github.com/OpportunityInsights/EconomicTracker/b https://www.wikidata.org/wiki/Wikidata:Main_Page
19
lob/main/docs/oi_tracker_data_dictionary.md https://www.w3.org/TR/owl-time/
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 6 of 12
4 Result and Case Study among these forecasts only. Once users have identi-
fied a relevant project and its models, e.g., by their
4.1 Statistics about COVID-Forecast-Graph past accuracy, more contextual information can be
queried via the graph such as the type of method used,
At the date of submitting this paper (February 2022), the funding resource, the geographic location of the
COVID-Forecast-Graph consists of 1, 835 forecasts in- team behind the project and so on (see Q2 for an ex-
volving 117 research projects that were made by 94 teams ample).
starting on 2021-07-01. These forecasts include predic- Q2: Which projects implement the assumption that [lo-
tions months into the future and the database is updated on cal social distancing policies will be kept in place]?
a rolling basis. We have also included 8 economic track- Which methods do these projects utilize?
ers that jointly generate about 1, 002, 750 economic statis-
Forecast models that are developed by different re-
tics in total that are related to COVID-19. These obser-
search projects vary in their applied assumptions and
vations are made at about 3, 200 places including states
underlying mathematical methods. Therefore, in or-
and counties in the United States in about 1, 241 tem-
der to understand how a model works or which types
poral snapshots including dates and weeks. There are in
of models are more robust over time and geographic
total 220, 120, 470 statements in COVID-Forecast-Graph.
space, obtaining knowledge about both the assump-
More detailed statistics can be found in Table 1. COVID-
tions and method types becomes imperative. A model
Forecast-Graph is updated weekly by including newly ob-
with an assumption that the social distancing will be
served forecasts, reported “ground truth”, and economic
lifted in the next 4 weeks or that vaccines will be dis-
statistics.
tributed rapidly would not work well for states that
do not plan to enact such a policy or where roll-
4.2 Competency Questions out is slow. Even though many data sources have
provided metadata files, they are often stored sepa-
One strength of the presented knowledge graph is that it rately from the core data, lack a consistent format,
makes complex queries across models and datasets avail- and are not organized in a way ready for ingestion by
able at users’ fingertips. In this section, we discuss multi- downstream models or visualizations. For example,
ple exemplary competency questions together with their such information often resides as unstructured text in
corresponding SPARQL (a standard query language for FAQs. This creates a significant barrier for end users
RDF-based knowledge graphs20 ) queries. These queries to efficiently capture the context of using a specific
result from discussions with humanitarian relief specialist model and its forecasts, putting the transparency and
at Direct Relief, an international non-governmental orga- trustiness of decision-makings based on these data
nization, as well as researchers in food systems and supply sources into questions. Our COVID-Forecast-Graph
chain, and hence only reflect a small, but representative, addresses this challenge by allowing end users to use
fraction of potential queries. These queries are also clus- just one query to answer questions such as Q2 that
tered into four main groups based on their goals. Note that requires meta-level information about a predication
the part inside of the bracket (i.e., [...]) can be replaced (e.g., underlying model assumptions and used meth-
by other instances from the same class. Plus, all discussed ods).
exemplary queries can be found at the published Github
repository as well as the graph endpoint and can be read- Group 2: Comparison between sensors and
ily executed. observations
Group 1: Description of sensors and observations Q3: Find all predicted [cumulative death] in [Califor-
nia] on [2022-02-12], and compare them with the re-
Q1: Which projects have forecasts about [cumulative ported [ground truth].
death] for [California] on [2022-03-12]? When have Given all the different forecasts, a decision-maker
these forecasts been made? Which of them used [re- might want to first figure out which model performs
gression analysis]? best for a specific region at a given time. By doing
Once a model is designed and published, it will be so, all forecasts at that place and given time have to
used to predict the variable of interest frequently. be found first, and then they have to be compared
However, since these projects are developed by vari- with the reported “ground truth”. This process can
ous teams and are maintained in different approaches, be achieved by running a simple query on COVID-
not all of them are applied for predictions at the same Forecast-Graph. Moreover, since most forecast mod-
frequency (e.g., once a week, twice a week, or ran- els report both the average prediction and its quan-
dom). Hence, it is common to pick forecasts at a spe- tile distribution (i.e., uncertainty), we also provide
cific place and time and then observe the changes an example query to help users evaluate a model on
whether it intends to predict a relatively short interval
20
https://www.w3.org/TR/rdf-sparql-query/ that includes the reported “ground truth”.
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 7 of 12
Economic
Economic Economic Temporal
Key class Project Forecast Target Statistics Place
Tracker Statistics Entity
Collection
Number of entities 117 1,835 12,920,792 9 116,448 1,265,922 3,200 1,241
Total number of statements: 221,339,868
Total number of entities: 17,217,023
Total number of properties: 62
Total number of classes: 29
Table 1. Statistics of COVID-Forecast-Graph as of 2022-02-18. Prefixes are removed from the class name for brevity.
Q4: Among all the [4-week ahead forecasts] of [cumula-
tive death] in [early January 2022 (i.e., before the
Omicron peak in the US)], which model performed
the best for each state across the US?
Similar to Q3, users might want to further compare
the best model across all states in the US. Since states
might have employed different policies, differ in pop-
ulation density, mobility, international connectivity,
and so on, which can be key assumption underlying
a model, we hypothesize that for different models,
Figure 7. The best forecast model to predict the cumulative death
their performances should be rather distinct across in the next 4 weeks for each state in the US (estimated in the
states. There are many ways to evaluate a model’s month of January 2022).
forecasting capability. In this work, we showcase one
possible solution using the inter-connected COVID-
Forecast-Graph. Specifically, we first extract the ear- Q5: In which US state does the [JHUAPL-Bucky] model
liest forecast date of each research project in January perform the best (and the worst) compared with other
2022, and then use the accuracy (e.g., absolute loss) models on [2022-02-05]? How do the results differ
of predicting the cumulative death in the next 4 weeks from forecasts for [2022-02-12] (another target fore-
to find the “best” model for each state. Figure 7 il- cast date)?
lustrates the striking results. It becomes clear that no
For a specific model, such as the Bucky by
single model performs best and that there are geo-
Johns Hopkins University Applied Physics Lab21 , a
graphic patterns underlying the performance of suc-
decision-maker may wonder in which states (or coun-
cessful models. Similar analyses can be conducted at
ties) this model works best and where it falls short.
the spatial scale of counties, as well as on the tem-
To answer this question, our COVID-Forecast-Graph
poral dimension as the performance of a model for a
allows one to query and rank the performance of
specific state might also change through time based
JHUAPL-Bucky in comparison to all other available
on ever-changing local factors such as policy, eco-
models for each state at a specific target date. More
nomic status, medical resource, and so on. Please
concretely, Figure 8 depicts the ranking class (e.g.,
note that these differences are substantial, i.e., the
top 10%) of JHUAPL-Bucky in the ordered list of all
best model for a state varies substantially from the
available models’ performances (i.e., absolute loss) in
second best model and so on. For instance, the best
forecasting the cumulative death on 2022-02-05 and
model for California is about 91 fatalities (per day)
2022-02-12, respectively, for each US state. For in-
off, the second best increases to 210, while the worst
stance, we observe that JHUAPL-Bucky remarkably
model is about 7909 cases off (See Table 2).
outperforms other available models in Mississippi for
the two selected dates with both being ranked in the
top 10% (there are about 35 available models for the
Absolute Delta
Rank Model target date, so top 10% indicates that the model is
(in fatalities)
ranked in the top 4 of the ordered list of model perfor-
1 MIT_ISOLAT-Mixtures 91.45
2 MIT_CritData-GBC 210.00 mances), while the model’s performance in Califor-
3 IHME-CurveFit 219.19 nia is consistently poor (i.e., ranked bottom 10%) for
... ... ... both dates. Nevada shows a different picture where
34 UMich-RidgeTfReg 7908.64 JHUAPL-Bucky’s performance positions in the bot-
Table 2. Model comparison of predicting COVID-19 cumulative tom 30% on 2022-02-05 but increases to the top 30%
death per day in January 2022 for California, US. on 2022-02-12. In fact, running this type of analy-
21
https://docs.buckymodel.com/en/latest/
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 8 of 12
sis on several models, we are able to demonstrate a
substantial variation of model performances across
space (regions) and time, and such an analysis can
be achieved readily using SPARQL query on top of
COVID-Forecast-Graph.
Figure 9. Time series of reported incident cases and people’s
time spent at retail and recreation locations (compared to January
2020) in New York.
Q7: Which model shows the largest deviation in accu-
racy of forecasting [cumulative death] as a function
Figure 8. Performance comparison of JHUAPL-Bucky model of [population density]?
across US states on 2022-02-05 and 2022-02-12.
In addition to economic indicators, we have also
linked census related data22 into COVID-Forecast-
Group 3: Integration across sensors and observations Graph. Consequently, one can explore the associa-
tion of model performances with census-based obser-
vations (e.g., population density). For instance, Fig-
Q6: What is the relation between [reported incident ure 10 compares the correlation between model ac-
cases] and the [time spent at retail and recreation curacy in forecasting the number of cumulative death
locations] in [New York]? in 2022-02-12 and the population density (at state
level) over different models. As illustrated, 34 out
When it comes to developing policies, the spread and of 35 models hold a positive correlation between
socio-economic impact of COVID-19 require knowl- the model accuracy and population density, with the
edge not only about epidemiology but also from do- model CU-scenario_low achieving the strongest cor-
mains such as human mobility, economy, as well as relation that reaches roughly 0.40. In contrast, the
local policy. Furthermore, local decision-makers and model of RobertWalraven-ESG have slightly negative
analysts do not only need the most accurate forecast correlations with population density. This exemplary
model (see Q4) but also other essential information experiment indicates that for these models that have
such as the employment rate, small business revenue, a strong correlation (either positively or negatively)
human activity, and credit card use in different retail between model accuracy and population density, they
sectors. Therefore, integrating all these data sources may consider exploring population-related effects in
into one platform and providing efficient query capa- order to improve their performances. Furthermore, in
bilities are beneficial to facilitate optimal decisions. an attempt to diagnose a specific model that shows a
By querying reported incident cases together with the strong correlation, one can refer to queries as shown
time people spent (relative to time spent in January in Q1 and Q2 to directly retrieve the involved method
2020) at retail and recreation locations as an ordered types and assumptions for deeper investigations.
time sequence for the state of New York, users can
subsequently build visualizations such as time series
(see Figure 9) to investigate the interaction of these 5 Conclusion and Discussion
two variables. For example, we can observe that the
trend for visiting retail or recreation locations in New
In this work we introduced the COVID-SO ontology to
York reaches to a relatively high level thanks to the
model sensors and observations related to the COVID-19
low incident cases in late 2021 while it dramatically
pandemic (e.g., forecasts, reported daily cases and deaths,
drops due to the peak of Omicron variant in January
as well as economic statistics). In order to enable in-
2022. This type of analysis can potentially help re-
teroperability with other graphs, COVID-SO reuses the
searchers investigate the impact of COVID-19 on the
SOSA ontology and its extensions that are standardized by
local economy and vice verse, which might further
W3C and OGC. On top of it, we further design a middle-
advance the development of new forecasting mod-
level and a lower-level ontology that are specific to three
els that take into account economic indicators. It is
commonly used data sources for COVID-19: forecasts,
worth emphasizing that in contrast to traditional silo-
“ground truth”, and economic data. Based on COVID-
ed databases, integrating more repositories can be
SO, we create a knowledge graph - COVID-Forecast-
done with ease by using COVID-Forecast-Graph as
it makes use of global identifiers and provides disam- 22
https://www.census.gov/data/datasets/time-series/demo/po
biguation, e.g., for places. pest/2010s-state-total.html
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 9 of 12
6 Data and Software Availability
The COVID-SO ontology, as well as the generated knowl-
edge graph - COVID-Forecast-Graph - are available at http
s://github.com/zhurui0509/COVID-Forecast-Graph. Ad-
ditionally, the graph is made available at an open SPARQL
endpoint (https://stko-roy.geog.ucsb.edu/covid) using
GraphDB25 . The code to generate the graph and support
the application of the results of this study is available in
the GitHub repository as well.
Figure 10. Comparison of the correlation between model accu- References
racy in forecasting cumulative death on 2022-02-12 and popula-
tion density across different models. Bailon, C., Goicoechea, C., Banos, O., Damas, M., Pomares, H.,
Correa, A., Sanabria, D., and Perakakis, P.: CoVidAffect, real-
time monitoring of mood variations following the COVID-19
Graph - that contextualizes and enriches the aforemen- outbreak in Spain, Scientific Data, 7, 1–10, 2020.
tioned data sources so that they can be efficiently consol- Brune, E.: Santa Barbara, San Luis Obispo, Ventura Counties re-
idated and queried. Several example queries demonstrate quest removal from southern California region stay home or-
the strengths of COVID-Forecast-Graph. der, available at: https://publichealthsbc.org/new-page-examp
While we included selected data sources in the current les/santa-barbara-san-luis-obispo-ventura-counties-request-r
version of COVID-Forecast-Graph, the three-tier COVID- emoval-from-southern-california-region-stay-home-order/,
last access: 20 February 2020, 2020.
SO ontology enables further incorporation of COVID-19
data sources. This is mainly attributed to the core compo- Chetty, R., Friedman, J. N., Hendren, N., Stepner, M., et al.: The
nents of COVID-SO: Place and Time. We leverage com- economic impacts of COVID-19: Evidence from a new public
mon information about places and time periods from dif- database built using private sector data, Tech. rep., national
Bureau of economic research, 2020.
ferent data “silos” to interlink them. For example, we have
integrated Wikidata into COVID-Forecast-Graph so that Cramer, E. Y., Huang, Y., Wang, Y., Ray, E. L., Cornell, M.,
basic information, e.g., population density, elevation, cap- Bracher, J., Brennen, A., Castro Rivadeneira, A. J., Gerding,
ital cities, connectivity, temperature, water area, and so A., House, K., Jayawardena, D., Kanji, A. H., Khandelwal,
A., Le, K., Niemi, J., Stark, A., Shah, A., Wattanachit, N.,
on, can be readily extracted to facilitate researchers to de-
Zorn, M. W., Reich, N. G., and Consortium, U. C.-. F. H.:
velop more regionally informed forecasting models. In ad-
The United States COVID-19 Forecast Hub dataset, medRxiv,
dition, the study of SARS-CoV-2 genome (He et al., 2020; https://doi.org/10.1101/2021.11.04.21265886, https://www.
Reese et al., 2020) is another main direction of COVID- medrxiv.org/content/10.1101/2021.11.04.21265886v1, 2021.
19 related research. Data sources, such as GISAID23 , have
Cuan-Baltazar, J. Y., Muñoz-Perez, M. J., Robledo-Vega, C.,
published the sequence of genes which involves infor-
Pérez-Zepeda, M. F., and Soto-Vega, E.: Misinformation of
mation such as where it originated, when samples were COVID-19 on the internet: infodemiology study, JMIR public
collected, which lab detected them, and so on. COVID- health and surveillance, 6, e18 444, 2020.
SO potentially has the capability of integrating these data
Desai, S., Baghal, A., Wongsurawat, T., Jenjaroenpun, P., Powell,
sources into COVID-Forecast-Graph so as to further em-
T., Al-Shukri, S., Gates, K., Farmer, P., Rutherford, M., Blake,
power COVID-19 related research with richer situational G., et al.: Chest imaging representing a COVID-19 positive
awareness. rural US population, Scientific data, 7, 1–6, 2020.
Apart from the science community, stakeholders like gov- Desvars-Larrive, A., Dervic, E., Haug, N., Niederkrotenthaler,
ernment agencies, industry, and NGOs can contribute to T., Chen, J., Di Natale, A., Lasser, J., Gliga, D. S., Roux, A.,
as well as benefit from the proposed COVID-SO and Chakraborty, A., et al.: A structured open dataset of govern-
COVID-Forecast-Graph as well. For example, the interna- ment interventions in response to COVID-19, medRxiv, 2020.
tional NGO - Direct Relief24 - has distributed over 51,000 Direct Relief: COVID-19 Relief, available at: https://www.dire
shipments of aids globally and implemented a platform for ctrelief.org/emergency/coronavirus-outbreak/, last access: 20
users to track these aid recipients (Direct Relief, 2022). February 2022, 2022.
However, the platform and its underlying data are not pre- Domingo-Fernández, D., Baksi, S., Schultz, B., Gadiya, Y.,
integrated with other data sources. Our approach of us- Karki, R., Raschka, T., Ebeling, C., Hofmann-Apitius, M.,
ing a semantically-enriched graph enables seamless cross- and Kodamullil, A. T.: COVID-19 Knowledge Graph: a
walks between cases, demographics, economic indicators, computable, multi-modal, cause-and-effect knowledge model
and delivered relief goods. of COVID-19 pathophysiology, Bioinformatics, p. btaa834,
2020.
23
https://www.gisaid.org/
24 25
https://www.directrelief.org/ https://www.ontotext.com/products/graphdb/
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 10 of 12
Dong, E., Du, H., and Gardner, L.: An interactive web-based Mondino, E., Di Baldassarre, G., Mård, J., Ridolfi, E., and Rusca,
dashboard to track COVID-19 in real time, The Lancet infec- M.: Public perceptions of multiple risks during the COVID-19
tious diseases, 20, 533–534, 2020. pandemic in Italy and Sweden, Scientific data, 7, 1–7, 2020.
Ellinger, B., Bojkova, D., Zaliani, A., Cinatl, J., Claussen, C., Porcher, S.: Response2covid19, a dataset of governments’ re-
Westhaus, S., Keminer, O., Reinshagen, J., Kuzikov, M., Wolf, sponses to COVID-19 all around the world, Scientific Data,
M., et al.: A SARS-CoV-2 cytopathicity dataset generated by 7, 1–9, 2020.
high-content screening of a large drug repurposing collection, Ray, E. L., Wattanachit, N., Niemi, J., Kanji, A. H., House, K.,
Scientific Data, 8, 1–10, 2021. Cramer, E. Y., Bracher, J., Zheng, A., Yamana, T. K., Xiong,
Espinoza-Arias, P., Poveda-Villalón, M., García-Castro, R., and X., et al.: Ensemble Forecasts of Coronavirus Disease 2019
Corcho, O.: Ontological representation of smart city data: (COVID-19) in the us, MedRXiv, 2020.
From devices to cities, Applied Sciences, 9, 32, 2019. Reese, J. T., Unni, D., Callahan, T. J., Cappelletti, L., Ravan-
Haller, A., Janowicz, K., Cox, S. J., Lefrançois, M., Taylor, K., mehr, V., Carbon, S., Shefchek, K. A., Good, B. M., Balhoff,
Le Phuoc, D., Lieberman, J., García-Castro, R., Atkinson, R., J. P., Fontana, T., et al.: KG-COVID-19: a framework to pro-
and Stadler, C.: The modular SSN ontology: A joint W3C and duce customized knowledge graphs for COVID-19 response,
OGC standard specifying the semantics of sensors, observa- Patterns, 2, 100 155, 2020.
tions, sampling, and actuation, Semantic Web, 10, 9–32, 2019. Roozenbeek, J., Schneider, C. R., Dryhurst, S., Kerr, J., Freeman,
He, Y., Yu, H., Ong, E., Wang, Y., Liu, Y., Huffman, A., A. L., Recchia, G., Van Der Bles, A. M., and Van Der Linden,
Huang, H.-h., Beverley, J., Hur, J., Yang, X., et al.: CIDO, S.: Susceptibility to misinformation about COVID-19 around
a community-based ontology for coronavirus disease knowl- the world, Royal Society open science, 7, 201 199, 2020.
edge and data integration, sharing, and analysis, Scientific Shimizu, C., Zhu, R., Mai, G., Fisher, C., Cai, L., Schildhauer,
Data, 7, 1–5, 2020. M., Janowicz, K., Hitzler, P., Zhou, L., and Stephen, S.: A
Hogan, A., Blomqvist, E., Cochez, M., d’Amato, C., de Melo, Pattern for Features on a Hierarchical Spatial Grid, in: The
G., Gutierrez, C., Gayo, J. E. L., Kirrane, S., Neumaier, S., 10th International Joint Conference on Knowledge Graphs,
Polleres, A., et al.: Knowledge graphs, Preprint at arXiv:2003 pp. 108–114, 2021.
.02320, 2020. Shiraef, M. A., Hirst, C., Weiss, M. A., Naseer, S., Lazar, N., Bel-
Honti, G. M. and Abonyi, J.: A review of semantic sensor tech- ing, E., Straight, E., Feddern, L., Taylor, N. R., Jackson, C.,
nologies in internet of things architectures, Complexity, 2019, et al.: COVID Border Accountability Project, a hand-coded
2019. global database of border closures introduced during 2020,
Janowicz, K., Haller, A., Cox, S. J., Le Phuoc, D., and Scientific data, 8, 1–11, 2021.
Lefrançois, M.: SOSA: A lightweight ontology for sensors, Sugaya, N., Yamamoto, T., Suzuki, N., and Uchiumi, C.: A real-
observations, samples, and actuators, Journal of Web Seman- time survey on the psychological impact of mild lockdown for
tics, 56, 1–10, 2019. COVID-19 in the Japanese population, Scientific Data, 7, 1–6,
Kang, Y., Gao, S., Liang, Y., Li, M., Rao, J., and Kruse, J.: Mul- 2020.
tiscale dynamic human mobility flow dataset in the US during Tasnim, S., Hossain, M. M., and Mazumder, H.: Impact of ru-
the COVID-19 epidemic, Scientific data, 7, 1–13, 2020. mors and misinformation on COVID-19 in social media, Jour-
Kostovska, A., Džeroski, S., and Panov, P.: Semantic Descrip- nal of preventive medicine and public health, 53, 171–174,
tion of Data Mining Datasets: An Ontology-Based Annotation 2020.
Schema, in: International Conference on Discovery Science, Wang, Q., Li, M., Wang, X., Parulian, N., Han, G., Ma, J., Tu, J.,
pp. 140–155, Springer, 2020. Lin, Y., Zhang, H., Liu, W., et al.: COVID-19 literature knowl-
Lefrançois, M., Kalaoja, J., Ghariani, T., and Zimmermann, A.: edge graph construction and drug repurposing report genera-
The SEAS Knowledge Model, Deliverable 2.2, ITEA2 12004 tion, Preprint at arXiv:2007.00576, 2020.
Smart Energy Aware Systems, p. 76, 2016. Yamada, Y., Ćepulić, D.-B., Coll-Martín, T., Debove, S.,
Liu, Y., Hur, J., Chan, W. K., Wang, Z., Xie, J., Sun, D., Han- Gautreau, G., Han, H., Rasmussen, J., Tran, T. P., Travaglino,
delman, S., Sexton, J., Yu, H., and He, Y.: Ontological model- G. A., and Lieberoth, A.: COVIDiSTRESS Global Survey
ing and analysis of experimentally or clinically verified drugs dataset on psychological and behavioural consequences of the
against coronavirus infection, Scientific data, 8, 1–12, 2021. COVID-19 outbreak, Scientific data, 8, 1–23, 2021.
Mathieu, E.: Commit to transparent COVID data until the WHO Zheng, Q., Jones, F. K., Leavitt, S. V., Ung, L., Labrique,
declares the pandemic is over, Nature, 602, 549, 2022. A. B., Peters, D. H., Lee, E. C., and Azman, A. S.: HIT-
COVID, a global database tracking public health interventions
Michel, F., Gandon, F., Ah-Kane, V., Bobasheva, A., Cabrio, to COVID-19, Scientific data, 7, 1–8, 2020.
E., Corby, O., Gazzotti, R., Giboin, A., Marro, S., Mayer, T.,
et al.: Covid-on-the-Web: Knowledge graph and services to Zhu, R., Ambrose, S., Zhou, L., Shimizu, C., Cai, L., Mai, G.,
advance COVID-19 research, in: International Semantic Web Janowicz, K., Hitzler, P., and Schildhauer, M.: Environmen-
Conference, pp. 294–310, Springer, 2020. tal Observations in Knowledge Graphs, in: 2nd Workshop on
Data and research objects management for Linked Open Sci-
Michelle, A.: Merkel says German coronavirus infections could ence, pp. 1–11, 2021a.
hit 19,200 a day, available at: https://www.reuters.com/articl
e/us-health-coronavirus-germany-infections/merkel-says-ger Zhu, R., Cai, L., Mai, G., Shimizu, C., Fisher, C. K., Janowicz,
man-coronavirus-infections-could-hit-19200-a-day-source- K., Lopez-Carr, A., Schroeder, A., Schildhauer, M., Tian, Y.,
idUSKBN26J1FP, last access: 20 February 2022, 2020. et al.: Providing Humanitarian Relief Support through Knowl-
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 11 of 12
edge Graphs, in: Proceedings of the 11th on Knowledge Cap-
ture Conference, pp. 285–288, 2021b.
AGILE: GIScience Series, 3, 21, 2022 | https://doi.org/10.5194/agile-giss-3-21-2022 12 of 12
File and source
- File
- zhu-et-al-2022-agile-giss-3-21-covid-forecast-graph.pdf
- Size
- 1,298,344 bytes
- SHA-256
- f5d69f696004ce306ae4a83eacf4d982cafbc16136600c0bf2e7833b544ef1fe
- Original
- No public link identified.
- Original
- web.archive.org