The availability of information from social media, such as tweets, and human rights monitors, such as Amnesty International, Human Rights Watch, and the US State Department has led to new opportunities to measure repression and human rights protections in higher resolution. In this paper, we present HR-VEAW, a Human Rights Violation Exploration, Analytics, and Warning system, to support understanding of social conflict dynamics and human rights violations/protections with quantitative data. After briefly discussing HR-VEAW's data acquisition and analysis components, we demonstrate how it visualizes rich spatio-temporal and conceptual information, enabling the examination of changes in patterns of violation and protection in aggregate over time, or across both space and time. This way HR-VEAW helps to explain social instability and conflicts and to guide decision-making, theorizing, and predictions.
array of rights and behaviors for specific groups. This
enables us to not only look at changes in patterns of
vioThrough a qualitative understanding of conflict dynamics, lations and protections in aggregate over time or across
human rights violations and protections, and ethnic poli- both space and time, but fundamentally explore which
tics and relations, policy analysts and conflict researchers groups are being targeted or privileged by the
governbuild mental models of the underlying grievances and ment and other actors and on what specific dimensions.
alliances that structure war and peace. Currently, there For example, it helps to explain why in Ethiopia, the
is no system that directly allowed policy makers and re- Tigrayan Peoples Liberation Front (TPLF) lost influence
searchers to both inform their visions of these grievances due to their broad repression as well as provides clues
and policies with systematic, quantitative data, as well for why the Oromo Liberation Army (OLA) and TPLF are
as share the resulting maps of the spatial and concep- now cooperating against the government and when that
tual patterns that guide decision-making, theorizing, and cooperation might end [
In this paper, we present a first Human Rights Violation tations, data visualization, and interpretable ML. In this Exploration, Analytics, and Warning system (HR-VEAW) demo paper, after briefly discussing HR-VEAW’s data acthat allows users to visualize rich spatial and conceptual quisition and analysis components (S 2), we demonstrate information that is relevant to both the escalation of in- how it visualizes rich spatio-temporal and conceptual stability, as well as to how negotiators might wind down information, helping users to explain social instability tensions, and with whom. We process textual data from and conflicts and to guide decision-making, theorizing, human rights reports and other social media that commu- and predictions (S 3). nicates both historical and contemporaneous information on who is alleged to have violated or protected a broad
Figure 1 illustrates an overview of the data flow of the
HR-VEAW system, consisting of two phases. The first
phase is Data Acquisition, in which data from diferent
sources are transformed by the ETL process and stored
into a data warehouse. The second phase is Data Analysis,
in which data visualization and interpretation are used
to explore the data and to discover patterns of social The data interpretation module aims to find the logical
instability for early warnings. relationship between diferent features in the data. One
way to accomplish it is to booleanise some of the
aggre2.1. Data Acquisition gated information from the data warehouse to get the
corresponding binary indicators. These binary indicators
This phase consists of a variety of ETL (Extract, Trans- can serve as learning features in machine learning
methform, Load) processes for diferent data sources. Human ods, such as the Tsetlin Machine [
The whole data acquisition phase is structured along The final module of HR-VEAW provides early warning the producer/consumer paradigm and topic channels, of future conflicts based on the data analysis results from powered by Apache Kafka [7]. There are two major the data interpretation and data visualization modules. advantages of using Kafka in HR-VEAW. First, Kafka can process high throughput data streams from diferent sources and can be easily integrated with the ETL pro- 3. Data Visualization in HR-VEAW cess as an event driven message bus. Second, while the data warehouse only stores the structured data after the This section provides implementation details of the data ETL process, Kafka can persist data/messages during the warehouse that drives the data visualization in HRwhole data acquisition phase, so that messages are never VEAW, describes the data visualization GUI and services, lost and can be retrieved at a later time as needed. and presents an example of data visualization use case.
TIME, VICTIM and ASPECT. The hierarchy (i.e., ontol- cell with valid data across time, victim and human rights
ogy) information for regions, victims and human rights aspect. The more opaque a cell is, the more negative the
aspects is from OSM, WordNet [
The fact table is called VALENCE. Valence is an output The user can pan, tilt, rotate and zoom the visualization value of PULSAR for each human rights record, repre- with simple mouse movements. Further, there is a legend senting the polarity of the record. A negative valance displaying valence ranges for diferent opacity levels. An represents human rights violation and a positive valence info box is also present to display the cell information represents human rights protection. The VALENCE table (i.e., the corresponding attribute values and the valence has five attributes: one dimension ID for each dimension sum) when a cell is clicked. and a valence value. If a human rights record contains The clicking of a cell will also trigger the rendering of more than one value in one dimension, then each value 1D and 2D plots which are displayed beside the 3D plot. will generate one record in VALENCE. For example, if These plots correspond to 1D and 2D data cube slices that a human rights record involves two regions, then each contain the cell. For example, the first 2D plot in Figure 2 region will generate a separate valence record. (3rd plot from top) displays the temporal valence change
In addition to the five tables above, the data warehouse across diferent aspects for the clicked region and victim. has one more basic table and one more view. The table is This feature helps the user to acquire clear and precise named BOUNDARY, and contains geographic boundary co- readings of data subspaces. ordinates for the regions in the REGION table. BOUNDARY The recommendation section is designed to help has a foreign key referencing the primary key in REGION user discover interesting patterns more easily. HR-VEAW and thus it has a many-to-one relation to REGION. The dynamically generates recommendation listings for the view is tentatively called valence_dimensions. It is ifltered data subset based on diferent ranking criteria. generated by joining VALENCE with all dimension tables Examples of two ranking criteria are: valence sum and so that the group-by attributes that can be requested by change point. The ranking scores are calculated for each the data analysis modules, namely all levels in all dimen- valence value time series (i.e., the column of cells perpensions, are present in the view. dicular to the map surface in the 3D visualization) that has a specific combination of region, victim and aspect
The valence sum criterion calculates the sum of the valence values in each time series and its ranking score is based on Equation 1: The main purpose of the data visualization module is to help data analysts discover interesting patterns in human rights situations. These patterns could include bdeifetrweenecne,rseigmioinlasr,ittiymaenrdancgoerrse,lsaotcioianl ginrovuaplse,nacnedvhaulumesan (,,) = ∑︁ (,,) (1) rights aspects, which are very helpful in understanding =1 and analyzing human rights situations. where (, , ) is a specific combination of region, victim The data visualization front-end is shown in Figure 2. and aspect values and (,,) are the time series values The GUI has three sections: a filter section, a visualization for that combination. section, and a recommendation section. The change point criterion calculates the maximum
The filter section contains hierarchical filters that of absolute changes between adjacent cells in each time allow the user to select attribute values in diferent levels series and its ranking score is based on Equation 2: of diferent dimensions. The levels and values in the ifldtiemrsernespiorensetanbtltehse.iBr ycoaullnotweripnagrttsheinutsheer ctoorsreelsepcotnvdailnuges (,,) = 1m≤a<x(|(+,1,) − (,,)|) (2) at diferent levels, this filter design enables cross-level where (· ) is the maximum operator and | · | is the display and comparison of the data. absolute value operator.
The visualization section displays the aggregate Each recommendation criterion will provide a listing of values (i.e., sum of valence) for the filtered data subset. the top 10 entries. By clicking an entry, the corresponding Specifically, for each selected region, a 3D visualization time column will be highlighted in the 3D visualization. will be generated displaying the valence sums in each
case, then dealing with the grievances reflected by those civil events could potentially prevent the increase of the instability events in the following years.
To conclude, through data visualization and exploration with the help of HR-VEAW, the analyst was able to form a hypothesis that the increase of civil events before year 2009 in the Somali region could be the cause to the increase of force events in 2009 and after in the same region, which could guide their further investigation.
An example data visualization use case is illustrated in Figure 3, in which the analyst tries to identify specific grievances that may have caused instability events in the history of Ethiopia.
Instability events are mainly reflected by human rights violations in the sub-aspect Force under the the aspect of Integrity, which contains primarily armed conflicts between the government and the ethnic groups. Therefore, the analyst first chooses to visualize that sub-aspect across all victims in Ethiopia as shown in Figure 3a. 4. Demonstration Scenario
They realize that there was a large increase in force events (i.e., human rights violation events under the Force The following demonstration introduces the key consub-aspect) in year 2009 and want to find out if this pat- cepts and visualization abstraction of HR-VEAW to the tern is universal across the country or specific to certain attendees. They will comprehend the performance of sub-regions. So they select all sub-regions in the filter our propositions during their interactions with our usersection to see the time columns across the sub-regions. friendly interface.
Then they check the top-1 recommendation from the valence sum criterion, and find out that the Somali region 4.1. Demo Artifact contributed greatly to the aforementioned force event increase, as shown in Figure 3b. The demo artifact is a web application prototype repre
So the data analyst focuses into the Somali region and senting the data visualization module of HR-VEAW. The selects all aspects as well as all sub-aspects under the software stack of the web app consists of Flask, MySQL, Integrity aspect. Then they check the top-1 recommen- JavaScript and HTML. The web app will be served on a dation from the change point criterion, and find out that dedicated HR-VEAW server, which can be publicly accivil events (i.e., human rights violation events under cessed during the demo. the Civil aspect) also had a large increase around year 2009 in the Somali region with a slightly earlier start, as 4.2. Demo Plan shown in Figure 3c. This indicates that the increase of the civil events could be the cause to the increase of the force events in the following years. And if this is the
portunity to interact with the web app through any web browser on a standard laptop or a tablet. (a)
The authors would like to thank Daniel Gustafson for his work on PULSAR. This work was partially funded by NSF grant SES-2017614 and reflects only the authors’ opinions.