arX

iv:1

705.

0340

5v3

[ph

ysic

s.so

c-ph

] 2

1 Ju

l 201

7

A complex network analysis of ethnic conflicts and human rights violations

Kiran Sharma,1, ∗ Gunjan Sehgal,2, † Bindu Gupta,2, ‡ Geetika Sharma,2, §

Arnab Chatterjee,2, ¶ Anirban Chakraborti,1, ∗∗ and Gautam Shroff2, ††

1School of Computational and Integrative Sciences, Jawaharlal Nehru University, New Delhi-110067, India2TCS Research, New Delhi, India

News reports in media contain records of a wide range of socio-economic and political eventsin time. Using a publicly available, large digital database of news records, and aggregating themover time, we study the network of ethnic conflicts and human rights violations. Complex networkanalyses of the events and the involved actors provide important insights on the engaging actors,groups, establishments and sometimes nations, pointing at their long range effect over space andtime. We find power law decays in distributions of actor mentions, co-actor mentions and degreesand dominance of influential actors and groups. Most influential actors or groups form a giantconnected component which grows in time, and is expected to encompass all actors globally in thelong run. We demonstrate how targeted removal of actors may help stop spreading unruly events.We study the cause-effect relation between types of events, and our quantitative analysis confirmthat ethnic conflicts lead to human rights violations, while it does not support the converse.

Keywords: complex network analysis, ethnic conflict; human rights violation, media reports

I. INTRODUCTION

The quantitative analyses of the gigantic amount of data associated with human social conditions have gainedmuch momenta in the recent years, especially through the multidisciplinary tools and approaches. The traditionaltheories of social science together with complex network analysis [1–3] and addition of new tools and paradigms fromseveral science disciplines like theoretical physics, applied mathematics, computer science, as well as psychology, havedeveloped into what is presently known as computational social science [4] and have helped uncover new patterns ofsocial behavior, including social dynamics [5, 6]. Data made available digitally, has drawn attention of researchersacross disciplines who have collaborated and contributed in their own ways to scientifically analyze and understandcomplex social phenomena in the recent years, previously not known to the present scale of detail.Social behavior, conditions, events, organizations, which have given shape to the history of human civilization, have

often been positive or constructive – building societies, shaping cultures, and at other times negative or destructive– conflicts, wars and battles followed by reorganization of countries or nations. Human civilization has evolved in acomplicated manner and created a variety in ethnic characters and cultures, in languages, beliefs and customs. It iswidely understood [7] that historically, certain groups pay priority to their linguistic, cultural, racial, and religiousties of individuals within the groups, which are then passed down through generations. Ethnic violence arises whengroups attempt to maintain their boundaries against pressure from historical enemies. In another possible scenario,when the authority of a multi-ethnic state declines, the central regime ceases to protect the interests of ethnic groups,creating a void in which ethnic groups start competing to establish and control a new regime that will protect theirinterests. Socio-political tension in such a situation is likely to incite violence.Ethnic heterogeneity is a characteristic feature observed in most countries and regions worldwide. With time, these

variations often reduce locally, where socio-economic and political forces are less dominant in favor of ethnic mixing.However, the course of political history along with that of religion and culture often encounters conflicts at variousscales. Ethnic conflicts are comparatively frequent in certain regions and rare elsewhere. Ethnic heterogeneity doesnot necessarily breed war and its absence does not ensure peace. Even today, ethnic wars continue to be globallythe most common form of armed conflicts, but the mechanisms that lead a society down the path of ethnic conflictare yet to be fully understood. What is intriguing, is the connection between democratization and the occurrence

∗Email: kiran34 sit@jnu.ac.in†Email: sehgal.gunjan@tcs.com‡Email: bindu.gupta2@tcs.com§Email: geetika.s@tcs.com¶Email: arnab.chatterjee4@tcs.com∗∗Email: anirban@jnu.ac.in††Email: gautam.shroff@tcs.com

2

of ethnic conflict. While stable democracies are unlikely to wage war with other democracies, a country that issocio-politically unstable may very well find conflicts between groups with opposing interests [8]. Human rights areinternationally agreed values, standards or rules regulating the conduct of states towards their own citizens andtowards non-citizens. Interestingly, the violation of human rights appears to be more associated with ethnic conflictsthan abuses of economic and social rights [9]. Other political, economic and social preconditions may also influencethe causes of ethnic conflicts, and the conscious promotion by the political actors of any polarizing dimension basedon these factors is sufficient to lead to conflict. It is often seen that changes in political regime can lead to conflictswhich are often ethnic in various parts of the world, sparking human rights violations. Therefore, spatio-temporalanalyses of regional conflict formations and political dynamics, and the statistical studies of the different variablesare important.Temporal data of events like human to human communications and physical contacts/proximity has been studied

in recent years (see e.g., Holme and Saramaki [10]) in connection to study of epidemics and contagion, spreading ofinformation using mobile phone communication data [11] as well as data from proximity sensors carried by humanentities [12].Here we provide a quantitative analysis of the scale and topology of ethnic tensions, related conflicts and violations

of human rights, using the reports in digital media. We look at data from a publicly accessible database, which keepsaccount of events from news available in media. We particularly focus on ethnic conflicts (EC) and human rightsviolations (HR) by suitably filtering keywords present in the digital text transcript of the news. With the availability ofhigh precision data containing precise spatio-temporal information, one can look towards finding correlations betweenevents, involved actors (individuals, groups, organizations or states), the geographical pattern of spreading of conflictsetc. Although there have been studies on conflicts, wars and terror attacks [13–16], and speculation of ethnic conflictsusing census data for segregated population [17], a comprehensive study of the actors involved in ethnic conflicts andhuman rights violation has been lacking. In this work, for the first time, we provide a quantitative and qualitativeunderstanding of the relative activity of the actors, frequently engaging actor pairs, the network of actors, and giveinsights into the static and dynamical aspects of the actor network in a scenario involving ethnic conflicts and humanrights violations.

II. RESULTS

GDELT Event Database [18] contains the database of news articles from all over the world in several languages.Using a query, it is possible to extract data for events, about ethnic conflicts (EC) and human rights violations (HR)happening around the world, spanning over a large time scale. Each event data contains information such as the pairof actors involved, a unique time stamp, names of individuals, organizations or groups, the location information ofthe event, as well as latitude, longitude data of actors and the event. The news event and subsequently the dataserves as a proxy for the actual event and its intensity (in terms of number of reports). Although we do not analyzethe actual news reports, we consider a report on EC to be an actual instance of ethnic conflict or its possibility, andsimilarly for HR. We analyzed 28, 055 events of EC and 36, 470 of HR for a 15 year period (2001-2015).

0

20

40

60

80

100

2001 2004 2007 2010 2013 2016

n

time (year)

(b)HR

0

40

80

120

160

2001 2004 2007 2010 2013 2016

n

(a)EC

10-4

10-3

10-2

10-1

100

100 101 102

cum

ulat

ive

prob

abili

ty Q

(n)

news reports on a single day, n

(c)

ECHR

FIG. 1: The time sequence of the number of events n reported daily for (a) EC, (b) HR during 2001-2015. (c) The cumulativeprobability (CCDF) Q(n) that n or more events are reported on a particular day. The data for EC seems to fit well to a powerlaw for the largest values, with decay exponent around 2.54 ± 0.03, HR fits well to a stretched exponential (exp[−anb] witha = 0.44 ± 0.12 and b = 0.62 ± 0.01).

3

The number of news entries n per day is a stochastic variable, and often there is a burst of activity noted (Fig. 1a,b).The amount of data cataloged is also observed to have increased since the late 2000’s. The number of news entries/reports on a single day n has a broad distribution due to the large inter-day fluctuations in the number of reports andbursty nature of the data. Fig 1c shows the complementary cumulative distribution function (CCDF) that a day hasmore than n news reports, Q(n). For EC, the distribution has a short lognormal body with a rather prominent powerlaw tail (decay exponent close to 2.54 ± 0.03); for HR, the bulk of the distribution fits to a stretched exponential(exp[−anb] with a = 0.44 ± 0.12 and b = 0.62 ± 0.01) The details of the database, data extraction and differentattributes are described in Supplementary Information. In our study, we focus on a relatively recent time span(2001-2015) and extract the data for actor pairs and geographical location of events from the database.

A. Static properties of the network

We construct the aggregate network over a given period of time, for the entire 15 year period (2001-2015) as wellas for each of the 15 individual years, for the two datasets. Each event is visualized as a link between the actorsinvolved, which are the nodes, thus creating a network of actors connected by events. The details of the constructionis given in the Methods section. A schematic representation of the network construction is given in Fig. 2a and atypical network is shown in Fig. 2b for EC data aggregated over the year 2001.

��

�����

�����

���

�����

�����

���

�����

����

�

�����

����

��

��

�

�� ���

���

��

��

�

��

�����

����

�

�

��

�����

�����

��

��

�

��

��

��

ALB_AL

MKD_AL

CVL_AL

CAF_CT00GOV_CT18

HRV_HRMOS_HR

GOV_AL

ALB_MK00

ARM_AM

AZE_AJ00

NGO_IS

PSE_IS

HRVCVL_HR

ISR_ISISRGOV_IS

ALBCVL_AL

BTN_BT

nep_NP00CVL_NL

HRV_NL

UIS_RB

IGOEURSCE_ALMKD_MKD4

nep_BT

JUD_BK02

GOV_BK02

ALBUAF_ALMKDCVL_MK

srp_RB USA_US

srp_BK02

FRA_MKFRAGOVMIL_MK

BLR_BO00

AZE_BO00

UAF_AL

srp_HRHRV_BK02

GOV_IS

ALB_AL47CVL_AL47

ALBGOV_AL

MKD_MK00

CVL_RB00

RUS_RS48ALB_RB00

MIL_ALUIS_AL

MKD_MJ

EGY_EG11

ALBRAD_AL

JUD_RB

GOV_RB

ALB_AL45SCGSRB_AL45

NLD_BK02

COP_BK02

ALBCVL_AL47

CVL_MK00

ALBREB_MKE8

MKD_MKE8

MOS_RB00

srp_BK01

NLD_BKBEL_RB00

CVL_RSCIRUS_RSJA

TUR_CY

CYP_CY

JEW_US

CVL_US

GRC_CH22

CYP_CH22

PAK_PK03CVL_PK03

MOS_BK02

HRVCVL_RBRUSMIL_RS48

RWA_BYBDI_BY

IRQ_IZ05

kur_IZ05

FRA_FRA5MOS_FRA5

CHN_RBALB_RB

HIN_IN36UAF_IN36

BTNREF_BT

ALB_MKCOP_MK

ARM_AJ00

IRQMED_IZ14IRQ_IZ14

JUD_IS

MOS_USPSE_BK

srpCHR_RBIGOWSTNAT_RB

ALB_AL42

MOS_BM17MMR_BM17

NGA_NI26NGALEG_NI26

PAK_PK05

UZB_UZ

MKD_MK

IRN_IR15

IRQ_IZ07NGOHRIAMN_IZ07

CVL_IS

AFG_AF13

TJK_TI

REB_MKE8

IGOWSTNATMIL_MK00

CVL_BK02

GBR_UKAZE_AJ61

ITA_BM17

IGOMEAARL_EG11

CVL_RS48

INDGOV_PKPAK_PK

ALB_MKE8

SCGSRB_RI00

CVL_RIHRV_RI

PSE_GZ

IDN_ID13mad_ID13

EUR_SFZWEGOV_SF

GRC_GR

CYPCVL_CY04

MOS_UK

IGOUNO_RB

GOV_AF13AFGCVL_AF13

MKD_MKE3

AZE_AJ57

znd_AJ00

AFR_SF

ZWEGOV_ZI

IRQGOV_IZ07

ISRMED_US

ISR_US

CRM_RB

MOS_RB

ALBRAD_MK00GOV_MK00

IGOEURSCE_MKD4

GBRMED_UKU2

GBR_UKU2

CVL_MYPSE_MY

RUS_UK

ALBMIL_MK00

MKDGOV_MKE3

MIL_PL

COP_PL

ISR_BE

HRV_HR21

srp_HR21NLDJUD_RB

bos_RB

MIL_UKche_RS17

ALB_AL40ALBUAF_AL40

PAK_IN36

IRQGOV_IZ05

MIL_RB00

HRV_RB00

srp_USHRV_US

AZE_AJ09

MOS_BK

TUR_TUCYPGOV_TU

NLDJUD_BK02

GBRGOV_ITJUDGOV_IT

ITA_IT07NLD_IT07

GOV_CQCHN_CQ

USA_AF

NGA_NI00

GOV_RB00

srp_BK

COP_RB00

PAK_IN12UAF_IN12

IRQ_IR15

IRQGOV_IR15

ARM_AM08

ISRGOV_IS00

PSEMED_IS

IMGSEAMOS_RP21

MOS_RP21

CVL_RB

RUS_RSCI

HRVGOV_HR08

HRV_HR08

MOSSFI_IZ05

HRV_HR12

CAF_CT18

GOV_ID00

IDN_ID00

IRQ_IZ13NGOHRIAMN_IZ13

CVL_NI00

CVL_SF

GOV_IR26

IRN_IZ

IRQ_IZ

GOV_IS00

PSE_IS00

RUS_RS17

MED_MK

IRQGOV_IZ13

ARMCVL_AM

LEG_US

GOV_NI00

NGOHRIAMN_IR15kur_IR15

MKD_NL

ALB_UK

ZAF_SF02GOV_SF02

PSEGOV_IS

MKDLEG_MK00

day_ID13

ISR_IS05PSE_IS05

UISLEG_MK00MKD_CH

ZWE_ASAGR_AS

MOS_YI

HRV_YI

MKDCOP_MK00

MKD_MKE2CRM_AL

MKDGOV_MK00

CVL_HR

COP_AL

CVL_BKGOV_BK

USA_USDC

MIL_USDC

ALB_BKEUR_BK

srp_RB00

MOS_BK01

IGOWSTNAT_MK00

HRV_HR04

srp_HR04

AGR_MK00

MED_PK03

ITA_IT09

ALB_RS47

MKD_RS47

IGOWSTNAT_RB00IGOWSTNATMIL_AL

ALB_RS48

NGA_US

ALB_AL46

RAD_MK00

srp_AL

MED_IZ15kur_IZ15

HRV_FRA5

IRQ_IZ11IRQGOV_IZ11

COP_BK

GOV_TT

NLD_AL

PSE_US

kur_IZ13

RUS_US

CVL_ID13

USAGOV_USDC

MED_IN12kas_IN12

IGOUNOREFHCR_AL

ALB_AL48

MKD_GM

UAF_MK

CVL_MK

MOS_MK00

RUS_RS12RUSPTY_RS12

UIS_RB00

ISR_IS03

COP_MK00

MED_MK00

AFR_SF02

CVL_USDCPSE_USDC

COP_ID00

TUR_TU68

AZE_AM

MIL_RS17

ISR_SF02

PSE_SF02

SDN_SU29

nba_SU29

ISRMED_USDCISR_USDC

HRV_HR19

MOS_HR19

HRV_RI00

MOSSFI_IZ18

IRQ_IZ18PAK_PK04CVL_PK04

MED_AL

MOS_FICVL_FI

MKDGOV_MK

GOV_IZ05

CVL_NI26

AZECVL_AJ00

IRN_IZ11

LEG_NI26

LEG_NI43

NGA_NI43

NGO_GM

IGOMEAARL_USDC

SCGSRB_RB00

PSE_EG11

CVL_SF02

JUD_RB00

srp_SZHRV_SZ

MKDCVL_MK00

IRQ_UKH9

IRN_IZ12

srpCRM_RB

COP_RS48

PTY_BK02

ALBMED_AL

IRQGOV_IZ

HRV_BK01

RUS_AF30TJK_AF30

PSE_RS48

MIL_NI00

kur_SF

PAK_UZUZB_AF

PTY_HR21

GOV_RIMOS_RI

UIS_RINGOHRI_RS17

GOVMIL_BM17

NLD_RI

BEL_RI

IGOUNOREFHCR_RB00

IGOEUREECGOV_MK00

kur_IZ

INT_MK

ALB_YI

IGOWSTNAT_YI

MKDGOV_MKE2

GOV_NI26

MOS_HR12

MED_UKU2

UAF_GMPSE_WE00

ISR_IS00IGOWSTNAT_AL

CVL_IS00

SIK010_PK05

HIN_PK05

HRV_RB

HRVCVL_HR21

kur_IZ11

nai_USDC

JEW_IS

HRV_BK

MED_KESCGSRB_KE

HIN_IN12 EUR_USDC

CVL_NI43

RUS_RS

RUSGOV_RS

LBR_LIEDU_TS

UAF_USGAUSA_USGA

JUD_AF13

IGOWSTNAT_BK

UAF_MKD4

AZE_AJREF_AJ

IGOUNO_BK01JUD_BK01RUS_AF

AGR_ZA03

GOV_ZA03

ALBREB_AL

MIL_RI00MILMIL_RI00

GOV_NL07

REF_ID00

NLD_MKJUD_MK

IDN_ID04

mad_ID04

PAKUAF_PK03

SIK_PK05

IRQGOV_IZ18

GOV_LY

MED_LY

CYPGOV_CY04

GOV_USDC

ARM_AJ61

AGR_ZI04

GOV_ZI04

RUS_RB

JUD_US

GOV_US

IGOUNO_BK02

GOVHRI_ID13

GOV_SO02SOM_SO02

CHN_HR

RUS_RB00

IRQ_TUIRQGOV_TU

AFR_ZIZWE_ZI

NLD_RB00

MKDGOV_AL

MLI_ML

MOSSFI_TU68

IRQ_TU68

GOV_FR

GRC_CY

IRQ_US

MKD_BU52

MKDGOV_BU52

ALBUAF_MKD4

ALBRAD_RB00

HRVJUD_HR21

MKDEDU_AL

EDU_RB00

MED_USDC

MOSSFI_NL

IRQ_NL

CAF_CT

FRACVL_FRB9

MKDMED_MK

USALEG_USNCnba_SU56

EDU_LI

NLDGOV_NL07

IGOWSTNAT_SU

RUS_RS47

SCGSRB_RI

SDN_SU56

REF_AL40

ARMGOV_AJ00

SCGSRBMIL_RI00

ISR_GZCOP_IS00

idg_ID00

madREF_ID00

GOV_SU56

NGAGOV_NI45NGA_NI45

RUS_RS20

GOV_EN01CHN_EN01

HRV_NL07

haz_AF10

NGALEG_US

IDN_ID12COP_ID12

GOV_IS02ISR_IS02

UISLEG_CH

LEG_IS

CVL_AG

GOV_AF11

AFGCVL_AF11

ALB_AL50

NLDJUD_BK

JEW_IS00

CVL_HR04NLDJUD_HR04

INT_IN12CVL_IN12

ALB_MKD4

GOVREB_GR35

REB_GR35

MED_IZ14IRQ_IZ08

ALB_NL

CRMJUD_IS

GOV_HR04srp_USDCGOV_MX11

ZWE_ZI05

MOS_NL07

CVL_SO02

PSEREB_MK

UAF_PK

MMRELI_BM11MMRMIL_BM11

ALB_LYIGOUNOHRIHCHJUD_IS05

MKDEDU_MK

MOS_TU00

TURCVL_TU00

IGOWSTNAT_US

srp_NL11

NLD_NL11

TUR_GR

CYP_GR

IGOUNOHRIHCHJUD_AL

MIL_RI

ALB_RI

NGACVL_NI00

CVL_RSJA

zap_MX25

CVL_MX31

MEX_MX11AGR_MX11

HRVJUD_RB

MCOELI_ID13

day_ID04

zap_MX31

IRN_IZ05

RUSCVL_RSCI

EC−2001(b)

FIG. 2: (a) The schematic visualization of events on a time-line and the subsequent evolution of the aggregated network ofactors: Events appear sequentially. Event E1 involving actors A1, A2 is followed by event E2 involving A2, A3 and so on. Thelinks depict the actors involved together in an event. New events (actors in yellow and links in red) add to actor mention (sizeof node) and actor co-mention (link weight; thickness of links). (b) The network of actors for EC for 2001. Each actor is anode and any two actors ever involved in an event are connected by a link. The size of the nodes are proportional to the totalnumber of mentions in the period, and the relative thickness of the links are proportional to the total number of co-mentionsof the actors (link weights). The largest connected component (giant component) is depicted in red.

In our study, the temporal granularity of the data is one day. We extracted (i) the number of mentions m of eachindividual actor and (ii) the number of co-mentions w of an unique pair of actors (actor pair mentions), as well as thenumber of unique actors k, one actor is involved with. In terms of the network, m measures node strength, w the linkweight and k the degree of a node. While m and k measure the importance, activity or visibility of a single actor, wmeasures the involvement of an actor pair in inciting an event. These quantities are measured for each year for the15 year period (2001-2015) as well as through the whole period (aggregate over 15 years).It is often perceived that certain actors frequently engage in reported events than others and the same is true for

pairs of actors engaging in ethnic conflicts or human rights violations. We computed the complementary cumulative

4

probability distribution (CCDF) for different quantities for the aggregated network in 2001−2015 (refer to Fig. 3a,b,c).For the actor mentions, the probability that an actor is mentioned m times or more fits to a power law for the

✶�✲✁

✶�✲✂

✶�✲✄

✶�✲☎

✶�✲✆

✶�✵

✶�✵

✶�✆

✶�☎

✶�✄

✶�✂

◗❛✭✝✞

♠

❆✟✠✡☛ ☞✌✍✠✎✡✍✏ ✑✒✒☛✌✒✑✠✌✷��✶✓✷�✶✔

✕✖✗

❊✘❍✙

✚✛✜✢

✚✛✜✣

✚✛✜✤

✚✛✜✥

✚✛✜✦

✚✛✧

✚✛✧

✚✛✦

✚✛✥

✚✛✤

★✩✪✫✬✮

✇

✯✰✱✳✴ ✸✹✺✴ ✻✼✽✱✺✳✽✾ ✹✿✿✴✼✿✹✱✼❀✛✛✚❁❀✛✚❂

❃❄❅

❇❈❉❋

●■❏❑

●■❏▲

●■❏▼

●■❏◆

●■❏❖

●■P

●■P

●■❖

●■◆

●■▼

❘❙❚❯

❦

❱❲❳❨❩ ❬❭❪❩❭❭ ❫❪❪❩❭❪❫❳❭❴■■●❵❴■●❜

❝❞❡

❢❣❤✐

FIG. 3: Structural properties of the aggregated network in the period 2001-2015: (a) Plot of the cumulative probability (CCDF)Qa(m) that an actor is mentioned at least m times. Except for an exponential decay at the very end of the tail, most of thedistribution has a the power law decay Qa(m) ∼ m−ν1 with exponents ν1 as 1.23±0.01 for EC and 1.28±0.01 for HR. (b) Plotof the cumulative probability Qab(w) that an actor pair is mentioned at least w times. The tail of the distributions fit well topower laws Qab(w) ∼ w−ν2 with exponents as 1.58± 0.01 for EC and 1.74± 0.03 for HR. (c) Plot of the cumulative probabilityQ(k) that an actor is connected to k others or more. The tail of the distributions fit well to power laws Q(k) ∼ k−ν3 withexponents ν3 as 1.52 ± 0.01 for EC and 1.48 ± 0.01 for HR.

largest values, Qa(m) ∼ m−ν1 , with decay exponents 1.23± 0.01 for EC and 1.28± 0.01 for HR. The exponents arecomputed, along with the values of the standard errors (uncertainties) and significances, using Maximum LikelihoodEstimates [19]. The probability that an actor pair is mentioned w times or more fits well to power laws Qab(w) ∼ w−ν2

with exponents ν2 as 1.58±0.01 for EC and 1.74±0.03 for HR data. The actor mentions and actor pair mentions arerespectively the node weights and edge weights in network terminology. The probability that an actor was involvedwith k or more actors during the time span (degree of the actor) fit well to power laws Q(k) ∼ k−ν3 with exponentsν3 as 1.52 ± 0.01 for EC and 1.48 ± 0.01 for HR data. For the individual years, due to less aggregation, the countsare less and data naturally seems noisy. However, the data still seem to exhibit the power law tail in the probabilitydistributions (see Supplementary Fig. S7). The fitting exponents for the individual years are given in SupplementaryTable S1. The correlation between actor degree and mentions, as well as the relationship between the power lawexponents ν1 and ν3 are shown in Supplementary Fig. S8.The above results quantitatively characterize the heterogeneity in the activity of actors, while most actors are

relatively less active: the power law distributions for actor mentions Qa indicate that there are a significant few whoconstantly engage in ethnic conflicts and human rights violations issues. The power law distributions in actor pairmentions Qab indicate similar characteristic for pairs of actors. The broad degree distributions Q(k) are indicative ofthe fact that the number of actors engaging with very large number of actors are also significant.

B. Clusters

The aggregated network is found to be composed of several disconnected components or ‘clusters’. Physically thismeans that the actors constituting one cluster have never been involved with any actor from a different cluster. Forboth sets, the largest connected component is 102 − 103 times the smaller clusters, which are large in number (seeSupplementary Table. S2). In fact, the largest cluster s1 grows superlinearly with the size of the network N , as wefound (from the data in Supplementary Table. S2), s1 ∼ N δ, with δ = 1.17 ± 0.01 for EC and 1.19 ± 0.02 for HR(Supplementary Fig. S9). This tells us that the fraction of nodes in the largest cluster grows with the total sizeof the network quite fast so that eventually the fraction of nodes outside the largest cluster will be negligible. Wecomputed the CCDF Q(s) of cluster size s and find that it roughly has a power law decay for components except thelargest one, with a decay exponent roughly close to 3 (see Fig. 4). The CCDF for individual years has been shown inSupplementary Fig. S10. The largest clusters in both sets have a slowly decaying clustering coefficient 〈C(k)〉 withdegree k (see Supplementary Fig. S11).

C. Network growth properties

In order to explore the dynamics of the process that leads to the broad distribution of the mentions and degrees,we investigate the dynamics of growth for these quantities over the span of 15 years (2001-2015). For good statistics,

5

10-4

10-3

10-2

10-1

100

100 101 102 103 104

Q(s

)cluster size s

2001-2015

ECHR

FIG. 4: Plot of the cumulative probability (CCDF) Q(s) that there is a cluster of size larger than s, for data aggregated overthe period 2001-2015. The size of the largest cluster is very large compared to the rest.

we ranked the actors according to the number of mentions and degrees, and analyzed the data for the top 10 actors.We first measured the growth of the degrees and mentions for the top 10 actors of EC and HR data, with respect to

the time t∗ when they were first mentioned. The long time behavior (asymptotic) indicates a growth law of (t− t∗)β

with β ≃ 3.

✶�✲✁

✶�✲✂

✶�✲✄

✶�✵

✶�✄

✶�✂

✶�✁

✶�✹

✶�✵ ✶�✄ ✶�✂ ✶�✁ ✶�✹

♣✭☎✆

♠✝✞✟✠✡✞☛ ♠

❊☞

✌✍✎

■✏✑✒■✏��P✏✓✒■✏��

■✏✑✒■✏❆✑✔✒❆✔❆✕✓✒❆✖✑❘✏✒✑✏P✏✓✒■✏

❆✕✓✒❆✖�✗❆✘✙✒❆✘

✑❘✏✒✑✏✚✛✜✄

✢✣✤✥

✢✣✤✦

✢✣✤✧

✢✣★

✢✣✧

✢✣✦

✢✣✥

✢✣✩

✢✣★ ✢✣✧ ✢✣✦ ✢✣✥ ✢✣✩

✪✫✬✮

✯✰✱✳✴✷✱✸ ✯

❍✺

✻✼✽

❯✾✿❀❯✾❁❂❃❯✾❀❃✾❄❅

❯✾✿❀❯✾❇❃❈❀❇❃❉❋❃❯✾❀❃✾❇❃❈❀❇❃

❇✾❃❀❇✾✣✣❂●❈❀❂●❉❉❏❑❃❀❯▲●▼

❏❑❃❀❯▲◆✧

FIG. 5: Cumulative growth rates π(m) for mentions m for (a) EC and (b) HR datasets. The curves asymptotically fit toπ(m) ∼ ma with a > 1. The precise fitting exponents are given in Supplementary Table. S4.

To extract the rate of growth for a quantity x, we computed Π(x) = ∆x∆t

. Π(x) in real data turns out to be very

noisy, so we computed the cumulative integral π(x) =∫ x

0 Π(x′)dx′ which turns out to be less noisy. We observethat data for degree π(k) has a shorter span and is noisier than that for mentions π(m). The plots suggest thatboth π(m) and π(k) are superlinear functions of their respective arguments, i.e., π(x) ∼ xα with α ∼ 1.2 − 1.4 (seeFig. 5 for mentions and Supplementary Fig. S12 for degree). Thus the asymptotic growth rate Π(x) ∼ xα−1 is stillweakly dependent on the respective arguments. We thus identify that the growth rate of the system (network) isnot independent of the size of the node (degree or mentions). The growth exponents are computed using MaximumLikelihood Estimates [19] and tabulated in Supplementary Table. S4. A superlinear growth rate implies that thenetwork will, in the long run, consist of a very large ‘condensate’ with a small fraction of nodes in isolated smallerclusters, which is consistent with our analysis of the cluster size distribution.

D. Tolerance to attack and failure

We also study how the network breaks down under attack, in order to investigate the possibility of preventing unrulyevents to spread [20]. The largest connected component of the network is subjected to targeted attack by removal of

6

the most connected nodes. We start by removing the node with the highest degree, followed by the next highest andso on. This results in rapid fragmentation or destruction of the network. We compute the fraction of nodes G presentin the largest cluster, which is observed to decrease very quickly (Fig. 6a). In fact, the network of actors can bedestroyed by targeted attack just by removing much less than 10% of the nodes compared to when randomly selectednodes are removed (random failure) one after another (Fig. 6b). This exercise indicates that targeted interventionmay help stop spreading of ethnic conflicts and human rights violations.

0

0.5

1

1.5

2

2.5

0 0.05 0.1 0.15 0.2 0.25 0.3

G, <

c>

Attack(a)

EC, GEC, <c>

0

0.5

1

1.5

2

2.5

0 0.05 0.1 0.15 0.2 0.25 0.3

G, <

c>

fraction of nodes removed

Attack(b)

HR, GHR, <c>

0

0.5

1

1.5

0 0.2 0.4 0.6 0.8 1

G, <

c>

Failure(c)

EC, GEC, <c>

0

0.5

1

1.5

0 0.2 0.4 0.6 0.8 1

G, <

c>

fraction of nodes removed

Failure(d)

HR, GHR, <c>

FIG. 6: The structure of the network under attack: Nodes are removed in the sequence of their degrees starting from thehighest degree. The plot shows the behavior of the giant component G (fraction of nodes in the largest connected component)and the average number of nodes in the clusters other than the giant component 〈c〉, with increasing fraction of removed nodes,for (a) EC and (b) HR. The structure of the network under random failure: Nodes are removed randomly. Results are shownfor (c) EC and (d) HR networks. The networks are destroyed very quickly by targeted node removal (attack), compared torandom node removal (failure). The results are for networks aggregated over 2001-2015.

E. Measuring causality

Considering that we have time series data for the number of mentions of ethnic conflicts and human rights violationsfrom news articles it would be interesting to study if there is a clear cause and effect relation between these events withtime. To determine the causal effect purely from the observations of the past data, we apply Granger Causality [21]which estimates the causal relationship by observing the changes in the distribution of the variables over time.Let us consider two random variables depicting counts of EC and HR. To say that EC causes HR, Granger causality

computes a regression of variable HR on the past values of itself and the past values of EC and then tests the significanceof coefficient estimates associated with EC. We consider a bivariate linear autoregressive model on EC and HR, andassume the L.H.S. to be dependent on the history of EC and HR,

HRt = a0 + a1ECt−1 + . . .+ ahECt−h + b1HRt−1 + . . .+ bhHRt−h + Et

where h is the maximum number of lagged observations (for both EC and HR). The coefficients ai, bi are thecontributions of each lagged observation to the predicted value of ECt−i and HRt−i respectively and Et is theprediction error.We set up a null hypothesis and to test the significance of the coefficients, compute the p value. If the p-value is

less than 0.05, one can reject the null hypothesis. We tested the counts for year wise as well as month wise mentions.Comparing the p values for both cases, we found that: (a) one cannot conclude that HR causes EC, but (b) EC causesHR. Hence, we can confirm that ethnic conflicts cause human rights violations, while human rights violations are notresponsible for ethnic conflicts. The details of the analysis are provided in the Supplementary Information.

7

III. DISCUSSIONS

News reports serve as a reasonable proxy for the importance and intensity of events. The intensity is reflectedby the number of reports and lingering span of time through which the reports follow. Our study focuses on eventsthat pointed at ethnic conflicts and human rights violations. The GDELT data is unique in the sense that it recordsevents and actors involved in them. The U.S. government had recently funded in a large-scale project, the IntegratedConflict Early Warning System (ICEWS) [22], which makes use of quantitative data and statistical methods inorder to forecast events of political instability, which include international and domestic crises, ethnic and religiousviolence, rebellion and insurgency. We have chosen the GDELT data instead of the ICEWS data to get a more globalperspective [23, 24]. The frequencies of mentions of actors and their co-mentions with others can be treated wellas proxies for their importance and influence, as well as involvement with others. The aggregate data enables usto construct a network of actors, and even finding disconnected groups. In fact, our study reveals that most eventsare disconnected in very small clusters while very large clusters of frequently engaging actors exist. One can studythe geographical localization of events and clusters to procure detailed information regarding the context, intensityand growth pattern of events [25]. Identifying important groups of actors, in terms of their intensities of activities,is important for possible intervention that may prevent the spread of such events. The probability distributions ofactor mentions, co-actor mentions and the degree of an actor have power law tails for the largest values, indicating astrong self-organizing principle behind the events. The growth properties of individual actor nodes indicate that inthe long run, very small fraction of disconnected clusters are left, while most of the actors belong to a giant connectedcomponent. The data on ethnic conflicts and human rights violations are found to be strikingly similar in terms ofstatic and dynamic properties of the network, and even in terms of network stability against failure as well as targetedattack. This may point at a very high degree of correlation between events. In fact, using a causality analysis, wecould quantitatively conclude that ethnic conflicts lead to human rights violations, while the reverse may not be true.These networks of ethnic conflicts and human rights violations reflect the negative aspects of human behavior andcooperation, and are certainly different from other social networks (friendship, collaboration, etc.) in the sense thatthere are not many triangles or closed communities, as observed in the distribution of the average clustering coefficientwith degree of the actors. This detailed scientific study of the network structure, dynamics, function and resiliencemay help policy makers, specifically in cases where there is a need for preventing the spread of ethnic tensions andconflicts. There are possibilities of similar analyses using data from online social media, e.g., Twitter, etc.We have deliberately avoided mentioning the real names and other details because the main aim of the paper was

to study the network properties and statistical regularities. Since we have just considered the reports as a proxy forthe events and not studied the reports in details to extract further information, it would not be proper to draw anyconclusions about the actors, the network relations or the detailed explanations/causes behind the conflicts or theviolations. Certainly, it would be interesting to address such sociological explanations, implications and policies inthe future. Specific studies on the geographical implications of actor networks, the geographical and socio-culturalinfluence of their robustness under possible intervention will be important issues to study.

Methods

Data acquisition and filtering

GDELT Event Database [18] contains the database of news articles from around the world in several languages,hosted through Google Cloud. Using Google BigQuery [26], it is possible to extract data for each event, having anunique time stamp, and providing the data about news about ethnic conflicts (EC) and human rights violations(HR) happening around the world spanning over a large time scale. The data contains information about a pair ofactors involved, the location information of the event, as well as latitude, longitude data of actors and the event.We procured 45, 942 events for EC and 48, 295 for HR for a 15 year period, 2001-2015. We filtered out those datafor which both actors were mentioned, along with their respective location information. The GDELT data has ahuge fraction of missing entries. We have filtered out and excluded those data rows, which have at least one entrymissing corresponding to the attributes we were interested in. Hence, after cleaning, we analyzed 28, 055 events ofEC and 36, 470 of HR. Details of the Cameo codes used for filtering the data (along with examples from the CAMEOcodebook [27]), data attributes and cleaning are provided in Supplementary Information.

8

Network construction

Given a period of time T , we construct the network of ‘connected’ actors in the following way: any two actorsA1 and A2 mentioned together in an event E1 reported at time t ∈ [t0 : t0 + T ] are ‘connected’ by a link of unitweight,where t0 is the beginning of an interval of time span T . If another event E2 within the same time windowinvolves actors A2 and A3, then A3 is connected to A2 with a link of unit weight. Thus A1 and A3 are both connectedto A2 (see Fig. 2a). Aggregating all such events over the time window T , connected components emerge, with linkweights increasing if the same pair of actors linking them appear in multiple events (actor pair mentions). Theseconnected components form a complex network of nodes (actors) and links (actor pair mentions). An actor may beco-mentioned with several other actors and thus have a larger ‘degree’, measured by the number of distinct co-actorsit has. In principle, the network aggregated over a time period can have several disconnected components or ‘clusters’.Fig. 2b shows the aggregated network for one year for actors mentioned in the EC dataset.

Acknowledgments

The authors thank L. Dey for useful comments.

9

Supplementary Information

Data Description

The data source for this quantitative analysis on the global structure of ethnic violence is GDELT, the GlobalDatabase of Events, Language, and Tone [18]. It is described [24] as “an initiative to construct a catalog of humansocietal-scale behavior and beliefs across all countries of the world, connecting every person, organization, location,count, theme, news source, and event across the planet into a single massive network that captures what’s happeningaround the world, what its context is and who’s involved, and how the world is feeling about it, every single day”.GDELT contains data since 1979 [28]. However, the data from the year 2000 onwards is more comprehensive,

reflecting the increase in the number of news media and the frequency of event recording. The entire GDELT datasetis available as a public dataset in Google Big Query [26]. GDELT event records are stored in an expanded versionof the dyadic CAMEO format, capturing two actors and the action performed by Actor1 upon Actor2. A widearray of variables break out the raw CAMEO actor codes into their respective fields to make it easier to interactwith the data, the Action codes are broken down into a hierarchical structure, a score describing the intensity ofconflict or cooperation is provided, an average tone score is provided for all coverage of the event, several indicatorsof “importance” based on media attention are provided, and an unique array of geo-referencing fields offer estimatedlandmark-centroid-level geographic positioning of both actors and the location of the action.However, for the purpose of this quantitative analysis we have extracted the data for the cameo codes 203 (engage

in ethnic cleansing) for Ethnic Conflicts (EC) and 092 (investigate human rights abuses), 1122 (accuse of humanrights abuses) for Human Rights Violations (HR) from the years 2001 to 2015.As mentioned in the CAMEO codebook [27], for the corresponding entries, we reproduce below the cited examples:

• Cameo code 203 for EC: Serb forces were engaged in ethnic cleansing in Kosovo against the majority Albanianpopulation of the province, according to the US government.

• Cameo code 092 for HR: Israel’s high court opened a landmark hearing Wednesday into the legality of secretinterrogation techniques used against Palestinian detainees.

• Cameo code 1122 for HR: Human rights watchdog Amnesty International accused the United States of vio-lating human rights, ignoring international law and sending a “permissive signal to abusive governments”.

Each data entry with a GLobalEventID had the following attributes [27]:

• Actor1Code, Actor2Code: The complete raw CAMEO code for Actor1 and Actor2 (includes geographic, class,ethnic, religious, and type classes). It may be blank if the system was unable to identify any actor.

• Actor1Name, Actor2Name: The actual name of the Actor1 and Actor2. In the case of a political leader ororganization, this will be the leader’s formal name (e.g., GEORGE W BUSH, UNITED NATIONS), for ageographic match it will be either the country or capital/major city name (e.g., UNITED STATES / PARIS),and for ethnic, religious, and type matches it will reflect the root match class (e.g., KURD, CATHOLIC, POLICEOFFICER, etc). It may be blank if the system was unable to identify an actor.

• Actor1Geo ADM1Code, Actor2Geo ADM2Code: This is the 2-character FIPS10-4 country code followed by the 2-character FIPS10-4 administrative division 1 (ADM1) code for the administrative division housing the landmark.

• ActionGeo FullName: Location of Event. This is the full human-readable name of the matched location.

• ActionGeo CountryCode: Location of Event. This is the 2-character FIPS10-4 country code for the location.

• ActionGeo ADM1Code: Location of Event. This is the 2-character FIPS10-4 country code followed by the 2-character FIPS10-4 administrative division 1 (ADM1) code for the administrative division housing the landmark.

• ActionGeo Lat, ActionGeo Long: This is the centroid latitude and longitude of the landmark for mapping.

• SQLDATE: Date the event took place in YYYYMMDD format.

There may be actors with similar ActorCode at different locations. So, for uniquely identifying each actor we haveconcatenated Actor1Code with Actor1Geo ADM1Code and Actor2Code with Actor2Geo ADM2Code. The data acquiredfrom queries had to further cleaned for missing entries. Each event entry mentions a pair of unique actors, but we alsofound rare instances (< 0.5%) where only one actor has been identified. Any row of data with missing actor names,

10

actor codes or location data were removed to created the working data set. The number of events were 45, 942 for ECand 48, 295 for HR and this was reduced to 28, 055 for EC and 36, 470 for HR after filtering. On actual inspectionof the data, we found that a very small fraction of entries do not actually report an actual case of ethnic conflict orhuman rights violations, yet contain the relevant keywords that dicusss the issues in a positive tone (e.g. absence ofethnic violence or human right violations etc.).

The structure of the network

10-4

10-3

10-2

10-1

100

100 101 102

Qa(

m)

m

EC, Actor mentions

(a)

200120022003200420052006200720082009201020112012201320142015

10-4

10-3

10-2

10-1

100

100 101 102

m

HR, Actor mentions

(b)

200120022003200420052006200720082009201020112012201320142015

10-4

10-3

10-2

10-1

100

100 101 102

Qab

(w)

w

EC, Actor pair mentions

(c)200120022003200420052006200720082009201020112012201320142015

10-4

10-3

10-2

10-1

100

100 101 102

w

HR, Actor pair mentions

(d)200120022003200420052006200720082009201020112012201320142015

10-4

10-3

10-2

10-1

100

100 101 102

Q(k

)

k

EC, Actor degree

(e)

200120022003200420052006200720082009201020112012201320142015

10-4

10-3

10-2

10-1

100

100 101 102

k

HR, Actor degree

(f)

200120022003200420052006200720082009201020112012201320142015

FIG. S7: Plots of the cumulative probability (CCDF) Qa(m) an actor is mentioned m times or more, for (a) EC and (b) HR;Qab(w) that an actor pair is mentioned w times or more for (c) EC and (d) HR, and Q(k) that an actor is co-mentioned withk actors or more for (e) EC and (f) HR, for each year in the period 2001-2015. The actual fits are given in Table S1.

11

100

101

102

103

104

100 101 102 103

Men

tion

m

Degree k

(a)

EC raw dataEC binned data

fit100

101

102

103

104

100 101 102 103

Men

tion

m

Degree k

(b)

HR raw dataHR binned data

fit

FIG. S8: Scatter plots showing the degree (k) versus mention (m) (red) for each actor, in the networks for (a) EC and (b) HR.The raw data (red) are then log-binned and plotted as black filled circles. The dashed lines are the best fits obtained usingordinary least squares for the log-binned data (black filled circles): m ∝ kγ with γ measured to be 1.22 ± 0.01 for EC and1.16±0.02 for HR. The slope of each best fit line implies strong correlation between the degree (k) and mention (m) for actors,as also reflected by the ratios of the power-law exponents of the probability distributions of the respective variables in Fig. 3:ν3/ν1 ≈ 1.23 for EC and ≈ 1.16 for HR, compares well to corresponding values of γ.

TABLE S1: Computed exponents power law distribution fits for (i) actor mentions, ν1 (ii) actor pair mentions, ν2 and (iii)degree, ν3 from the data for Ethnic Conflicts (EC) and Human Rights Violations (HR) for different years (2001-2015) as wellas for the 15 year aggregate.

YearActor pair mention Actor mention Actor degreeEC HR EC HR EC HR

2001 2.61 ± 0.07 3.47 ± 0.35 1.56 ± 0.05 1.79 ± 0.04 2.13 ± 0.05 2.01 ± 0.062002 3.47 ± 0.23 3.55 ± 0.22 1.90 ± 0.06 2.06 ± 0.07 2.44 ± 0.07 2.16 ± 0.092003 3.04 ± 0.37 3.32 ± 0.35 1.71 ± 0.03 2.03 ± 0.07 2.01 ± 0.05 2.15 ± 0.042004 2.53 ± 0.11 3.81 ± 0.39 1.42 ± 0.02 1.90 ± 0.04 1.66 ± 0.04 2.06 ± 0.032005 3.35 ± 0.24 3.48 ± 0.18 2.15 ± 0.11 2.08 ± 0.05 2.29 ± 0.15 2.16 ± 0.102006 3.25 ± 0.20 3.14 ± 0.17 2.15 ± 0.07 1.99 ± 0.08 2.24 ± 0.09 2.30 ± 0.072007 3.04 ± 0.18 3.04 ± 0.14 1.64 ± 0.04 1.86 ± 0.07 1.85 ± 0.05 2.25 ± 0.052008 2.27 ± 0.07 2.28 ± 0.13 1.43 ± 0.03 1.74 ± 0.04 1.69 ± 0.03 2.11 ± 0.052009 2.07 ± 0.17 2.97 ± 0.09 1.50 ± 0.02 1.54 ± 0.03 1.93 ± 0.04 1.69 ± 0.052010 1.47 ± 0.10 2.61 ± 0.15 1.30 ± 0.03 1.49 ± 0.02 1.71 ± 0.03 1.62 ± 0.042011 1.67 ± 0.07 2.81 ± 0.07 1.32 ± 0.01 1.38 ± 0.03 1.62 ± 0.04 1.56 ± 0.032012 2.86 ± 0.13 2.60 ± 0.13 1.54 ± 0.02 1.52 ± 0.02 1.80 ± 0.03 1.72 ± 0.042013 1.92 ± 0.07 1.41 ± 0.05 1.40 ± 0.02 1.21 ± 0.02 1.86 ± 0.02 1.57 ± 0.022014 1.83 ± 0.04 2.48 ± 0.05 1.29 ± 0.01 1.37 ± 0.01 1.69 ± 0.02 1.60 ± 0.032015 1.82 ± 0.06 2.06 ± 0.07 1.31 ± 0.01 1.45 ± 0.01 1.68 ± 0.02 1.64 ± 0.01

2001-2015 1.58 ± 0.01 1.74 ± 0.03 1.23 ± 0.01 1.28 ± 0.01 1.52 ± 0.01 1.48 ± 0.01

12

Clusters

TABLE S2: Table showing the number of actors and the actors in the largest connected cluster for the different data sets:Ethnic Conflicts (EC) and Human Rights Violations (HR) for different years (2001-2015) as well as for the 15 year aggregate.

YearEC HR

Total nodesLargest

cluster sizeTotal nodes

Largestcluster size

2001 540 193 917 3332002 396 104 636 722003 355 117 788 1842004 633 329 823 2852005 263 48 714 1102006 451 56 907 1832007 721 288 1,102 3092008 1,279 696 1,375 4512009 1,204 544 2,267 1,0112010 951 474 1,887 9542011 1,412 746 2,719 1,5912012 1,644 899 2,423 1,1422013 1,728 892 2,707 1,5762014 2,560 1,620 3,447 2,1522015 3,432 2,229 3,816 2,322

2001-2015 10,394 7,875 15,899 12,106

The size of the largest cluster is computed as the maximum number of nodes s1 in the giant component or largestsubgraph and shown in Supplementary Table. S2 and the variation of s1 with the size of the entire network is shown inFig. S9, and the asymptotic fit is found to be s1 ∼ N δ. The degree distribution of the largest cluster / giant componenthas a power law tail. The power law exponents for the asymptotic fits are given in Supplementary Table. S3.

101

102

103

104

105

102 103 104 105

size

of l

arge

st c

lust

er s

1

total actors N

ECHR

EC fit, δ=1.17HR fit, δ=1.19

FIG. S9: Variation of the size of the largest cluster s1 with network size N , for different years (2001-2015) and well as 15 yearaggregate for EC and HR. The power law fits to s1 ∼ Nδ are δ = 1.17 ± 0.01 for EC and 1.19 ± 0.02 for HR.

13

10-3

10-2

10-1

100

100 101 102 103

Q(s

)

cluster size s

(a)

EC10-3

10-2

10-1

100

100 101 102 103

cluster size s

HR

(b)20012002200320042005200620072008

2009201020112012201320142015

FIG. S10: Plots of the cumulative probability (CCDF) Q(s) that there is a cluster of size larger than s, for different years(2001-2015) for (a) EC and (b) HR.

TABLE S3: Computed exponents ν3 of power law distribution to asymptotic fits to the degree distribution Q(k) ∼ k−ν3 forthe largest connected cluster (giant component) for the two categories of data – Ethnic Conflicts (EC) and Human RightsViolations (HR) for individual years and the 15 year aggregate (2001-2015).

Year EC HR2001 1.99 ± 0.06 1.71 ± 0.072002 1.84 ± 0.08 1.65 ± 0.052003 1.84 ± 0.08 1.56 ± 0.052004 1.58 ± 0.05 1.83 ± 0.062005 1.75 ± 0.29 1.48 ± 0.092006 1.50 ± 0.05 1.97 ± 0.082007 1.71 ± 0.06 1.84 ± 0.042008 1.64 ± 0.04 1.81 ± 0.082009 1.88 ± 0.05 1.56 ± 0.042010 1.62 ± 0.04 1.49 ± 0.032011 1.50 ± 0.03 1.41 ± 0.042012 1.75 ± 0.03 1.60 ± 0.042013 1.83 ± 0.02 1.53 ± 0.022014 1.67 ± 0.03 1.57 ± 0.042015 1.69 ± 0.02 1.60 ± 0.01

2001-2015 1.51 ± 0.01 1.48 ± 0.01

Biggest cluster: clustering coefficient

To find the cohesion of ethnic conflict and human rights violation, we computed the clustering coefficient of both(5 years aggregated) networks, which shows the extents to which the nodes of a network are closely connected withone another. The clustering coefficient for a node i in an undirected graph is computed as: Ci =

2Ni

ki(ki−1) , where ki

is the degree of the node i, and Ni is the number of links between the neighbors of i. We plot the average clusteringcoefficient of a node with degree k in Fig. S11, and find that higher degree nodes are less clustered compared to lowdegree nodes.

14

10-3

10-2

10-1

100

100 101 102 103<

C(k

)>

k

ECHR

FIG. S11: Plot of average clustering coefficient 〈C(k)〉 with degree k of the biggest cluster for the two categories of data – ECand HR for the 15 years aggregate (2001-2015). The data for all sets exhibit a slow decay with degree.

Dynamics of network growth

10-3

10-2

10-1

100

101

102

100 101 102 103

π(k)

degree k

EC

(a)

ISR_IS00PSE_IS00

ISR_ISARM_AMRUS_RSALB_ALPSE_IS

RUS_RS48AZE_AJ09

AZE_AJm1

10-3

10-2

10-1

100

101

102

103

100 101 102 103

π(k)

degree k

HR

(b)

USA_USDCUSA_US

RUS_RS48IRN_IR26

GBR_UKH9IRN_IR

RUS_RSGBR_UKISR_IS00

CHN_CH22m1

FIG. S12: Plots showing the cumulative growth rates π(k) for degree k for (A) EC and (B) HR datasets. The curves fit toπ(k) ∼ kα with α > 1. The precise fitting exponents are given in Table. S4.

TABLE S4: Computed exponents of asymptotic power law fits for the cumulative growth rates π(x) ∼ xα for the top ten highestcounts for (i) actor mentions, (ii) degree, for the 2 categories of data – Ethnic Conflicts (EC) and Human Rights Violations(HR) for the 15 year aggregate.

EC HRActor Degree Mention Actor Degree Mention

ALB AL 0.86 ± 0.01 1.04 ± 0.01 CHN CH22 0.71 ± 0.02 1.11 ± 0.01ARM AM 1.55± 0.02 1.51± 0.01 GBR UK 1.11± 0.02 1.33± 0.01AZE AJ 0.50± 0.02 0.96± 0.01 GBR UKH9 1.68± 0.02 1.79± 0.01AZE AJ09 0.96± 0.04 1.28± 0.01 IRN IR 0.57± 0.01 1.10± 0.01ISR IS 0.89± 0.02 1.23± 0.01 IRN IR26 1.58± 0.03 1.47± 0.01ISR IS00 1.26± 0.01 1.38± 0.01 ISR IS00 1.86± 0.02 1.71± 0.01PSE IS 1.05± 0.02 1.31± 0.01 RUS RS 1.27± 0.02 1.41± 0.01PSE IS00 1.82± 0.02 1.40± 0.01 RUS RS48 1.04± 0.02 1.22± 0.01RUS RS 0.47± 0.01 0.92± 0.01 USA US 1.33± 0.02 1.40± 0.01RUS RS48 0.53± 0.02 0.79± 0.01 USA USDC 1.29± 0.01 1.35± 0.01

15

Measuring causality

Let us consider the two random variables depicting counts of EC (Ethnic Conflicts mentions) and HR (HumanRights Violation mentions). To say that EC causes HR, Granger causality [21] computes a regression of variable HRon the past values of itself and the past values of EC and then tests the significance of coefficient estimates associatedwith EC.We consider a bivariate linear autoregressive model on EC and HR, and assume the L.H.S. to be dependent on the

history of EC and HR,

HRt = a0 + a1ECt−1 + . . .+ ahECt−h + b1HRt−1 + . . .+ bhHRt−h + Et (1)

where h is the maximum number of lagged observations (for both EC and HR). The coefficients ai, bi are thecontributions of each lagged observation to the predicted value of ECt−i and HRt−i respectively while Et is theprediction error.If b1 = b2 = . . . = bh = 0, we call it a null hypothesis HR0 which implies that EC does not cause HR. In other

words the coefficients of EC are not significant enough to cause HR. But if the null hypothesis gets rejected we saythat the coefficients of EC are significant enough to cause HR.For testing this significance of the coefficients, we compute the p-value. If the p-value is less that 0.05 one can reject

the null hypothesis, and hence conclude that EC causes HR (HR ∼ EC).Applying the above process on the EC and HR year wise mentions, we found that:

• On testing EC ∼ HR for h = 4, we find p ≃ 0.721, and the null hypothesis cannot be rejected. Hence, onecannot conclude that HR causes EC.

• On testing HR ∼ EC for h = 4, we find p ≃ 0.029, and the null hypothesis can be rejected. Hence, we can saythat EC causes HR.

Applying the above process on the EC and HR month wise mentions, we found that:

• On testing EC ∼ HR for h = 5, we find p ≃ 0.193 and thus the null hypothesis cannot be rejected. Hence, wecannot say that HR causes EC.

• On testing HR ∼ EC for h = 5, we find p ≃ 0.036, and thus the null hypothesis can be rejected. Hence, we cansay that EC causes HR.

Hence, we can definitely conclude from the above quantitative analysis that Ethnic conflicts cause Human Rightsviolations.

[1] M. E. J. Newman, A.-L. Barabasi, and D. J. Watts. The structure and dynamics of networks. Princeton Univ. Press,Princeton, 2006.

[2] R. Albert and A.-L. Barabasi. Statistical mechanics of complex networks. Rev. Mod. Phys., 74(1):47–97, 2002.[3] A.-L. Barabasi. Network Science. Cambridge Univ. Press, Cambridge, 2016.[4] D. Lazer, A. Pentland, L. Adamic, S. Aral, A.-L. Barabasi, D. Brewer, N. Christakis, N. Contractor, J. Fowler, M. Gutmann,

T. Jebara, G. King, M. Macy, D. Roy, and M. Van Alstyne. Computational social science. Science, 323(5915):721–723,2009.

[5] C. Castellano, S. Fortunato, and V. Loreto. Statistical physics of social dynamics. Rev. Mod. Phys., 81:591–646, May 2009.[6] P. Sen and B. K. Chakrabarti. Sociophysics: An Introduction. Oxford University Press, Oxford, 2013.[7] M. D. Toft. The geography of ethnic violence: Identity, interests, and the indivisibility of territory. Princeton Univ. Press,

Princeton, 2005.[8] J. Vorrath, L. F. Krebs, and D. Senn. Linking ethnic conflict & democratization: An assessment of four troubled regions.

National Center of Competence in Research, Challenges to Democracy in the 21st Century, Working Paper, (6), 2007.[9] O. N. T. Thoms and J. Ron. Do human rights violations cause internal conflict? Human Rights Quarterly, pages 674–705,

2007.[10] P Holme and J Saramaki, editors. Temporal networks. Springer, 2013.[11] J.-P. Onnela, J. Saramaki, J. Hyvonen, G. Szabo, D. Lazer, K. Kaski, J. Kertesz, and A.-L. Barabasi. Structure and tie

strengths in mobile communication networks. Proc. Nat. Acad. Sci., 104(18):7332–7336, 2007.[12] C. Cattuto, W. Van den Broeck, A. Barrat, V. Colizza, J.-F. Pinton, and A. Vespignani. Dynamics of person-to-person

interactions from distributed RFID sensor networks. PloS ONE, 5(7):e11596, 2010.[13] L. F. Richardson. Statistics of deadly quarrels. Boxwood Press, Pittsburgh, 1960.

16

[14] A. Clauset, M. Young, and K. S. Gleditsch. On the frequency of severe terrorist events. J. Conflict Resol., 51(1):58–87,2007.

[15] O. Becerra, N. Johnson, P. Meier, J. Restrepo, and M. Spagat. Natural disasters, casualties and power laws: A comparativeanalysis with armed conflict. In Proc. Ann. Meet. Am. Political Sc. Assoc. Citeseer, 2012.

[16] A. Chatterjee and B. K. Chakrabarti. Fat tailed distributions for deaths in conflicts and disasters. Rep. Adv. Phys. Sc.,1:1740007, 2017.

[17] M. Lim, R. Metzler, and Y. Bar-Yam. Global pattern formation and ethnic/cultural violence. Science, 317(5844):1540–1544, 2007.

[18] The GDELT Project, retreived October, 2016. www.gdeltproject.org/.[19] A. Clauset, C. R. Shalizi, and M. E. J. Newman. Power-law distributions in empirical data. SIAM review, 51(4):661–703,

2009.[20] R. Albert, H. Jeong, and A.-L. Barabasi. Error and attack tolerance of complex networks. Nature, 406(6794):378–382,

2000.[21] C. W. J. Granger. Investigating causal relations by econometric models and cross-spectral methods. Econometrica,

37:424–438, 1969.[22] M. D. Ward, A. Beger, J. Cutler, M. Dickenson, C. Dorff, and B. Radford. Comparing GDELT and ICEWS Event Data.

Analysis, 21:267-297, 2013.[23] B. Arva, J. Beieler, B. Fisher, G. Lara, P. A. Schrodt, W. Song, M. Sowell and S. Stehle. Improving Forecasts of

International Events of Interest. Working paper, 2013.[24] K. Leetaru and P. A. Schrodt. GDELT: Global data on events, location, and tone, 1979–2012. In ISA Ann. Convention,

volume 2. Citeseer, 2013.[25] B. Gupta, G. Sehgal, K. Sharma, G. Sharma, A. Chatterjee, A. Chakraborti, and G Shroff. Spatial analysis of networks

of ethnic conflicts and humans rights violations. in preparation.[26] World’s largest event dataset now publicly available in BigQuery. https://cloudplatform.googleblog.com/2014/05/worlds-largest-event-dataset-now-publicly-available-in-google-bigquery.html.[27] GDELT - DATAFORMATCODEBOOKV 1.03, as on 8/25/2013. http://data.gdeltproject.org/documentation/GDELT-Data_Format_Codebook.pdf.[28] J Dana Stuster. Mapped: Every protest on the planet since 1979. ForeignPolicy. com, 2013.