GETTING HELP IN A CROWD

A Social Emergency Alert Service

Andreas Geyer-Schulz, Michael Ovelg¨onne and Andreas C. Sonnenbichler

Information Services and Electronic Markets, Institute of Information Systems and Management

Karlsruhe Institute of Technology (KIT), D-76128 Karlsruhe, Germany

Keywords:

Emergency recommender, Social network analysis, Modularity clustering, Restricted random walk clustering,

Bystander effect.

Abstract:

This paper presents the conceptual design of an emergency alert service which addresses the problem of getting

help in a crowd. The design is based on the results of more than 30 years of research in social psychology on

the bystander effect and it addresses the obstacles for helping by directing an alert directly to the members of

the victim’s social group who happen to be near the location of the emergency event. The proposed emergency

recommender design combines the general availability of geo-location services with social clusters available

either from the analysis of social web-sites or from communication analysis. For this purpose, two recently

developed innovative clustering methods are used. The feasibility of the design is evaluated by simulating

emergency events on the MIT reality mining data set.

1 INTRODUCTION

On Saturday, the 12th of September 2009, 50-year

old Dominik Brunner was brutally murdered in a Mu-

nich S-Bahn station (Spiegel Online, 2009b). The

attack on Dominik Brunner was observed by 15 pas-

sengers (Spiegel Online, 2009c) and transmitted and

recorded by his mobile-phone on the open police

emergency channel (Spiegel Online, 2009a).

This tragic incident reminds social psychologists

of the murder of Kitty Genovese on March 13th, 3:20,

1964 in Queens, New York. Kitty Genovese was

stalked, stabbed, and sexually assaulted near her own

apartment building. During the attack on her she

screamed and broke free twice. 38 of her neighbors

witnessed the attack, but no one intervened. After 45

minutes one man called the police, but at this point in

time Kitty Genovese was dead.

This incident motivated the study of social pro-

cesses in emergency situations by Darley and Latan´e

(Latan´e and Darley, 1970) and it points to the short-

comings and problems of real emergency response

organizations and their management which very re-

cently have become the object of scientiﬁc research

e.g. (Yuan and Detlor, 2005), (Faraj and Xiao, 2006),

(Chen et al., 2008), and (Comfort et al., 2009). The

mobile phone of the victim transmitting and record-

ing to the end conﬁrms Palen and Liu’s thesis of the

increasing availability of ICT and its use in an emer-

gency by citizens and also their observation, that the

traditional hierarchical command-and-control report-

ing system of emergency response organizations may

not be adequate and “does not include built-in consid-

erations for the important roles that members of the

public play as participants” (Palen and Liu, 2007, p.

729). Public participation in emergencies and disas-

ters is active and altruistic. First responders are of-

ten not the trained professionals of an emergency re-

sponse organization who are sent to the incident, but

the people from the local and surrounding communi-

ties. They provide ﬁrst-aid, transport victims to the

hospital, and begin search and rescue (Palen and Liu,

2007, pp. 728-729). In the case of Dominik Brunner

we may speculate whether a fast activation of his so-

cial community parallel to the emergency call could

have saved his life.

The paper starts with a short review of the social

processes which Darley and Laten´e have identiﬁed

as obstacles for helping in emergencies in section 2.

These obstacles have become known as the bystander

effect: The more bystanders, the less likely the victim

will receive help. We address the bystander effect by

a social emergency alert service and discuss how so-

cial emergency alert services may help in improving

these processes in emergency settings by activating

the nearby members of the victim’s social network.

207

Geyer-Schulz A., Ovelgöenne M. and S. Sonnenbichler A. (2010).

GETTING HELP IN A CROWD - A Social Emergency Alert Service.

In Proceedings of the International Conference on e-Business, pages 207-218

DOI: 10.5220/0002982102070218

 SciTePress

Step 1

Notice that

something

is happening

Step 2

Interpret

event as an

emergency

Step 3

Take

responsibility

for helping

Step 4

Decide

how to

help

Emergency

Step 5

Provide

help

Distraction

Self-concerns

Stop fooling around kids,

we are here to eat.

I’m late for a very

important date.

Audience

inhibition

Costs

exceed

rewards

I’ll look like a fool.

What if I do something

wrong? He’ll sue me.

Lack of

competence

I’m not trained

to handle this,

and whom can I call?

Diffusion of

responsibility

Someone else must

have called the

police.

Ambiguity

Is she really sick

or just drunk?

Relationship

between

attacker and

victim

They have to resolve

their own family

problems.

Pluralistic

ignorance

No one else

seems

concerned.

Figure 1: The social help process: 5 steps leading to assistance.

For the process of giving and receiving help, we

propose to monitor social interactions and to iden-

tify the social clusters of the victims and to locate

the nearest members of the social cluster of the vic-

tim in an emergency for notiﬁcation purposes. In sec-

tion 3 we present details on the social emergency alert

service for getting help in a crowd and in section 4

we discuss implementation variants of realizations of

such services based on readily available technology

by the telecommunication and Internet industry.

The architecture for the prototype of the service

which is currently implemented is presented in sec-

tion 5.

For social cluster identiﬁcation, recently devel-

oped fast algorithms for graph clustering are pre-

sented in section 6.

Section 7 of this paper aims at assessing the

chances that the social emergency alert service pre-

sented has in reality. For this purpose, a ﬁrst attempt

is made to answer three questions which play a cru-

cial role for the success of a social emergency alert

service:

1. Has the victim in an emergency a chance to trans-

mit a request for help?

2. Is someone of his social network nearby?

3. Will this person really help?

2 THE UNHELPFUL CROWD:

FIVE STEPS TO HELPING IN

AN EMERGENCY

The murder of Kitty Genovese in 1964 in Queens,

New York, in front of 38 witnesses who did not in-

terfere led Darley and Latan´e to start investigating

the social psychological processes at work in this in-

cident. Their research revealed that the more by-

standers, the less likely the victim will be helped. This

is the bystander effect: The presence of others inhibits

helping. One of the reasons for this is the diffusion of

responsibility (Darley and Latan´e, 1968). Darley and

Latan´e (Latan´e and Darley, 1970) provided a care-

ful analysis of the process of emergency intervention

shown in ﬁgure 1. Unfortunately, many obstacles to

helping must be overcome. At each step, however,

psychological factors are at work which explain why

people fail to help:

1. Noticing. The presence of others distracts at-

tention from the victim. People who live in big

cities may ﬁlter out people lying on sidewalks or

screams (stimulus overload (Milgram, 1970)).

2. Interpreting. People must interpret the meaning

of what they perceive. Their perception may be

ambiguous: Cries of pain may be taken for laugh-

ter, hypoglycemia may be taken for drunkenness

(e.g. (Clark and Word, 1972), (Piliavin et al.,

1975)). A perceivedrelationship between attacker

and victim may lead the observers to think that

everything is OK (Shotland and Straw, 1976). If

an emergency happens, the most powerful infor-

mation available is often what other people do.

However, if everybody looks on everybody else

to get clues on what to do, the whole group is suf-

fering from pluralistic ignorance and everybody

concludes that help is not needed (e.g. (Miller and

McFarland, 1987), (Monin and Norton, 2003)).

3. Take Responsibility. When help is needed, who

is responsible for providing it? The diffusion of

responsibility means that people belief that oth-

ers will or should help. The effect usually is

strengthened by anonymity and considerably re-

duced by a reduction in psychological distance.

Groups with members who know each other are

more helpful than strangers. See e.g. (Garcia

et al., 2002), (Rutkowski et al., 1983), (Baumeis-

ter et al., 1995).

ICE-B 2010 - International Conference on e-Business

208

4. Decide how to Help. Bystanders are more likely

to offer direct help if they feel competent to per-

form the actions required (e.g. (Shotland and

Heinold, 1985), (Cramer et al., 1988)).

5. Provide Help. Some people may feel too em-

barrassed to provide help in a public setting (au-

dience inhibition). However, when people think

they will be scorned by others for failing to pro-

vide help, the presence of an audience will in-

crease their helpful actions. See (Schwartz and

Gottlieb, 1980).

In addition, a series of other variables have a high

inﬂuence on helping behavior as experiments in so-

cial psychology have shown: Time pressure reduces

the tendency to help (e.g. (Darley and Batson, 1973),

(Batson et al., 1978)). Group membership and em-

pathy and attractiveness interact: empathy is a posi-

tive predictor for help for in-group members, whereas

attractiveness works for out-group members (St¨urmer

et al., 2005). Group membership positively inﬂuences

help for in-group members, and the group bound-

aries can be shifted by proper priming (Levine et al.,

2005). Group status and group identiﬁcation inﬂu-

ence the willingness of receiving help (Nadler and

Halabi, 2006).

But what can you do to receive help in a crowd?

Try to counteract the ambiguity of the situation by

making it clear that you need help, and try to reduce

the diffusion of responsibility by addressing a speciﬁc

individual for help, keep eye contact, point or direct

requests (e.g. (Moriarty, 1975), (Shotland and Steb-

bins, 1980)). Consistent with this is a recent study

of P. Markey (Markey, 2000) of people in Internet

chat rooms: If the number of individuals is large in a

chat room, individuals react slower to a plea for help.

However, addressing a speciﬁc individual by his name

leads to considerably faster help and eliminates this

effect.

Research on the bystander effect in social psychol-

ogy showed that even weak social links matter and in-

crease the chance of a victim to receive help consid-

erably. This fact is the main motivation to send alerts

to the geographically close members of the victim’s

social group.

The asymmetric perception of social links (e.g.

(Hoser and Geyer-Schulz, 2005)), the role of weak

ties, and the cultural norms of the community play a

major role in the formation of the social group of the

victim. The asymmetric perception of social links im-

plies that a person may not be really aware of possible

helpers in his loose social contacts. Taken together

with cultural norms, even professional acquaintances

are potential helpers. The role of weak ties for net-

works has been studied by M. Granovetter ((Gra-

novetter, 1973) and (Granovetter, 1982)). In the con-

text of information diffusion on open jobs Granovet-

ter observed that “it is remarkable that people receive

crucial information from individuals whose very exis-

tence they have forgotten” (Granovetter, 1973, p. 14).

This is an indication that an explicit list of emergency

contacts provided by the subscriber of such an emer-

gency service will considerably limit the effectiveness

of such a service, because of these social phenomena.

3 GETTING HELP IN A CROWD

- A SOCIAL EMERGENCY

ALERT SERVICE

Getting help in a crowd is a social emergency alert

service selecting nearby members of the social group

of the victim and notifying them about the victim’s

need for help and the victim’s location. With this

service we aim to activate the locally available social

network of the victim and to eliminate the bystander

effect.

The UML sequence diagram depicted in ﬁgure 2

shows the generic process in an emergency incident.

It is designed on a high-level, abstract way allowing

a variety of industrial implementations. We will ad-

dress this issue in section 4.

An emergencynotiﬁcation is submitted by the vic-

tim by starting an application on his mobile device

(

LaunchEmergenyApplication

in ﬁgure 2) e.g. by

pressing the help-button shown in ﬁgure 5. The ap-

plication retrieves the current geo-position. Both,

the ID of the emergency caller and his geo-location

are then transmitted to the emergency alert service

(

EmergencyAlert

in ﬁgure ﬁgure 2).

For discovering the most likely person to help

in the victim’s social network, his social network

has to be known and possible helpers identiﬁed

(

GetHelperCluster

in ﬁgure 2). The network is ei-

ther built from social interaction data from e-mail,

sms, phone, and mobiles where the number of in-

teractions is taken as an indication of social near-

ness or from social web sites as for example Face-

book or Xing. However, the number of social interac-

tions may be ambiguous as a recent incident (Rungg,

2010) of a woman threatened by her former husband

with Googles Buzz has shown: So the possibility to

check for such unwanted relations must be provided

for the participants of the emergency alert service

(

FilterCandidates

in ﬁgure 2). Usually social net-

works tend to be very large. As emergencies are often

time-critical, it might take too long to calculate such a

network on-the-ﬂy. The network is pre-built and up-

GETTING HELP IN A CROWD - A Social Emergency Alert Service

209

Figure 2: Sequence Diagram of a Generic Social Emergency Alert Service (in-incident).

dated regularly for all service subscribers. To ﬁnd out

which persons in the social network are likely to help,

a clustering of the network is performed. Details for

this clustering are addressed in section 6.

Next, the current geo-position of the can-

didates is retrieved from a geo-position service

(

GetGeoPositionofCandidates

in ﬁgure 2). The

alert service uses the geo-data as a ﬁlter on the vic-

tim’s social cluster to ﬁnd out, who of the possible

helpers is locally close enough. Section 7 deals with

details of having at least one member of the victim’s

social cluster in range.

The possible helpers in range and the emer-

gency response center are informed about the emer-

gency situation of the victim (several invocations of

HelpRequest

in ﬁgure 2). All possible helpers in

range are informed simultaneously. The victim’s

name, his geo-location and the shortest route as well

as the contact data for the emergency response center

are provided.

Finally, possible helpers and the emergency re-

sponse experts at the police’s emergency response

center may communicate, because of the information

forwarded by the emergency alert service (not shown

in ﬁgure 2). This facility has the potential of provid-

ing expert guidance to the socially close ﬁrst respon-

ders on the scene. However, it also reveals the identity

of potential helpers to the emergency response center.

The privacy impact of this must be addressed for such

a service.

In section 2 obstacles to the ﬁve steps leading to

assistance have been described. The emergency alert

service presented addresses these obstacles directly:

1. The distraction obstacles can be avoided by notic-

ing, that an emergency incident takes place: Clear

signal words are part of the personal message to

the helpers. This makes it obvious, an emergency

case is happening and this is made clear to the

helper.

2. Self concerns are also addressed by the service:

Since the potential helper is directly addressed

and others know this from the incident protocol,

social norms lead to pressure to help.

3. Ambiguity is by-passed by the clear and unmistak-

able help request sent to the helpers.

4. As this message is personal, pluralistic ignorance

is eliminated. Experimental evidence for these

effects in internet chat rooms is provided by P.

Markey (Markey, 2000).

5. The relationship between attacker and victim can

not lead the helper to overlook the emergency

event, because of the unambiguous emergency

message.

6. Diffusion of responsibility is also reduced, since

the emergency alert message is directly and per-

ICE-B 2010 - International Conference on e-Business

210

sonally addressed to the helper. Because of this,

he is responsible and because of the incident pro-

tocol, others will know this and hold him respon-

sible.

7. Lack of competence may be addressed by provid-

ing fast expert backup for helpers from the police

emergency response center.

8. Expert guidance of how to help also addresses the

problems of audience inhibition and costs exceed

rewards.

4 IMPLEMENTATION VARIANTS

In this section we present implementation examples

how the Social Emergency Alert Service can be im-

plemented in an industrial environment.

4.1 Emergency Alerting

To be able to use the service, the user has to possess

a mobile device (e.g. a mobile phone). He can then

subscribe to the service. In case of an emergency, he

starts an application on his mobile device. Of course,

the start of the application must be made simple and

fast, as we do not expect it likely that the victim is able

to deal with complex applications in an emergency

situation. For the implementation third-party plat-

forms like Android can be used. Android (Android,

2010) is a mostly free and open-source OS platform

developed and driven by the Open Handset Alliance

(Open Handset Alliance, 2010). Further platforms

like Apples iPhone may be supported as well. Special

mobile devices combined with body-sensors, e.g. for

elderly people, linking the start of the application to a

hardwired button can be offered.

4.2 Geo-position Service

Geo-positions of both the victim and all possible

helpers of the victim’s social network need to be cal-

culated. Many of todays mobile phones are able to

calculate their geo-position by GPS (Global Position-

ing System). The service Google Latitude is an ex-

ample for a service that users can publish their cur-

rent geo-position and share it with friends. If a mo-

bile device does not include such a feature, several al-

ternative techniques have been described and imple-

mented. Even speed vectors can be calculated (for

example (Kikiras and Drakoulis, 2004; Borkowski

and Lempiainen, 2006)). By this, the expected geo-

position of somebody moving in a train can be found

out.

4.3 Building Social Networks and

Identifying Possible Helpers

For discovering the most likely person to help in an

emergency case the social network has to be known.

We present three possible realizations.

The social network can be built by monitoring

outgoing and incoming calls on the mobile device

of the subscriber. The emergency alert application

running on the mobile device collects this call data,

pools it and regularly (e.g. once a week) transmits it

via HTTP to the social cluster service (see ﬁgure 2).

There the call logs of all service subscribers are com-

bined and the network is calculated: Telephone num-

bers are represented as nodes, the calls are weighted

ties. Each call strengthens a tie. The advantage of

this solution is, that the network is independent from

the telephone provider. It works depending just on the

emergency alert application. The disadvantage is, that

the calculated social network consists only of sub-

scribers and their direct connections. Ties between

non-subscribers can not be observed technically.

Alternatively not the mobile devices monitor

the calls, but connection records from telephone

providers are used. Connection records are stored for

billing purposes. In the European Union an directive

forces the provider to save call logs from six month

up to two years (EU, 2006). These connection data

can be used to calculate the social networks. Every

connection is represented by a tie between the calling

parties (more concrete, their telephone numbers) as

nodes. Of course, the resulting network will be huge.

In section 7 we will show that feasible cluster algo-

rithms for such huge data sets exist. The advantage

of this solution is, that much more network data can

be collected so that the problem of missing links is

smaller. On the other hand this alternative can only be

realized if the calling logs are available to the emer-

gency alert provider. As we do not expect network

providers to give such information away, the most

likely approach for this alternative is, if the network

provider is identical with the emergency alert service

provider. The provider can then use the service as

an additional opt-in feature. Another disadvantage is,

that one network provider will probably not exchange

network or call log data with other providers. By this

the social network is limited to the customers of the

provider plus their direct links.

As an additional feature for both alternatives ad-

dress books in the customers’ devices can be used to

group telephone numbers. Different telephone num-

bers of one person can be combined and fused to one

node in the social network.

A third approach to build the social network is

GETTING HELP IN A CROWD - A Social Emergency Alert Service

211

to cooperate with existing social network platforms.

Data from Facebook, Xing, Myspace can be used.

The advantage of this solution is, that no subscriber or

network provider boundaries exist. The disadvantage

is, that people tend to accept more ’friends’ in social

platforms than they would accept ofﬂine. Addition-

ally most of these platforms do not weight their ties,

which makes it difﬁcult to apply cluster algorithms.

In practice all three alternatives used to build a so-

cial network as discussed above can be complemented

with a list of emergency contacts provided by each

subscriber and, if available, with a list of dedicated

helpers for an event or for a community. In a Ger-

man small rural community, the community’s volun-

tary ﬁre-ﬁghters are an example of such a community.

We expect, that people in the same social group are

likely to help each other.

5 ARCHITECTURE OF THE

PROTOTYPE

In ﬁgure 3 we present the architecture of a proto-

type of the service. The mobile application (Emer-

gency Alert Widget, GPS-Push Service, Communica-

tion Capture Service in ﬁgure 3) is implemented on

a Motorola Milestone smart phone using Android as

OS. The Service Provider is realized in Java using

Apache Service Mix as OSGi

provider.

Figure 4 depicts the sequence diagram of the pro-

totype. Each subscriber is identiﬁed by his unique

telephone number (ID). In the Pre-Incidence Phase

each subscriber submits his geo-location data to the

GPS Cache (step 1). We use a GPS Push Service

running as background service on the Android smart

phone to transmit the GPS coordinates via REST to

the GPS Cache. The transmission is done on a reg-

ular basis, e.g. every 10 minutes and if the position

changed by more than 10 meters. The GPS Cache

stores the latest transmitted geo-position of each sub-

scriber in a database. The Communication Capture

Service runs on the mobile device. It monitors all in-

going and outgoing calls and messages. This com-

munication data is collected in a local cache. Once a

day the smart phone submits the cached communica-

tion data via SOAP to the Communication Collector

(step 2). The Social Cluster Calculator pulls new data

from the Communication Collector on a daily basis. It

builds a communication network, clusters the groups

and stores the results in the Social Group Provider’s

The Open Services Gateway initiative is a module

system and service platform for Java. For details see

http://www.osgi.org

database (step 3 and 4). OSGi is used as the internal

Service Provider’s protocol.

In the In-Incident Phase the user in need (in our

example this is Alice) starts a widget on her smart

phone (see ﬁgure 5). A emergencyalert is transmitted

via REST to the Emergency Alert Service (step A).

The service pulls the helpers candidate list from the

Social Group Provider for Alice via OSGi (step B).

For each candidate the GPS Cache is searched for the

latest geo-position data of the candidate (step C). If

close enough, a MMS is sent out to the helper with

information about the victim and the victims geo-

position (depicted on a map).

Note, that Figure 4 is restricted to a proper emer-

gency alert. False alerts can be revoked by a similar

process (not shown in Figure 4) which is password

protected. However, pragmatically a set of passwords

is provided which act as silent signals. A small so-

lution consists of three passwords, the ﬁrst signaling

a false alarm, the second signaling that the victim is

forced to revoke the alarm, and the third that there is

danger for the helpers.

6 IDENTIFYING SOCIAL

GROUPS BY CLUSTERING

SOCIAL NETWORKS

Calling persons willing to help is crucial for the pro-

posed system but their identiﬁcation is not trivial.

Communication networks or ’friend’ networks of on-

line social network sites usually contain many links

that do not result from close personal relations. Links

may connect business partners or co-workers. On so-

cial network sites people ’friend’ others they rarely

know. Therefore, identifying social groups is an ap-

proach to separate close personal contacts from other

distant contacts that are less willing to help in a case

of emergency.

The appropriate cluster algorithms depend on the

network that needs to be analyzed. All algorithms

need to be highly efﬁcient as the mentioned networks

are very huge (several million vertices). From com-

munication data weighted networks could be created

where an edge connects caller and the callee respec-

tively sender and receiver of a text message. The

edges can be weighted by the number of calls or mes-

sages. Walk context clustering is a suitable method

for this kind of network. It generates overlapping

clusters and can reﬂect that people might have several

groups of close contacts (family, friends, neighbors)

that are almost not connected with each other.

Cluster algorithms based on optimizing the mod-

ICE-B 2010 - International Conference on e-Business

212

Mobile Device

Emergency

Alert Widget

GPS Push

Service

Communication

Capture Service

GPS Cache

Communication

Collector

Social

Cluster

Calculator

Social Group

Provider

Emergency Alert Service

REST

Location data

<<every n minutes>>

SOAP

Communication data

<<daily>>

REST

Emergency Alert!

Emergency Alert,

Victim, Location

MMS

OSGi

Comm. data

<<daily>>

Social Groups

<<daily>>

OSGi

Location data

<<on request>>

Group data

<<on request>>

OSGi

Service Provider

Cache

Figure 3: Architecture of the Prototype.

Figure 4: Sequence Diagram of the Prototype.

ularity measure can not reﬂect the multitude of social

groups one person is integrated in. However, they are

able to deal with unweighted networks like the friend-

ship networks of social network sites.

GETTING HELP IN A CROWD - A Social Emergency Alert Service

213

Figure 5: Alert Widget on Phone.

6.1 Walk Context Clustering

Walk context clustering consists of two stages (see

algorithm 1). In the walk stage, a set of restricted

random walks is generated by starting a number of

walks at each vertex and repeatedly choosing the fol-

lowing vertices randomly among those vertices that

are linked by an edge which has a higher weight than

the previously taken one (see ﬁgure 6).

In the cluster construction stage, clusters get gen-

erated from the walks. Walk context clustering as-

signs a vertex to the cluster of another vertex if both

are part of the same walk. A level parameter l speci-

ﬁes the fraction of vertices at the beginning of a walk

that are disregarded. The later a pair of vertices ap-

pears in a walk the stronger is their connection. The

interesting feature of walk-context clustering is that

the closeness of two persons can be measured by the

maximal level that assigns one person to the cluster of

the other one.

A recently developed database-backed update al-

gorithm for the walk stage maintains asymptotically

optimal clusters in near real-time (< 0.2 sec for a

single update on graphs with approximately 500000

nodes and 20 million edges) (Franke and Geyer-

Schulz, 2009).

Algorithm 1: Walk context clustering.

Data: undirected, weighted graph G = (V, E), con-

stant p

H Walk generation

walkSet ←

0 forall v ∈ V do

for counter ← 1 to p do

walk ← (i); last ← 0; i ← v while N =

{x|ω

> last} 6=

0 do

j ← random element of N last ← ω

ap-

pend j to walk w i ← j

walkSet ← walklist ∪walk

Data: walkSet ws, vertex v, level l

H Cluster Construction

cluster ←

0 forall w ∈ ws|v ∈ w∧ pos(v,w) > l do

forall x ∈ w|pos(x,w) > l do

cluster ← cluster ∪ x

Figure 6: Narrowing search space for successive vertices of

the restricted random walk algorithm. The walks terminates

when no neighbor is within the search space. Solid arrows

symbolize used transitions. Dashed arrows symbolize links

to possible successors that have not been chosen by the ran-

dom process.

6.2 Modularity Clustering

Newman and Girvan (Newman and Girvan, 2004)

proposed a quality function for graph clusterings. The

modularity Q of a clustering C = C

,...,C

(∀i, j :

∩C

0 and ∪

= V) of a graph G = (V,E) is

Q =

∑

− a

) (1)

with e

{{v

}∈E|v

∈C

∧v

∈C

}

2|E|

and a

∑

The randomized greedy modularity clustering al-

gorithm (Ovelg¨onne and Geyer-Schulz, 2009) (see al-

gorithm 2) is a fast agglomerative hierarchical clus-

tering algorithm. It places each vertex in a separate

cluster and builds the complete dendrogram by re-

peatedly merging clusters. In each step it randomly

selects k clusters and searches among the clusters and

ICE-B 2010 - International Conference on e-Business

214

their neighbors for the pair that yields in the high-

est increase in modularity. The ∆Q of the merge of

two clusters i and j is ∆Q(i, j) = e

+ e

− 2a

2(e

− a

). This algorithm is able to process even

very large networks in reasonable time. A graph with

about 300.000 vertices and 1 million edges is clus-

tered in roughly 10 seconds.

Algorithm 2: Randomized greedy modularity clustering.

Data: undirected, connected graph G = (V,E), constant k

H Initialize

forall v ∈ V do

forall neighbours n of v do

e[v, n] ← 1/(2∗edgecount)

a[v] ← rowsum(e[v])

H Build Dendrogram (Randomized Greedy)

for i = 1 to rank(e)-1 do

maxDeltaQ ← −∞ for j = 1 to k do search among

k communities for best join

c1 ← random community for all communi-

ties c2 connected to c1 do

deltaQ ← 2(e[c1,c2] − (a[c1] ∗a[c2])) if

deltaQ > maxDeltaQ then

maxDeltaQ ← deltaQ next join ←

(c1,c2)

join(nextjoin)

7 ASSESSMENT OF BENEFITS

7.1 Ability to Transmit Request for

Help

Emergency situations can result of various incidents,

e.g. crime, accidents, medical emergencies. By

their nature, accidents happen unexpected and sud-

den. Some medical emergencies as heart attacks do as

well. The ability to make an emergency call in these

cases will depend on the physical condition of the per-

son in need.

For victims of violent crime their ability to send an

emergency signal depends on the progress of crime.

In 25%, respectively 22%, of the robberies analyzed

by Smith (Smith, 2003) instant violence or attempts

to snatch property don’t give the opportunity to call

for help. However, in 37% of the robberies the of-

fender approached the victim and demanded money

or valuables without immediate violence. In many

cases later victims were also aware of an upcoming

threat. In those cases it would be possible to send an

emergency message.

7.2 Chance for Nearby Help

The helpfulness of the proposed system depends on

the availability of close social contacts in the proxim-

ity of the site of the emergency. The actual number of

persons in ones proximity in general and the number

of close contacts with a particular motivation to help

will surely depend on several factors, e.g. place and

time.

To gain insight into the availability of potential

help the MIT reality mining experiment (Eagle and

Pentland, 2005) provides an interesting data set. For

this experiment a group of 100 persons (75 students

or faculty of the MIT Media Laboratory and 25 stu-

dents of the MIT Sloan Business School) has been

equipped with smart phones with special software ap-

plications preinstalled. These applications recorded

phone numbers of incoming and outgoing calls, text

messages, and the id of the cellular tower the phones

were connected to during one academic year.

The phone call and text message data from the re-

ality mining data set can be used to construct a com-

munication network. Based on the assumption that

the communication intensity of two people reﬂects the

closeness of their relation, social groups can be iden-

tiﬁed by clustering this weighted network. The avail-

ability of nearby help from within the social group of

a person in need can be estimated by the number of

persons from the same social group whose phones are

connected to the same cellular tower as the person in

need.

For three consecutive months all communica-

tion prior to the speciﬁc month has been used to

build an undirected, weighted communication net-

work. The edge weights are the number of commu-

nication events (calls, text messages). This network

has been clustered by the walk context cluster algo-

rithm with the level parameter set to 0.8. The gener-

ated clusters had an average size of about 5.

The MIT reality mining data set contains a history

of time-stamped connection records of the participat-

ing persons. For each month 1000 connection records

have been randomly selected. Then, for each persons

in a connection record the number of persons of his

social cluster who have been connected to the same

cellular tower at the same time have been counted.

This simulation showed that on average more than

one close fellow student was available for help at any

time (see table 1). E.g. for September 2004 the prob-

ability of having at least one person from one’s social

group in the proximity was 78% - independent of time

of the day.

This is just a basic assessment for a particular

group of people that has several shortcomings. Due to

the lack of more detailed data it was not possible to as-

GETTING HELP IN A CROWD - A Social Emergency Alert Service

215

Table 1: Average number of other persons of same social group connected to same cellular tower for 1000 randomly selected

persons and points of time.

Day (6am-6pm) Evening (6pm-11pm) Night (11pm-6am)

Group Others Group Others Group Others

09/2004 1.16 3.74 1.17 2.73 0.95 1.51

10/2004 1.33 3.12 1.29 3.4 1.14 1.82

11/2004 2.2 3.18 1.53 2.16 1.76 1.03

sess if other social contacts than fellow students were

available for help. Proximity could just be estimated

by radio cells which have a radius of a few hundred

meters in urban areas and a radius up to about 35 km

in rural areas. But it is fair to regard the results as an

indication that people living in urban areas will usu-

ally have at least one of their social contacts in their

proximity.

7.3 Chance to Actually Receive Help

Once a request for help has been transmitted to a po-

tential helper in the proximity of the site of the emer-

gency the chance to actually receive help depends on

the willingness and the ability of the informed per-

sons to get to that place. The ability to get to a

speciﬁc site can be supported in various ways. For

example a map and route directions could be dis-

played together with the emergency message. The re-

search of Markey (Markey, 2000) showed that also in

computer-mediated communication settings help re-

quests that were directed to speciﬁc persons had a

high probability to receive fast response and the by-

stander effect was virtually eliminated.

8 CONCLUSIONS

In this paper a novel emergency alert service has been

introduced which addresses all obstacles to providing

help identiﬁed in the social emergency intervention

process discussed in section 2. The service is de-

signed to reduce psychological barriers that result in a

bystander effect and inhibit effective help for persons

in need. The analysis of emergency situations and

whereabouts of persons in relation to their respective

social group suggest that the described service can ac-

tually provide a beneﬁt in practice. A ﬁrst prototype

of this service is currently under development.

The main emphasis of this paper is on the reduc-

tion of the bystander effect. However, the following

challenges which are beyond the scope of this paper

are examples of what must be addressed thoroughly

for concrete industrial service offerings:

1. Legal issues: The service raises e.g. the problem

that potential helpers become liable to help and

failure to do so may be prosecuted.

2. Privacy: The service should be designed in order

to minimize the intrusion of privacy of service-

subscribers.

3. Emergency dialog: The emergency dialog could

be further automated e.g. by providing an auto-

matic classiﬁcation of the incident type.

4. Geo-positioning problems: Geo-positioning is

still problematic in large buildings, tunnels, sub-

terranean areas (e.g. subway). Enhancements

could be based e.g. by embedding geo-position

senders in such structures or by image recogni-

tion techniques which exploit public geo-coded

images of such spaces.

ACKNOWLEDGEMENTS

The research leading to these results has received

funding from the European Community’s 7th Frame-

work Program FP7/2007-2013 under grant agreement

◦

215453 – WeKnowIt.

REFERENCES

Android (2010). Android. http://www.android.com/.

Batson, C. D., Cochran, P. J., Biederman, M. F., Blosser,

J. L., Ryan, M. J., and Vogt, B. (1978). Failure to help

when in a hurry: Callousness or conﬂict? Personality

and Social Psychology Bulletin, 4:97 – 101.

Baumeister, R. F., Chesner, S. P., Sanders, P. S., and Tice,

D. M. (1995). Who’s in charge here? group leaders

do lend help in emergencies. Personality and Social

Psychology Bulletin, 14:17 – 22.

Borkowski, J. and Lempiainen, J. (2006). Practical

network-based techniques for mobile positioning in

UMTS. EURASIP J. Appl. Signal Process., 2006:149–

149.

Chen, R., Sharman, R., Rao, H. R., and Upadhyaya, S. J.

(2008). Coordination in emergency response manage-

ment. Communications of the ACM, 51(5):66 – 73.

Clark, R. and Word, L. E. (1972). Why don’t bystanders

help? because of ambiguity? Journal of Personality

and Social Psychology, 24:392 – 400.

ICE-B 2010 - International Conference on e-Business

216

Comfort, L. K., Kilkon, K., and Zagorecki, A. (2009). Co-

ordination in rapidly evolving disaster response sys-

tems: The role of information. American Behavioral

Scientist, 48(3):295 – 313.

Cramer, R. E., McMaster, M. R., Bartell, P., and Dragna, M.

(1988). Subject competence and the minimization of

the bystander effect. Journal of Applied Social Psy-

chology, 18:1133 – 1148.

Darley, J. M. and Batson, C. (1973). ”From Jerusalem to

Jericho”: A study of situational and dispositional vari-

ables in helping behavior. Journal of Personality and

Social Psychology, 27(1):100 – 108.

Darley, J. M. and Latan´e, B. (1968). Bystander intervention

in emergencies: Diffusion of responsibility. Journal

of Personality and Social Psychology, 8(4):377 – 383.

Eagle, N. and Pentland, A. (2005). Reality mining: Sens-

ing complex social systems. Personal and Ubiquitous

Computing, 9:1 – 14.

EU (2006). Directive 2006/24/EC of the European Parlia-

ment and of the Council on the retention of data gener-

ated or processed in connection with the provision of

publicly available electronic communications services

or of public communications networks and amending

Directive 2002/58/EC.

Faraj, S. and Xiao, Y. (2006). Coordination in fast-response

organizations. Management Science, 52(8):1155 –

1169.

Franke, M. and Geyer-Schulz, A. (2009). An update al-

gorithm for restricted random walk clustering for dy-

namic data sets. Advances in Data Analysis and Clas-

siﬁcation, 3(1):63 – 92.

Garcia, S. M., Weaver, K., Moskowitz, G. B., and Darley,

J. M. (2002). Crowded minds: The implicit bystander

effect. Journal of Personality and Social Psychology,

83:843 – 853.

Granovetter, M. (1982). The strength of weak ties: a net-

work theory revisited. chapter 5, pages 105 – 130.

Sage Publications, Inc., Beverly Hills, California.

Granovetter, M. S. (1973). The strength of weak ties. The

American Journal of Sociology, 78(6):1360 – 1380.

Hoser, B. and Geyer-Schulz, A. (2005). Eigenspectral anal-

ysis of hermitian adjacency matrices for the analysis

of group substructures. The Journal of Mathematical

Sociology, 29(4):265 – 294.

Kikiras, P. and Drakoulis, D. (2004). An integrated

approach for the estimation of mobile subscriber

geolocation. Wireless Personal Communications,

30(2):217–231.

Latan´e, B. and Darley, J. (1970). The Unresponsive By-

stander: Why doesn’t he help? Appleton-Century-

Crofts, New York.

Levine, M., Prosser, A., Evans, D., and Reicher, S. (2005).

Identity and emergency intervention: How social

group membership and inclusiveness of group bound-

aries shape helping behavior. Personality and Social

Psychology Bulletin, 31(4):443 – 453.

Markey, P. M. (2000). Bystander intervention in computer-

mediated communication. Computers in Human Be-

havior, 16:183 – 188.

Milgram, S. (1970). The experience of living in cities. Sci-

ence, 167:1461 – 1468.

Miller, D. T. and McFarland, C. (1987). Pluralistic igno-

rance: When similarity is interpreted as dissimiliarity.

Journal of Personality and Social Psychology, 53:298

– 305.

Monin, B. and Norton, M. (2003). Perceptions of a ﬂuid

consensus: Uniqueness bias, false consensus, false

polarization, and pluralistic ignorance in a water con-

servation crisis. Personality and Social Psychology

Bulletin, 29:559 – 567.

Moriarty, T. (1975). Crime, commitment, and the respon-

sive bystander: Two ﬁeld experiments. Journal of Per-

sonality and Social Psychology, 31:370 – 376.

Nadler, A. and Halabi, S. (2006). Intergroup helping as sta-

tus relations: Effects of status stability, identiﬁcation,

and type of help on receptivity to high-status group’s

help. Journal of Personality and Social Psychology,

91(1):97 – 110.

Newman, M. E. J. and Girvan, M. (2004). Finding and eval-

uating community structure in networks. Physical Re-

view E, 69(2):026113.

Open Handset Alliance (2010). Open handset alliance.

http://www.openhandsetalliance.com/.

Ovelg¨onne, M. and Geyer-Schulz, A. (2009). Randomized

greedy modularity clustering. Technical report, Karl-

sruhe.

Palen, L. and Liu, S. B. (2007). Citizen communication

in crisis: Anticipating a future of ict-supported public

participation. In CHI 2007 Proceedings, pages 727 –

736. ACM.

Piliavin, I. M., Piliavin, J. A., and Rodin, J. (1975). Costs,

diffusion, and the stigmatized victim. Journal of Per-

sonality and Social Psychology, 32(3):429 – 438.

Rungg, A. (2010). Privatsp¨are 2.0. Technical report, Finan-

cial Times Deutschland, 17.2.2010, p.25.

Rutkowski, G. K., Gruder, C. L., and Romer, D. (1983).

Group cohesiveness, social norms, and bystander in-

tervention. Journal of Personality and Social Psychol-

ogy, 44:545 – 552.

Schwartz, S. H. and Gottlieb, A. (1980). Bystander

anonymity and reaction to emergencies. Journal of

Personality and Social Psychology, 39:418 – 430.

Shotland, R. L. and Heinold, W. D. (1985). Bystander

response to arterial bleeding: Helping skills, the

decision-making process, and differentiating the help-

ing response. Journal of Personality and Social Psy-

chology, 49:347 – 356.

Shotland, R. L. and Stebbins, C. A. (1980). Bystander re-

sponse to rape: Can a victim attract help? Journal of

Applied Social Psychology, 10:510 – 527.

Shotland, R. L. and Straw, M. K. (1976). Bystander re-

sponse to an assault: When a man attacks a woman.

Journal of Personality and Social Psychology, 34:990

– 999.

Smith, J. (2003). The nature of personal robbery. Home

Ofﬁce Research Study 254, Home Ofﬁce Research,

Development and Statictics Directorate.

GETTING HELP IN A CROWD - A Social Emergency Alert Service

217

Spiegel Online (2009a). Notruf-Mitschnitt belegt Bru-

talit¨at der S-Bahn-Pr¨ugler. Technical report, Spiegel

Online.

Spiegel Online (2009b). S-Bahn-Fahrgast stirbt nach bru-

taler Attacke. Technical report, Spiegel Online.

Spiegel Online (2009c). T¨odlicher Angriff: Mehrere Men-

schen beobachteten M¨unchner S-Bahn-Attacke. Tech-

nical report, Spiegel Online.

St¨urmer, S., Snyder, M., and Omoto, A. M. (2005). Proso-

cial emotions and helping: The moderating role of

group membership. Journal of Personality and Social

Psychology, 88(3):532 – 546.

Yuan, Y. and Detlor, B. (2005). Intelligent mobile cri-

sis response systems. Communications of the ACM,

48(2):95 – 98.

ICE-B 2010 - International Conference on e-Business

218