THE ROLE OF DATA MINING TECHNIQUES IN EMERGENCY

MANAGEMENT

Ning Chen

GECAD, Instituto Superior de Engenharia do Porto, Instituto Politecnico do Porto, Portugal

An Chen

Institute of Policy and Management, Chinese Academy of Sciences, Beijing, P. R. China

Keywords:

Emergency management, Data mining, Intelligent decision support system, Assessment, Planning.

Abstract:

Emergency management is becoming more and more attractive in both theory and practice due to the fre-

quently occurring incidents in the world. The objective of emergency management is to make optimal deci-

sions to decrease or diminish harm caused by incidents. Nowadays the overwhelming amount of information

leads to a great need of effective data analysis for the purpose of well informed decision. The potential of

data mining has been demonstrated through the success of decision-making module in present-day emergency

management systems. In this paper, we review advanced data mining techniques applied in emergency man-

agement and indicate some promising future research directions.

1 INTRODUCTION

Emergency management (EM) is described as a pro-

cess integrating various social resources to control

and manage emergency based on the analysis of cau-

sation, evolution and consequences (L. Ji, 2006). The

objective of EM is to make optimal decisions to de-

crease or diminish harm caused by incidents. Gen-

erally, incidents can be categorized into natural disas-

ter, accident, public sanitary incident, and social secu-

rity incident. There are some concepts closely related

to emergency management in literature, such as risk

management, disruption management, alert manage-

ment, crisis management, early warning etc. From

the generalized perspective they can be regarded as

parts of EM though the information and strategies are

slightly different. Risk management is a very general

concept. There are many kinds of risks in our life, in-

cluding economics risk, social risk, natural risk, etc.

In our opinion, risk has different forms regarding the

intensity. The lightest is disruption situation which

can be recovered to the original status easily. The

worst is emergent situation which must have caused

loss of the value, and only a part of the whole value

can be recovered. The intensity of crisis lies between

disruption and emergency. Crisis management is a

turning point in the sense that a successful interven-

tion can transfer the bad situation to a better one,

while a failing intervention often induces a worse sta-

tus. Early warning and alert management mainly fo-

cus on early provision before incidents. Early warn-

ing is to ﬁnd the symptom of an incident or disaster.

It is always based on the prediction of the incident,

and able to give an alert for the coming event. Alert

Management is to manage the signal and early warn-

ing information.

Generally, emergency management can be de-

scribed as a four-step process, namely, assessment,

planning, action and monitoring (M. de la Asuncion,

2005). As outlined in Figure 1, assessment is to eval-

uate the occurrence or evolution of incidents with re-

spect to the scenario of incidents and situation. Plan-

ning is to construct the response (plan), i.e., a pro-

posal of control activities, based on both current situ-

ation and foreseeable evolvement, and then dispatch it

to the responsible person or organization (emergency

manager). The successive step is to implement the ac-

tions in the plan and evaluate the impact to decrease

the damages caused by incident or overcomethe prob-

lematic situation. Finally, the monitoring module de-

tects the execution of plan in dynamic environment,

concerning incomplete information and uncertainty

factors during the episode. If speciﬁed conditions sat-

isfy, the original plan is revised and a new process is

118

Chen N. and Chen A. (2009).

THE ROLE OF DATA MINING TECHNIQUES IN EMERGENCY MANAGEMENT.

In Proceedings of the 11th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

118-123

DOI: 10.5220/0001961601180123

 SciTePress

triggered. The emergency manager is responsible for

the execution of the whole process by means of vali-

dating the plan and modifying it at necessary.

evaluation

scenario impact

plan

Emergency Manager

Figure 1: Emergency management process.

Traditional emergency management mostly relies

on the skill and experience of emergency manager

or domain expert to evaluate the situation, deﬁne

the goal of intervention and choose the strategies to

achieve the goal. Nowadays, due to the rapid de-

velopment of information techniques, the increasing

amount of data makes the decision-making task more

difﬁcult and complicated. The data involved includes

numerical data captured by sensors or communication

devices reﬂecting current situation, qualitative data

from domain experts expressing personal preference,

and historical data from data base. The data needs to

be converted into valuable knowledge which helps to

generate well informed plans or strategies. Due to the

complexity of emergency management problems and

explosive growth of data processed, there is a great

need to enhance and supplement the capabilities of

emergency management under the support of special-

ized problem-solving tools. Data mining (DM) is a

good candidate for this purpose, in particular in as-

sessment and planning phases.

Since effective management of incidents is based

on reasonable assessment and decisions, decision

making is the kernel of emergency management. The

success of current emergency management systems

is largely dependent on the decision-making module

which enables managers to take advantage of the ca-

pabilities of DM techniques. In literature, a variety

of data mining techniques were employed in the area

of emergency management to provide the functional-

ity for the purpose of improving the intelligence of

emergency management systems. In such systems,

data mining techniques analyze the information, track

the evolution of episode of emergency and share the

decisions with emergency manager. The potential of

data mining techniques have been demonstrated for

valuable plan construction in both military domains,

e.g., air camping design, noncombatant evacuation,

and civil domains, e.g., oil spill, ﬂood, forest ﬁre. As

a related work, (U. Cortes, 2000) shows the usage of

artiﬁcial intelligence techniques in environmental de-

cision support system. However to our knowledge,

there is little attempt on reviewing DM applications

in the ﬁeld of emergence management.

The remainder of the paper is organized as fol-

lows. Section 2 addresses state-of-the-art DM tech-

niques, including association rule mining, classiﬁca-

tion and sorting, clustering, and case-based reason-

ing as well as their contributions in emergency man-

agement. Some emergency management systems in-

tegrated with DM techniques are introduced in sec-

tion 3. Section 4 concludes the paper and presents

some interesting future research directions.

2 DATA MINING TECHNIQUES

IN EMERGENCY

MANAGEMENT

Formally, data mining is an analytical technique to

discover hidden, implicate, and nontrivial knowledge

from database. Data mining, especially predictive

data mining acts as a model-driven process for risk

prediction in pre-incident phase and evolvement fore-

cast in post-incident phase. A broad range of DM

techniques have been employed in the management

of emergencies and play important roles to provide

emergency manager with high-level information in

order to take proper decisions at right time. The con-

tributions of DM can be summarized in four aspects:

(1) In assessment phase, it is used for knowledge ac-

quisition to identify potential incident and predict the

possible consequence from relevant factors and their

relationship; (2) In response planning phase, it is used

to suggest the desirable plan to responsible manager;

(3) In plan action phase, it is used to evaluate the per-

formance of plan concerning the accuracy, reliability,

and usefulness. (4) In monitoring phase, it is used to

acquire high quality information from raw data. In

this section, we introduce advanced data mining tech-

niques and their representative applications in emer-

gency management.

2.1 Association Rule Mining

Association rule (AR) mining is one of the most

widely applied technique to discover interesting re-

lations between variables in large databases. Mining

association rules is introduced for the ﬁrst time to dis-

cover regularities between products from basket data

recorded by point-of-sale (POS) systems in supermar-

kets (R. Agrawal, 1993). In a market basket database,

each transaction contains the set of products (items)

purchased by a customer. Association rule mining at-

tempts to discover important association among items

THE ROLE OF DATA MINING TECHNIQUES IN EMERGENCY MANAGEMENT

119

such that the presence of some items in a transaction

will imply that of other items in the same transaction.

So far, association rule has been extended to different

types and applied in a vast variety of areas.

Some experiences show that effective perfor-

mance of DM is the key to the success of risk man-

agement, which relies on the accurate prediction of

emergency to some extent. For this purpose, associ-

ation rule mining is employed to predict the possibil-

ity of occurrence of particular incidents. A geospa-

tial decision support system (S.K. Harms, 2001) is

developed to improve the quality and accessibility

of draught risk management. The relationship be-

tween user deﬁned target episodes and other climate

events is exploited from geospatial and temporal data

and consequently used to predict the occurrence of

target episodes. It was reported that the association

rule between weather station data, crop yields and sea

surface thermal properties provides reasonable strate-

gies like adjusting the planting date and population to

farmers in US.

In network event management, abnormal event is

seen as a kind of emergencyessential to network man-

agement. Three types of events, namely, burst pat-

terns, periodic patterns and mutually dependent pat-

terns, that indicate underlying problems are mined

from historical event data (J. L. Hellerstein, 2002).

These patterns do not occur frequently enough in

well-maintained production environments, but they

are of particular interest to event management. Min-

ing burst patterns is achieved by ﬁrst ﬁnding periods

in which event rates are higher than a speciﬁed thresh-

old, then mining patterns common to the periods iden-

tiﬁed. Periodic patterns are explored by ﬁnding pe-

riod lengths for each event type followed by temporal

associations. Mining mutually dependent patterns is

similar to association rule mining except that it fo-

cuses on infrequent patterns related to problem situa-

tions.

2.2 Classiﬁcation and Sorting

Classiﬁcation is a typical supervised learning method

to derive models between features (independent vari-

ables) and class (target variable) in the sense that

the class information is considered during learning.

The process of classiﬁcation is to ﬁrst develop mod-

els through training on the preclassiﬁed data and then

assign new data to a class on the basis of the model

developed. Sorting also involves the assignment of

an alternative to a predeﬁned group expect that the

groups are ordered in an ordinal way. In emergency

management, classiﬁcation and sorting are particu-

larly useful in the assessment of incidents, which ana-

lyze whether the incident occurs or to what extent the

damage is caused.

Decision tree is one of the fastest and easily in-

terpreted algorithms to solve the task of classifying

instances. The tree structure is constructed by a “di-

vide and conquer” strategy, with leaves representing

classiﬁcation and branches representing conjunctions

of feature tests. Decision tree algorithms have many

variations differing in alternate methods of choosing

the split attribute and value, pruning method, and data

access methods used for large training dataset. In

(A. Gerbesioti, 2001), the classiﬁcation model de-

rived by C4.5 decision tree algorithm (Quinlan, 1993)

relating to some problem-speciﬁc instances is used by

an expert system for ranking infected buildings.

Artiﬁcial neural network (ANN) is a computa-

tional modelisation inspired from biological neural

network to explore nonlinear patterns from compli-

cated and large data. It consists of a set of intercon-

nected artiﬁcial neurons for data computation through

a connectionist approach. ANN has been widely

used in various categories of risk assessment prob-

lems. E.g, an early warning system is composed of

knowledge acquisition, knowledge representation and

knowledge-based reasoning using ANN (B. Yang,

2001). The applicability of ANN is demonstrated on

predicting storm surge and surge deviation in real-

world data from Suao Harbor station of Taiwan (Lee,

2008).

2.3 Clustering

Clustering is an unsupervised process to partition a

set of data into homogeneous clusters in the sense that

there is no preclassiﬁed data as it were in classiﬁca-

tion. The objects are grouped together on the basis of

intrinsic similarity to maximize the intraclass similar-

ity and minimize the interclass similarity. Clustering

is able to generate compressed representatives of raw

data, especially for large, complex data set with many

variables and a lot of internal structures. Applications

of clustering in emergency management mainly focus

on high quality information acquisition from a mass

of raw data with redundancy and noise.

Since real data usually contains reduplicate and

false alarms, the alarms should be ﬁltered and aggre-

gated to achieve a better understanding of situation.

Through a density-based clustering algorithm which

ﬁnds dense regions separated by low-density regions

and clusters together the objects in the same dense

region, the alarms form into clusters based on the as-

sociated attributes and then false alarms are identiﬁed

from the uniformly behaving neighborhoods. As a re-

sult, the aggregated alerts are generated from a large

ICEIS 2009 - International Conference on Enterprise Information Systems

120

number of incoming alarms (Vandana, 2005).

Kohonen’s self-organizing map (SOM) (Kohonen,

1997) is called topology mapping in the sense that

the similarity of input data is preserved in the out-

put layer through a competitive learning. As a visual-

ized approach, SOM makes possible to detect internal

data structure from the map grid. In (Y. Atoji, 2000),

SOM is used as an information ﬁltering method to se-

lect useful information from a large amount of data

related to emergency. Utilizing the clustering abil-

ity of SOM, the incoming data that contains similar

content is mapped to a particular category and then a

number of representative keywords are extracted for

each pattern. Also, SOM is used for security assess-

ment of power system (K.S. Swarup, 2006). The line

ﬂows under different component cases are input to

a two-dimensional grid network to obtain the cluster

of components based on their loading limits. From

the output patterns, the violation of constraints can be

identiﬁed and consequently the secure and insecure

status are classiﬁed accordingly. The utilization of

SOM eliminates the repetitive computation required

in traditional methods.

2.4 Case-based Reasoning

Case-based reasoning (CBR) originates from tradi-

tional behavior of human problem solving to solve

new problems based on past cases. For example, a

doctor diagnoses the disease of a patient by recall-

ing another patient who has similar symptoms. Sim-

ilarly, a lawyer advocates a case in a trial based on a

judge in the past. Formally, the process of CBR con-

sists of four steps (A. Aamodt, 1994): (1) Retrieve:

extract relevant cases from case base based on a dis-

tance measurement; (2) Reuse: map the similar cases

to the target case by adapting to current situation; (3)

Revise: apply the solution in the real world and test

the inﬂuence with respect to the feedback; (4) Retain:

store the new case and solution if accepted. Com-

pared with rule-based reasoning, CBR does not need

the complicated phase of rule extraction, however, it

is critically depended on the distance metric for sim-

ilar case retrieval and sensitive to the computational

cost when a large number of cases are available for

comparison. Moreover, the adaption is important to

solution reuse taking into account the difference be-

tween matched cases and current situation, which is

probably more difﬁcult than generating a new solu-

tion.

In emergency management, CBR methodology is

regarded as a powerful method to create the emer-

gency plan based on the solutions of similar past inci-

dents. The popularity of CBR beneﬁts from the pres-

ence of cases available in the databases. For example,

CBR is applied to plan the initial attack to forest ﬁres

with respect to past interventions stored in the case

base (P. Avesani, 1999). The adaption of retrieved

plan to domain requirements and constraints is imple-

mented by a constraint propagationalgorithm. CBR is

also used to generate incident response plans automat-

ically in the domain of computer and network security

(G. Capuzzi, 2006). The past cases including attacks

(incidents) and corresponding responses (plans) are

stored as a sequence of concrete events. When a new

attack occurs, it searches the case memory for similar

attacks on the basis of a distance metric, then reuses

the past response by substituting the concrete action

and parameters with current action type and parame-

ters.

3 EMERGENCY MANAGEMENT

SYSTEMS AND APPLICATIONS

Emergency management system (EMS) is an intelli-

gent decision support system (IDSS) capable of en-

hancing the ability of human to manage the emergent

situations. It plays an increasing role to reduce the

risks by appropriate and proactive preparation or min-

imize the negative impact caused by incidents by im-

mediate and efﬁcient response. Although a number

of emergency management systems have been devel-

oped, the capabilities of emergency management sys-

tems should be addressed:

• Acquire, ﬁlter, select, transform and visualize data

in diverse types and from different resources to

provide well understanding of relevant factors and

identify the potential problems;

• Explore, represent, structure and manage back-

ground knowledge, expert subjective knowledge

and predictable models by means of statistical

analysis, data mining or other approaches;

• Provide functions of problem formulization, diag-

nosis, planning and optimization using rule-based

reasoning or case-based reasoning;

• Assist emergency manager to evaluate the impact

and cost of strategies, choose appreciate plans for

a given problem;

• Monitor the performanceof plans and adapt to dy-

namic environment taken into consideration un-

certainty factors;

• Provide user-friend interface which allows users

to offer subjective criteria, justify decision impact,

and intervene the process interactively.

THE ROLE OF DATA MINING TECHNIQUES IN EMERGENCY MANAGEMENT

121

Data Acquisition

Assessment

Monitoring

Planning

Action

Filter Aggression Transformation

AR Sorting Classification

...

Rule-based

Reasoning

Cased-based

Reasoning

...

Database

Rule base

Case base

Real-time data

Expert

Figure 2: Emergency management ﬂow.

As shown in Figure 2, the ﬂow of emergencyman-

agement starts from data acquisition followed by data

preprocessing of ﬁlter, aggression and transformation.

Then using data mining based analysis tools, rules

are extracted and reserved in rule base for the usage

of assessment and planning. Finally, the decisions

are made through rule-based reasoning, case-based

reasoning or hybrid approaches. So far a variety of

EMS at regional, national, or international level have

been designed in centralized or distributed architec-

ture for various domains. Some representative emer-

gency management systems are listed in the follow-

ing.

• ARTEMIS (J.Z. Hernandez, 2001): a general

architecture of knowledge-based environmental

emergency management system. The evolution of

incident is represented by a collection of causal

models, classiﬁcation models and temporal mod-

els structured in a hierarchy. The effect of primary

causes is propagated top-down through a rule in-

ference process and the expected impact of plans

is estimated by simulators.

• PHOENIX (P.R. Cohen, 1989): a multi-agent sys-

tem capable of providing a distributed, adaptive

and real-time planner to control simulated forest

ﬁres.

• CHARADE (P. Avesani, 1999): an interactive de-

cision support system for situation assessment and

activities planning in forest ﬁre emergencies ap-

plication domain. This system is also used for

knowledge acquisition from the cases, which en-

able users to directly extract various forms of

knowledge, e.g., dependencies between features,

clusters of cases, feature statistical description.

• SIADEX (M. de la Asuncion, 2005): an interac-

tive ﬁre ﬁghting planner, integrating several artiﬁ-

cial intelligence techniques for designing and re-

designing plans against forest ﬁres. The system

provides an easy-to-use tool capable of not only

assisting the manager during the whole process of

decision making in an online fashion, but also re-

producing the episodes for manager training in an

ofﬂine fashion.

• RODOS (W. Raskob, 2005): a real-time decision

support system for nuclear emergency manage-

ment, providing countermeasure and remediation

strategies from pre-release phase, release phase to

long-time phase.

• AGrIP (J. Luo, 2007): a ﬂood decision support

system integrating a multi-strategy data mining

tool MSMiner in the algorithm library.

• SIPE (Bienkowski, 1995): a crisis management

system for plan generation to oil threats in the sea

requiring users to provide knowledge or interac-

tion to the system.

• CubeView (C.T. Lu, 2005): a road transportation

system applying advanced data mining and visu-

alization techniques to road trafﬁc data for emer-

gency situation control and management plan-

ning. Visualization of loop-detector trafﬁc data

helps to recognize potentially important patterns

embedded in the data. Outlier detection discov-

ers trafﬁc patterns applicable in the detection of,

prevention of and recovery from road crisis.

• IMASH (E. Iakovou, 2001): an intelligent inte-

grated dynamic information management tool to

provide comprehensive data pertaining to emer-

gency planning and response for hurricane disas-

ters. It can offer support for a wide range of hurri-

cane disaster related activities including hurricane

preparedness, hurricane response and restoration.

4 CONCLUDING REMARKS AND

FUTURE PERSPECTIVES

With the increasing demand of data analysis involved

in emergencymanagement, there is an urgent need for

automatic and effective tools which allow emergency

managers to make quick and desirable intervention to

incidents and disasters. The intelligent decision sup-

port systems based on data mining techniques become

more and more popular in diverse domains related to

emergency management. This paper reviews the up to

date data mining techniques and their applications in

the ﬁeld of emergency management.

A future research direction is the investigation of

fuzzy data mining for emergency management. The

data related to the incident and situation are usually

represented in linguistic assessments instead of de-

terministic numerical values, thus, the fuzzy strate-

gies are needed to deal with the imprecise and un-

ICEIS 2009 - International Conference on Enterprise Information Systems

122

certain information. Since decision-making is usually

carried out in a distributed environment to achieve a

satisfactory solution, organizational learning and col-

laborative learning is another direction of need in fu-

ture study. Additionally, special attentions should be

given to the implementation of emergency manage-

ment system. As pointed out by several authors, an

interactive system is more realistic than a completely

automated system, in which human is responsible of

personnel knowledge speciﬁcation, knowledge inter-

pretation and decision selection. The selection can

be performed through a multi-criteria evaluation from

several perspectives such as cost, effort, feasibility,

public acceptance, psychological and political impli-

cation, preference of decision makers (W. Raskob,

2005). Since EM managers are not experts on DM,

the comprehension is quite important for easy access,

e.g, derived rules are easily understandable and appli-

cable in decision reasoning. The scalability of DM

is worth noting for decision-making due to the infor-

mation ﬂood occurring at the inception of emergency,

when real-time response becomes difﬁcult.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the ﬁnancial

grant of GECAD/ISEP-Knowledge Based, Cognitive

and Learning Systems (C2007-FCT/442/2006).

REFERENCES

A. Aamodt, E. P. (1994). Case-based reasoning: Founda-

tional issues, methodological variations and system

approaches. Artiﬁcial Intelligence Communications,

1:39–52.

A. Gerbesioti, V. Delis, Y. T. S. A. (2001). Developing de-

cision support tools for confronting seismic hazards.

In Proc. of the 8th Panhellenic Conference on Infor-

matics, pages 247–253.

B. Yang, L.X. Lib, H. J. J. X. (2001). An early warning

system for loan risk assessment using artiﬁcial neural

networks. Knowledge-Based Systems, 14(5-6):303–

306.

Bienkowski, M. (1995). Demonstrating the operational fea-

sibility of new technologies: the arpi ifds. IEEE Ex-

pert, 10(1):27–33.

C. T. Lu, L.N. Sripada, S. S. R. L. (2005). Transportation

data visualization and mining for emergency manage-

ment. International Journal of Critical Infrastruc-

tures, 1(2/3):170–194.

E. Iakovou, C. D. (2001). An information management sys-

tem for the emergency management of hurricane dis-

asters. International Journal of Risk Assessment and

Management, 2(3-4):243–262.

G. Capuzzi, E. Cardinale, I. P. L. S. (2006). An incident re-

sponse support system. International Journal of Com-

puter Science and Network Security, 6(10):72–78.

J. L. Hellerstein, S. Ma, C.-S. P. (2002). Discovering ac-

tionable patterns in event data. IBM Systems Journal,

41(3):475–493.

J. Luo, L. Xu, J.-P. J. L. Z. Z. S. (2007). Flood decision

support system on agent grid: Method and implemen-

tation. Enterprise Information Systems, 1(1):49–68.

J. Z. Hernandez, J. S. (2001). Knowledge-based models for

emergency management systems. Expert Systems with

Applications, 20:173–186.

Kohonen, T. (1997). Self-organizing Maps. Springer Ver-

lag, Berlin.

K. S. Swarup, P. C. (2006). Power system static security

assessment using self-organizing neural network. J.

Indian Inst. Sci., 86:327–342.

L. Ji, H. Chi, e. (2006). Emergency Management. High

Education Press.

Lee, T.-L. (2008). Prediction of storm surge and surge de-

viation using a neural network. Journal of Coastal

Research, 24(sp3):76–82.

M. de la Asuncion, L. Castillo, F. J. G. A. P. F. (2005).

Siadex: an interactive knowledge-based planner for

decision support in forest ﬁre ﬁghting. AI Commu-

nications, 18(4):257–268.

P. Avesani, A. P. (1999). Cases on ﬁre: Applying cbr to

emergency management. New Review of Applied Ex-

pert Systems, 6:175–190.

P. R. Cohen, M.L. Greenberg, D. H. A. H. (1989). Trial

by ﬁre: Understanding the design requirements for

agents in complex environments. AI Magazine,

10(3):32–48.

Quinlan, J. (1993). C4.5: Programs for Machine Learning.

Morgan Kauffman, Los Altos, CA.

R. Agrawal, T. Imielinski, A. S. (1993). Mining association

rules between sets of items in large databases. In Proc.

of ACM SIGMOD, pages 207–216.

S. K. Harms, S. Goddard, S. R. W. W. T. T. (2001). Data

mining in a geospatial decision support system for

drought risk management. In Proc. of the National

Conference on Digital Government Research, pages

9–16.

U. Cortes, M.S. Marre, L. C. I. R. (2000). Artiﬁcial intel-

ligence and environmental decision support systems.

Applied Intelligence, 13:77–91.

Vandana, P. (2005). Dm-ams: Employing data mining tech-

niques for alert management. In Proceedings of Na-

tional Conference on Digital Government Research,

volume 89 of ACM International Conference Pro-

ceeding Series, pages 103–111.

W. Raskob, V. Bertsch, J. G. S. B. F. G. (2005). Demands

to and experience with the decision support system

rodos for off-site emergency management in the de-

cision making process in germany. In Proc. of the 2nd

International ISCRAM Conference.

Y. Atoji, T. Koiso, S. N. (2000). An information ﬁlter-

ing method for emergency management based on self-

organizing map. Human Interface, 2(3):1–6.

THE ROLE OF DATA MINING TECHNIQUES IN EMERGENCY MANAGEMENT

123