EXPLORING THE POTENTIAL FOR USING ARTIFICIAL

INTELLIGENCE TECHNIQUES IN POLICE REPORT ANALYSIS

A Design Research Approach

Fredrik Bengtsson

1,2

, Amadeus Hein

1,3

and Carl Magnus Olsson

Software Engineering and Management, University of Gothenburg, Gothenburg, Sweden

Acando, Gothenburg, Sweden

Diadrom, Gothenburg, Sweden

Department of Computer Science and Engineering, Chalmers, University of Gothenburg, Gothenburg, Sweden

Keywords: Artificial Intelligence, Data Mining, Neural Networks, Criminology, Design Research.

Abstract: Storing digital data is increasingly affordable and attractive for many organizations, thus allowing

longitudinal postum analysis of events and for identifying trends that may hold interest for predicting future

scenarios. Results of manual data analysis suffer from high time consumption and human error due to the

complexity or volume of data. Responding to this, our study explores advances in artificial intelligence

techniques by presenting experiences from the iterative development of a prototype that assists intelligence

officers in identifying trends in serial crimes. This study contributes by illustrating the first steps that may

be taken towards diffusing advances in artificial intelligence into a practice area serving the general public.

1 INTRODUCTION

Storing large sets of data is no longer a problem for

organizations due to the increasing affordability of

digital storage solutions, as well as technical

advancements in computing power, bandwidth for

data retrieval. The problem instead lies in how we

may leverage advantages from the data we collect

(Shaon & Woolf 2008). Analyzing the data to gain

benefits requires more time to complete than before,

as the data volume grows exponentially over time

(Chen et al. 2004; Liang & Austin 2005;

Rajagopalan & Isken 2001). Furthermore, another

challenge in performing longitudinal analysis lies in

interpreting the raw data (Nath 2006). Data mining

is one field linked with artificial intelligence (AI)

that strives to address these challenges (Williams

1983; Liang & Austin 2005) that holds the potential

for saving time for the analyst (Charles 1998; Chen

et al. 2004). However, mining complex data is

difficult and often requires a skilled data miner and

an analyst with good domain knowledge (Nath

2006) to ensure a low rate of human errors (Chen et

al. 2004; Charles 1998).

This highlights an opportunity for software

systems to reduce the data volume to only include

data sets that are most likely relevant to the analysis.

Completing the analysis manually therefore becomes

less time consuming when the raw data has been

pre-processed. Additionally, software may be used

to replicate repetitive existing human analysis steps

to provide human analysts with higher quality data

to work with, e.g. showing data trends, narrowing

the frame of analysis, and making decisions easier

through suggesting likely beneficial answers. The

potential for practical application of this in software

systems to complement and assist analysis has been

established (cf. Chen et al. 2004; Liang & Austin

2005; Charles 1998). It remains unclear, however, as

to how an effective interplay between AI systems

and human actors may be found for these

organizations, as well as how such balance may be

struck as adapted system designs based on AI

techniques are introduced.

In response to this, the research question for this

paper is: how can AI techniques be used and adapted

to assist in identifying data trends that are likely to

be of relevance for further investigation by human

agents? To answer this, we use the increasingly

recognized design research approach (cf. Hevner et

al. 2004; Hevner & Chatterjee 2010). The main

contribution of this paper subsequently lies in

213

Bengtsson F., Hein A. and Magnus Olsson C..

EXPLORING THE POTENTIAL FOR USING ARTIFICIAL INTELLIGENCE TECHNIQUES IN POLICE REPORT ANALYSIS - A Design Research

Approach.

DOI: 10.5220/0003710802130218

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 213-218

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

illustrating how advances in AI may be approached

to diffuse and adapt these into practice.

2 THEORETICAL

BACKGROUND

Two attributes for data analysis are of particular

interest: volume and complexity. We refer to volume

as the amount of data being analyzed, and

complexity as the challenge in interpreting data.

Using these, we develop a matrix model which

illustrates how they relate and roles that different

actors are likely to play in negotiating them.

2.1 Data Volume

Over the past decades, the world has fully entered

into the information age, resulting in exponential

increases in computing power and communication

bandwidth (Shaon & Woolf 2008). In turn, this has

resulted in a rise of the amount of data that can be

accessed (Rajagopalan & Isken 2001), and therefore

also increasing the time it takes to perform data

analysis. Organizations involved in data analysis

have to face the challenge of accurately and

efficiently analyzing the growing data volume (Chen

et al. 2004; Liang & Austin 2005). One reason for

the challenges is that the relevant data is often

hidden in a larger set of irrelevant data, making it

difficult to find for human actors. Data mining is one

field that strives to address these challenges by

efficiently extracting information from large data

sets (Liang & Austin 2005).

2.2 Data Complexity

Complex data may be differently expressed and

structured in different data sets, changed

periodically, be generally diffuse, and/or span long

periods of time (Tanasescu et al. 2005; Chen et al.

2004). Furthermore, volume also brings complexity

as the permutations of possible interpretations

increases (Nath 2006). Humans are inherently good

at recognizing complex patterns in data sets due to

our brain being perceptive of patterns (Kingston

1992). However, the more complex data, the more

time the analysis will take while the risk of human

error increases (Chen et al. 2004; Charles 1998).

2.3 Illustrating the Relationship

between Complexity and Volume

The relation between data volume and complexity

can be shown in what we refer to as a Complexity-

Volume matrix (CVmatrix). A CVmatrix consists of

the four permutations of volume and complexity that

a data set may exist in. The suitability for an agent

(human or AI) to analyze a data set depends on the

configuration of the attributes in the specific

permutation. Human actors are good at analyzing

complex data in smaller volumes, while an AI is

more efficient with large volumes of less complex

data (Charles 1998). In the case when the data set is

both complex and contains a large volume of data

the risk of errors are highest (Chen et al. 2004).

Thus, finding a method for reducing either (or both)

of the attributes is desirable to make the analysis of

large data sets more efficient and correct.

The CVmatrix presents two plausible paths from

the full set of raw data towards a more manageable

subset: (1) having human actors focus only on a

small subset of the complex data for in-depth

analysis, or (2) using computer based algorithms for

precise but more basic and repeated analysis of the

full data set. For the purposes of this paper, and as

shown in our CVmatrix (Figure 1), our interest lies

in exploring the first option where the role that the

software-based system assist human actors in the

selection process of data subsets, and defer

interpretation of this data set to human actors based

on AI-informed suggested links.

Figure 1: A CVmatrix showing the approach taken for

reducing volume and complexity in this paper.

3 METHOD

3.1 Research Setting

This study is done in collaboration with the police

intelligence unit in Gothenburg, Sweden. As the

Swedish police are under government directive to

work against ‘crimes of quantity’ such as burglary,

physical abuse, and vandalism, the challenge in

effectively analyzing the growing number of police

reports is of particular interest to address. In

addition, the Gothenburg police have lately noted

ICAART 2012 - International Conference on Agents and Artificial Intelligence

214

increasing problems with ‘roaming burglaries’.

Roaming burglaries are crimes where the criminals

travel quickly around the country while committing

burglaries which makes it particularly difficult for

the police districts to apprehend the criminals and

notice patterns in these crime tours. Statistics show

that only 4% of burglaries were resolved in 2009

(BRÅ 2009), meaning the risk for being

apprehended is currently very low and is a growing

problem from a societal perspective.

This paper reports from the first stage of the

collaboration and focuses on exploring the impact of

a software prototype (Sherlock) using advances in

AI to assist in negotiating the large volume of crimes

and identify patterns in police reports that are of

particular interest for analysts to assess.

3.2 Research Approach

This research combines qualitative interpretation

(Creswell 2003) with an iterative design research

approach (Hevner et al. 2004). Design research is

characterized by the focus on improving current

practices through design of artifacts based on best

practices identified in research (Hevner et al. 2004).

Through the design artifact, practice needs and

research findings are provided with a vessel for

contributing to both practitioners and researchers.

The approach taken holds benefits given that our

experience as researchers with criminology is low

compared to our collaborating partner, and the

iterative nature of design research (Kuechler &

Vaishnavi 2007) affords opportunities for learning-

by-doing (Jeffries et al. 1981) and reflection-in-

action (Schön 1983). These practices are inherently

qualitative acts of testing hypotheses that are guided

by interpretation based on the active collaborative

participation and prototype driven approach.

3.3 Data Collection

Data collection during the iterative development of

Sherlock has been a combination of informal

discussions on a daily basis and open-ended

interviews (Creswell 2003; Wolcott 1994). The

open-ended interviews were further semi-structured,

holding some general questions prepared in advance

to stimulate interviewees to freely share reflections,

experiences, local knowledge and practices (Myers

& Newman 2007). This allows the interviews to be

guided by both research interest and practitioner

experiences, in the same way as intended by design

research. The interview results are presented in this

paper as an integrated part of the three design

iterations, and are discussed in section 4.

4 DISCUSSION OF RESEARCH

PROCESS AND OUTCOMES

Our design process is based on three iterations (I1

through I3) and the following subsections discuss

the research outcomes of each such iteration,

followed by a summary of the outcomes as a whole.

4.1 I1: Data Preparation

Through interviews with our intelligence analysts,

we identified that the modus operandi (crime scene

behavior) has a major impact on the potential for

recognizing crimes that are part of a larger series.

We therefore opted for using neural networks in our

prototype, as extant research outlined this AI

technique as promising for crime analysis (cf.

Charles 1998; Chau et al. 2002; Chen et al. 2004).

Specifically, we decided to rely on a neural network

architecture called Self-Organizing Map (SOM). It is

based on unsupervised training and is commonly

used for classifying patterns (Heaton 2008).

Before relying on this approach, the data from

the police reports had to be pre-processed and

filtered (as suggested by Goodwill & Alison 2006).

As also noted by Helberg (2002), the preparation

process plays an extensive role in the analysis

procedure, and the initial development of our study

was thus focused on preparing the data for neural

network analysis.

To achieve this, we designed and implemented a

way of normalizing all crime data to be better suited

as input for the neural network. The normalizing

process’ primarily objective is to translate the

relevant data from police reports to be used as input

for the neural network. We started by assigning each

attribute a predefined numerical value with a unique

meaning. For example, the gender attribute was

awarded three numerical values: ‘0’, ‘1’ and ‘2’,

which represents ‘unknown’, ‘man’ and ‘woman’.

This was then repeated for all attributes, forming

a complex matrix of numerical values with unique

meaning associated with each crime. Normalizing

the data thus provided a chance to discuss and define

ways in which crime types could be characterized

based on the existing practice of our analysts, and

was the foundation for input to be used by the neural

network.

The SOM network was evaluated together with

the developed input structure. The input structure

consisted of two normalized crimes that the network

compared. However, when testing the network with

two normalized crimes we experienced a redundant

and unmanageable number of different patterns

EXPLORING THE POTENTIAL FOR USING ARTIFICIAL INTELLIGENCE TECHNIQUES IN POLICE REPORT

ANALYSIS - A Design Research Approach

215

suggested. Furthermore, the network had too little

information to base its analysis on, meaning

extracting more data from the police reports and

revising the input structure was necessary.

4.2 I2: Main Development Phase

To revise the input structure, additional interviews

were held to understand which data needed to be

retrieved from the police reports and why. Based on

this, we turned to the suggestions by Chau et al.

(2002), Nath (2006) and Bache et al. (2007) who

emphasize that the free text open-entry field of

police reports contain much valuable.

The open-entry information may be difficult to

interpret, however, as the data is expressed and

structured depending on who writes the report. Such

differences are part of what defines data sets as

complex (Tanasescu et al. 2005). Interpretation of

the open-entry field was thus decided on as a focal

point for the software prototype as advances in

pattern recognition in such fields are likely to

provide particular value to the analysis. The second

iteration thus emphasized a broadened use of data

mining through exploring an efficient way for

retrieving relevant data from a complex data set.

To address the challenge of interpreting open-

entry fields and reducing the volume of data to

review for the human analysts, we implemented an

open-entry interpreter based on a lexical-lookup

(Chau et al. 2002) extraction approach which we

adapted to establish a modus operandi. The lexical

lookup details were defined in collaboration with the

intelligence analysts, and we implemented the open-

entry interpreter to softly match keywords using an

algorithm called Q-gram (Younghoon et al. 2010) to

recognize different tense of words and misspelled

words as part of the match.

As a result of the evaluation performed in the

first iteration, in which the SOM network was

unable to produce the expected output, we realized

that there was a need of modifying the input

structure in order for the neural network to find

crime patterns. The proposal was to create an

algorithm that merged all attributes of two

normalized crimes to generate an input pattern for

the neural network that would indicate the

similarities and differences between the crimes. The

neural network then interprets high peaks of the

pattern as differences and low peaks as similarities

to decide whether crimes are related or not.

The advantage of using neural networks for

recognizing and classifying complex patterns is that

the network can be taught over time. This means that

the neural network’s ability to analyze effectively

increases over time, similar to how human actor

experience works and may positively influence

results. Training a neural network involves gathering

both training and evaluation data (Heaton 2008).

Sherlock’s neural network is trained and evaluated

through data retrieved by the intelligence analysts

for previous crime series where the manual work has

already been done to be able to interpret the success

of the Sherlock implementation in relation to the

known manual analysis results. Throughout the

development of Sherlock it was important that the

intelligence analysts contributed with their

perspective of how to analyze and classify crimes.

We evaluated the implementation of the SOM

neural network with predefined input based on the

new structure and discovered that the SOM network

did not function as well as needed. This was similar

to our initial evaluation, and the reason for why the

structure of the neural network input was

emphasized and why the merge algorithm and the

open-entry interpreter were developed. At this stage

it was evident that the network still had difficulty

mapping crimes itself, due to the input structure

required, resulting in output that was very hard to

perceive any patterns of value in. After evaluating

the open-entry interpreter by running it on 60

manually pre-analyzed police reports, the results

showed that all modus operandi of these reports

could be correctly structured by the open-entry

interpreter.

4.3 I3: A New Neural Network Model

We were now at a point where we required more

control over the learning step of the neural network

and thus started the search for a new neural network

model. This led us to the feedforward

backpropagation network (FB). Feedforward is a

method for recalling patterns and backpropagation is

a supervised training method (Heaton 2008). By

using a supervised training method, rather than

automated as described in iteration two, we could

now manipulate the neural network more actively

during the training session to assist in rejecting or

accepting suggested patterns.

To evaluate the implementation of an FB based

Sherlock, the intelligence analysts provided us with

a crime series they identified in 2009 containing a

total of 58 linked crimes in the same crime series.

We evaluated the FB network analysis capabilities

by using a training set consisting of a subset of 10

crimes to train the FB network as our hypothesis was

that it should then be able to recognize most of the

remaining 48 crimes in the series when executed on

the full set of crimes. To provide realism to this

ICAART 2012 - International Conference on Agents and Artificial Intelligence

216

scenario, we included the entire 58 known and

linked crimes in the police report database of

burglaries in the Swedish region of Skaraborg 2009.

This brought the number of crime reports to a total

of 318 crimes, with the 58 linked crimes included.

Executing Sherlock on this data set based only on

the 10 linked crimes it had been trained on, the

software identified 41 out of the 58 crimes that the

human analysts had already established as linked.

Sherlock’s failure to identify 17 crimes indicates

that further training and potential tuning of the FB is

still needed, but the results are nevertheless a

substantial improvement over the SOM based

approach. Given the manipulation based training of

the FB based neural network, the human agents have

an important role to play for increasing the

reliability of Sherlock. As the intelligence analysts

train the network with more crime series, the

network should learn to better determine whether the

crimes belong to a crime series, thus reducing the

error rate in relation to manual analysis by human

actors.

4.4 Reflection on Iteration Outcomes

From the evaluation stage of the FB network in the

third iteration, this study demonstrates that it is

possible to reduce the volume of a complex data set

through combining a set of AI techniques in a

practice situation. The AI techniques that Sherlock

uses are open-entry interpretation, a crime merge

algorithm, and aided by an FB neural network, our

results indicate that it is possible to follow the path

indicated in the CVmatrix earlier (Figure 1).

As a result of the growing body of research in

neural networks, future work may both improve our

current approach and offer new alternatives

altogether. We strongly believe that assessing such

advances for extracting more of the relevant data in

practice situations is likely to play an important role

in moving them from potential options to best

practice examples in different and specific contexts.

Our study marks one such example, but as with

neural networks themselves, many more data points

are needed to capture a more complete set of linked

needs between different practices. We are, however,

in this study taking an active stance that practice

driven needs are of particular relevance for such

continued work.

The open-entry interpreter which we developed

and used in the Sherlock implementation is itself a

technique that could be used separately from the

neural network to reduce the volume of complex

data sets. As the purpose of this paper has been

towards exploring and demonstrating the potential of

applying advances in AI in practice situations, rather

than technical discussions of the design artifact, we

hold further elaboration on the open-entry interpreter

as part of future papers. When generalized for use

towards all crime types, the interpreter would allow

the human analysts to run more specific search

queries, for example matching the modus operandi

of a crime to get more precise data to analyze.

Discussing the long-term impact of the

prototype, the police analysts reflect that

exhaustively analyzing ‘crimes of quantity’ is a task

not performed at present due to the human resource

demands this holds. Subsequently, the potential of

Sherlock is perceived as very strong: “We know that

Sherlock is still in its early stage of development,

but it is still able to provide great help to us since it

is performing a task that we cannot do at present. By

introducing a software system that takes care of the

collecting and cleaning process, it is possible for us

to focus on what we were trained for – analysis.”

They further recognized how the collaborative

marriage between research and actual practices

brings visible benefits to them when adopting the

system into their daily work routines: “Another

important factor with Sherlock is that it is based on

the same coding schemes as we already use,

meaning it does not bring additional overhead work

to integrate into our current work processes.”

5 CONCLUSIONS

This study set out to explore the assisting role AI

techniques may have in identifying data trends that

are likely to be of relevance for additional

investigation by human agents. We found

practitioners who struggle with large volumes of

complex data and collaborated actively with them to

adapt the developed prototype to their needs, based

on best practices and advances in extant research.

The focus on serial crime was due to such crime

marking the majority of incidents and is of great

societal impact due to the repeated nature of them

and the difficulty in terms of resources and time

faced by manual analysis of such crime.

The contributions of this paper are linked to the

practice implications and potential of the Sherlock

implementation experiences. Not only is our

research approach one that argues for the mutual

contributions to practice and research, but we also

strongly feel that advancements in the field of

artificial intelligence stands much to gain from

studies that also expose practice needs and

challenges for diffusing theoretical advances.

EXPLORING THE POTENTIAL FOR USING ARTIFICIAL INTELLIGENCE TECHNIQUES IN POLICE REPORT

ANALYSIS - A Design Research Approach

217

REFERENCES

Bache, R, Crestani, F, Canter D, Youngs, D 2007

‘Application of Language Models to Suspect

Prioritisation and Suspect Likelihood in Serial

Crimes’, IAS ‘07 Proceedings of the Third

International Symposium on Information Assurance

and Security, IEEE Computer Society Washington,

DC, USA, pp. 399-404.

BRÅ 2009. Brottsförebyggande Rådet (Translation: ‘The

Swedish National Council for Crime Prevention’),

Available in English at http://www.bra.se/extra/pod/

?action=pod_show&id=1&module_instance=11

Charles, J 1998, ‘AI and law enforcement’, Intelligent

Systems and their Applications, IEEE, vol. 13, no. 1,

pp. 77-80.

Chau, M, Xu, J, Chen, H 2002, ‘Extracting meaningful

entities from police narrative reports’, Proceedings of

the 2002 annual national conference on digital

government research, Digital Government Society of

North America, pp. 1-5

Chen, H, Chung, W, Xu, J.J, Wang, G, Qin, Y, Chau, M

2004, ‘Crime data mining: a general framework and

some examples’, Computer, vol. 37, no. 4, pp. 50-56.

Creswell, J W 2003, Research design: qualitative,

quantitative and mixed methods approaches, 2nd edn,

Sage Publications Inc, London.

Goodwill, A, Alison, L 2006, ‘The development of a filter

model for prioritising suspects in burglary offences’,

Psychology, Crime and Law, vol. 12, no. 4, pp. 395-

416.

Heaton, J 2008, Introduction to neural networks for Java,

2nd edn. Heaton Research, Chesterfield.

Helberg, C 2002, Data mining with confidence – SPSS

(2nd edn). SPSS Inc, Chicago, Illinois.

Hevner, A. R, Chatterjee, S 2010, Design research in

information systems: Theory and practice, Springer,

New York; London.

Hevner, A. R, March, S, Park, J, Ram, S 2004, ‘Design

science in information systems research’, MIS

Quarterly, vol. 28, no. 1, pp. 75-105.

Jeffries, R, Turner, A. A, Polson, P. G, Atwood, M. E

1981, ‘The processes involved in designing software’,

in Anderson, J. R. (ed.), Cognitive Skills And Their

Acquisition, Hillsdale, NJ: Lawrence Erlbaum

Associates, pp, 255-283.

Kingston, C 1992, ‘Neural networks in forensic science’,

Journal of Forensic Sciences, vol. 37, no. 1.

Kuechler, W. and Vaishnavi, V. 2007, ‘Design [science]

research in IS: A work in progress’ Proceedings of 2nd

International Conference on Design Science Research

in Information Systems and Technology (DESRIST

‘07), Pasedena, CA, May 13-16, 2007.

Liang, B, Austin, J 2005, ‘A neural network for mining

large volumes of time series data’, Conference on

industrial technology, pp. 688-693, 14-17 December

2005.

Myers, M, Newman, M 2007 “The qualitative interview in

IS research: examining the craft”, Information and

Organization, vol. 17, no. 1, pp. 2-26.

Nath, S V, 2006 ‘Crime pattern detection using data

mining’, Proceeding of the 2006 IEEE/WIC/ACM

International Conference on Web Intelligent Agent

Technology, Oracle Corporation, Florida, Atlantic

University, pp. 41-44.

Rajagopalan, B, Isken, M.W. 2001, “Exploiting data

preparation to enhance mining and knowledge

discovery,” IEEE Transactions on Systems, Man, and

Cybernetics, Part C: Applications and Reviews,

vol.31, no.4, pp.460-467.

Schön, D. A 1983, The reflective practitioner: How

professionals think in action, Basic Books, New York.

Shaon, A, Woolf, A 2008, ‘An OASIS based approach to

effective long-term digital metadata curation’,

Computer and Information Science, vol. 1, no. 2, pp.

2-12.

Tanasescu, A, Boussaid, O, Bentayeb, F 2005, ‘Preparing

complex data for warehousing’ The 3rd ACS/IEEE

International Conference on Computer Systems and

Applications, pp. 30-33.

Williams, C 1983 ‘A brief introduction to artificial

intelligence’, OCEANS’83 Proceedings, vol., no., pp.

94-99.

Wolcott, H.T. 1994, Transforming qualitative data:

description, analysis, and interpretation, Sage,

Thousand Oaks, CA.

Younghoon, K, Kyoung-Gu, W, Hyoungmin, P, Kyuseok,

S 2010, ‘Efficient processing of substring match

queries with inverted q-gram indexes’ International

Conference on Data Engineering (ICDE), vol., no., pp.

721-732, 01-06 March 2010.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

218