University of Sindh Journal of Information and Communication Technology

(USJICT)
Volume 2, Issue 1, January 2018

ISSN-E: 2523-1235, ISSN-P: 2521-5582 © Published by University of Sindh, Jamshoro

Website: http://sujo.usindh.edu.pk/index.php/USJICT/

A Survey of Data Mining Techniques for Crime Detection

Shamaila Qayyum, Hafsa Shareef Dar
Department of Software Engineering, International Islamic University, Islamabad
shamaillla@gmail.com, hafsa_darr@yahoo.com

_______________________________________________________________________________________________________
Abstract: In large datasets, data mining is one of the most powerful ways of knowledge extraction or we can say it is one of the
best approaches to detect underlying relationships among data with the help of machine learning and artificial intelligence
techniques. Crime Detection is one of the hot topics in data mining where different patterns of criminology are identified. It includes
variety of steps, starting from identification of crime characterization till detection of crime pattern. For this purpose, various crime
detection techniques have been discussed in literature. In this paper, we have selected widely adapted data mining techniques that
are specifically used for crime detection. The analytical study is presented with an extraction in form of strengths and weakness of
each technique. Each technique is specific to its use. This survey would serve as a helping guide to researchers to get state of the
art crime detection techniques in data mining along with pros and cons.

Keywords: Data mining, Crime detection, Classification, Clustering, Association, Prediction, Constraint, Association rules
_____________________________________________________________________________________________________________________

I. INTRODUCTION through computers is more convenient and less costly than

With the advancement of technology, criminals are also hiring and training people for the purpose of collecting and
adapting smarter ways to commit crimes. With time, the analyzing existing crime information. Data Mining can
crime rate has been tremendously increased instead of help the investigators, no matter if they are experienced or
decreasing. The technology where, has helped the people not, to explore large databases efficiently[10]. This not
in making their lives easier, has also helped the criminals only helps them in tracking and solving crimes but also
in making new plans for their crimes. It is becoming trivial predicting crimes in advance[11]. Data mining techniques
to get smarter and technical ways to investigate and prevent offers some predictive models that manipulate the hidden
these crimes. One of the most common ways of crime that information and can predict the trends [12].
can be seen nowadays, is not just within the streets, but in
Hosseinkhani et al.,[13] has suggested some data mining
the world of connectivity; the internet. It is deemed
techniques that can be used for crime detection. These
important to discover and adapt the ways that help in
techniques are clustering, association rule mining,
efficient discovery and prevention of such crimes. The
deviation detection, classification and string comparator.
criminal acts range from the street crimes to massive terror
The crime detection data mining techniques as presented
attacks and undoubtedly to offend large databases and
by Hsinchun et al., are entity extraction, clustering,
systems. This all lie under the umbrella of crime[1]. Crimes
association rule mining, sequential pattern mining,
can be defined for example as network attacks [2], Hacking
deviation detection, classification, string comparator and
[3], network intrusion [2], cyber fraud [2, 4],deceptive
social network analysis [14]. Hossein Hassani et al., have
information [5].
later presented a review of some existing crime data mining
Data mining refers to the extraction, discovery and analysis techniques which included; entity extraction, cluster
of meaningful patterns and rules from a very large amount analysis, association rule, classification and social network
of data. It is emerging as very useful tools for crime analysis [15]. The aim of this study is to get an insight of
detection [6]. Data mining is a very powerful tool to the techniques that are followed for crime data mining.
undermine the activities of criminals by analyzing the Previously, survey on data mining techniques was
criminal’s record and information and preventing the conducted with only five techniques that covered small
crimes in future[7]. Data mining for crime detection is scope of data mining. In this survey, we intend to cover
considered one of the most important research area as each possible technique used for crime detection.
described in [8] despite data mining being a new and
In this study the existing techniques of data mining for
evolving field itself [9]. Data mining is thought of being
crime detection and investigation are thoroughly observed.
very helpful and accurate in understanding the crime trends
Different data mining techniques are discussed and
as compared to Humans. Humans are often error prone
analyzed for their usage. They are then further compared
specially when overworked; computers prove to be more
for their strengths and weaknesses under different
accurate as compared to human. Mining information
 University of Sindh Journal of Information and Communication Technology (USJICT), Vol.2(1), pg.: 1-6

circumstances of usage. The differences and This is very helpful in investigating the simultaneous
commonalities of these techniques have also been occurrences of events[27]. The strength of association can
discussed. be measured in terms of support; the applicability of rule
to given dataset and confidence; the frequency of
II. EXISTING CRIME DATA MINING TECHNIQUES: appearance of one data in transactions that contain another
Crime can range from the simple street crimes to data[28].
internationally planed crimes[8]. Crime data mining, as
compared to usual data mining, is more concerned with D. Sequential pattern mining
privacy [16]. In structured data, the patterns are identified Sequential pattern mining discovers frequently
through different traditional data mining techniques such occurred sequence of items at different intervals of times.
as association, classification, prediction, clustering and It is helpful in network intrusion detection. For meaningful
outlier analysis[17]. Advanced data mining handles both results, a large amount of structured data is required[2].
structured and unstructured data for pattern recognition Ayres et al., have provided an algorithm that finds all
[14, 18]. In this section we analyze the existing data possible sequences in the transactions data, very quickly.
mining techniques that are used for crime detection and They have used depth first traversal combined with the
investigation. bitmap representation to achieve this [29].

A. Entity extraction E. Deviation detection

It is used to identify persons, vehicles, texts basically It is also known as outlier detection. It studies data that
by identifying patterns [19]. It is the process of extracting has clear distinction from rest of the data set. This
data from text documents [15]. In computer forensics it can technique is very helpful in fraud detection and other crime
help in identifying programs write by hackers. This is done analysis[2, 4]. Aggarwal, in his research has proposed an
by grouping similar programs. However it requires a huge evolutionary outlier detection algorithm which works by
amount of clean data to produce good results[3]. The main selection then crossover and then mutation methods [30].
approaches for entity extraction are machine learning, Whereas, Arnind et al., have proposed a linear method of
statistic based, rule based and lexical lookup [19]. deviation detection in larger database, which provides
Machine learning techniques use algorithms to extract solution by simulating a mechanism that is familiar to
knowledge and derive patterns e.g, decision trees, neural human beings [31].
networks[20], entropy maximization [21] and hidden
markov models [22]. Rule based systems are the structural, F. Classification
contextual or lexical hand crafted rules to identify Classification has some predefined classes. This data
entities[23]. Statistical based systems used training dataset mining technique finds out some common characteristics
to obtain statistics for identifying the occurrence of of different crime entities. These are then organized into
particular patterns[24]. Lexical loops systems maintain the predefined classes. A predefined classification scheme
popular entities of interest and look up in text the phrases is required for this approach[4]. Classification is very
that are specified in their lexicons[21]. helpful in predicting crimes and identifying crime entities
with a very less amount of time. However, it needs a
B. Clustering techniques predefined scheme for classification and complete training
They help in maximizing or minimizing the interclass with testing data, as only this could bring accuracy to the
similarities by grouping the data items into the classes predicted results [4]. Classification aims to discover a set
based on their characteristics. In criminal investigation, it of rules from the dataset. Classification can be carried out
can help in identifying criminals that follow a set pattern either by Decision Trees[32-34], Support Vector Machines
for committing a crime [25]. Clustering functions by [35], [35]Naïve Bayes Rules[36] or Neural Networks [37-
obtaining the distance measurement among objects, such 42].
as Euclidean distance, Minkowksi distance and Manhattan
distance. Different algorithms of cluster group the data into G. String comparator
hierarchical manner or partition them as per It computes the similarity among the database records
requirements[26]. Apart from these two approaches, other by comparing their textual fields. It helps identifying
main approaches for clustering are: density based, grid deceptive information. However it requires a large amount
based, model based and constraint based. of computations [5]. It returns a numerical value by
comparing two strings [23].
C. Association rule mining
It presents the patterns as the rules by finding frequent H. Social network analysis
occurrences of items within a dataset. This is helpful in Social network analysis analyses the role and
identifying network attacks [2]. It was initially built to interaction of nodes within a conceptual framework. It can
observe interesting co-occurrences in market data[23]. be used to identify criminal’s roles by creating a network.

2
 University of Sindh Journal of Information and Communication Technology (USJICT), Vol.2(1), pg.: 1-6

It can also help in analyzing the flow of information among The mentioned techniques have been deeply studied to
these entities, though it won’t help in identifying networks’ know the pros and cons of any technique while adapting it
true leaders[5].it reveals the structure within some text, by in crime detection. Entity Extraction is enhanced by
presenting some interlinked entities[43]. This shows that machine learning techniques but polluted data can be a
people have participated or communicated somewhere[44]. hurdle to it so its weakness is requirement of clean data.
The most widely used techniques for SNA are: Degree; The strength of clustering is detection of outliers without
number of nodes connected to any node [45], Density; labeled data but as this process is costly and its
number of edges in a specific area as compared to the effectiveness depends on selected method as well.
overall number of edges[46] and Centrality; the importance Association Rule Mining is yet another technique which
of a node within a structure[47]. basically supports classification and its weakness is its
specific nature to classification rules only. Sequential
Hossein Hassani et al., have reviewed the data mining pattern mining has wide range of applicability in all areas
techniques for crime. This review cover the techniques: and hence, large amount of structured data is required for
entity extraction, cluster analysis, association rule, its execution. Deviation detection is widely used in fraud
classification and social network analysis [15]. In [48] a detection but data dependency in some areas is still a
tool was discussed which is based on Natural Language question unsolved. Classification technique is
Processing technique for detection of white collar crimes. conventional technique with very less time consumption
However, a comparative analysis of all above mentioned and weakness is predefined scheme of classification that
crime data mining techniques is still missing in the requires complete training data set. String comparator
literature. accuracy is great when we consider numerical values but it
requires high computations. Social network analysis
III. ANALYSIS focuses on relationships between actors rather than their
In this section, comparative analysis of each technique is attribute which makes it more direct but unfortunately it
presented on the basis of its strength and weakness. The doesn’t identify network’s true leader in the system.
strength and weakness of each technique has been
extracted from literature review in which authors and IV. CLASSIFICATION OF EXISTING TECHNIQUES
researchers have identified positive and negative impacts In order to understand the techniques used for crime data
of particular technique. Table 1 is a complete analysis of mining, we first present classification of these techniques.
each technique: This classification contains the data mining techniques that
are specifically used for crime data mining and are stated
TABLE 1 STRENGTH AND WEAKNESS OF EACH CRIME
DATA MINING TECHNIQUE in the literature. The classification of the techniques along
with the methods they use is shown in Fig 1.
TECHNIQUE STRENGTH WEAKNESS
Entity extraction Machine learning large amount of clean
makes it easier data required
Clustering Detect outliers Computational cost is
without any required high. Its effectiveness
label data also depends upon the
method used
Association Rule Support It is used for the most
mining accurate classification
rules
Sequential pattern Wide range of large amount of
mining applicability structured data is
required
Deviation Widely applicable in Sometimes its data
detection fraud detection dependency becomes a
hurdle
Classification Very less time Predefined scheme of
consumption classification and
complete training
dataset required
String Accuracy in terms of Large amount of
Comparator numerical value computation required
Social network focus on relationships Won’t identify
analysis between actors rather network’s true leaders
than attributes of
actors

3
 University of Sindh Journal of Information and Communication Technology (USJICT), Vol.2(1), pg.: 1-6

Identifying Criminals Social Network [42, 43, 44, 45, 46,

role analysis 47]

This classification not only helps researchers but also

investigators, to choose the right technique for their
scenario, without wasting time. To understand these
techniques in more details and to have a comparative
analysis of these details, we have identified some attributes
of these techniques, after a thorough study of the
techniques. These attributes are defined as follow:

i. Timeliness is one of the most important attribute.

Most investigators are concerned with the timely
response of a data mining technique.
ii. Data dependency is another important attributes.
At times the techniques require a lot of data to
produce accurate results, but the required amount
of data may not be available.
iii. Accuracy is very important in identifying the true
criminals. So this attribute is also very important
for analyzing crime data mining techniques.

A. Comparison of Existing Crime Data Mining

Figure 1. Classification of crime data mining techniques
Techniques
This section aims to compare and analyze the existing
techniques for crime data mining, on the basis of proposed
In order to clearly understand the usage of these
thematic taxonomy. We have learnt that there are various
techniques, it must be clear to the researchers and
techniques used for detection of crime, and every technique
investigators that what kind of usage each technique
possess. In order to take the full advantage of these is used in different scenario. So the real comparison is
techniques and to make their selection easier for the according to the scenario in which they are used, and they
may not need to be compared for the common attributes.
researchers and investigators, we have thoroughly studied
But however, it is important to understand the comparison
the techniques that are helpful in crime data mining; we
of these techniques for effective selection and flawless
have then prepared a comparison of these techniques based
investigation. The comparison is based on three values –
on their usage. We have identified all the areas in which
crime data mining techniques can be used and have further timeliness, data dependency and accuracy of the technique.
identified which data mining technique is used in each Table 3 shows the comparison of these techniques based
on the proposed taxonomy.
scenario. We have thus, further classified these techniques
based on their usage. The usage of each crime data mining TABLE 3 COMPARISON OF CRIME DATA MINING
technique is presented in Table 2. TECHNIQUES

TABLE 2 USAGE BASED CLASSIFICATION OF CRIME COMPARISON OF CRIME DATA MINING TECHNIQUES
DATA MINING TECHNIQUES ATTRIBUTE/TECHNIQ TIMELINES DATA ACCURAC
UE S DEPENDENC Y
USAGE DATA MINING REFERENCE Y
TECHNIQUE Entity extraction Less time Huge Accurate
Identification of Entity extraction [11, 15, 16, 17, 18, consumptio amount of
programs written by 19, 20, 21] n data required
hackers Clustering Moderate Less data Accurate
Identifying criminals Clustering [22, 23] Time dependency
following a set pattern consumptio
Identifying network Association rule [20, 24, 25, 26] n
attacks mining Association Rule Moderate Moderate Accurate
Intrusion detection Sequential pattern [24, 27] mining
mining Sequential pattern Moderate Huge Accurate
Fraud detection Deviation detection [24, 28, 29, 30] mining amount of
Predicting crimes Classification [28, 31-33, 36-41] data required
Identifying deceptive String Comparator [20, 42] Deviation detection Less time Moderate Accurate
information consumptio
n

4
 University of Sindh Journal of Information and Communication Technology (USJICT), Vol.2(1), pg.: 1-6

Classification Less time Complete Accurate VI. FUTURE WORK

consumptio training set on data
n data required availabilit Crime data is an important area in which efficient crime
y detection data mining techniques play vital role for analyst
String Comparator More time Moderate Accurate and law enforcers to proceed the case in investigations and
consumptio
n help resolving criminal cases. The scope of this study can
Social network Moderate Moderate No true be further enhanced by working on criminal investigation
analysis leader data set like FBI and crime detection of counter terrorism
identified measures. Another enhancement of this technique is to
[2] implement in any integrated ERP software
The comparison presented in table 2 is quite elaborative in REFERENCES
terms of attribute’s timeliness, data dependency and [1] P. Kanellis, Digital crime and forensic science in cyberspace:
accuracy. Entity extraction is first attribute which has high IGI Global, 2006.
data dependency with minimum time consumption and [2] W. Lee, S. J. Stolfo, and K. W. Mok, "A data mining
optimal accuracy. Clustering is one of the renowned framework for building intrusion detection models," in Security and
techniques in data mining, but for crime detection its time Privacy, 1999. Proceedings of the 1999 IEEE Symposium on, 1999, pp.
120-132.
consumption is moderate having less data dependency and
[3] A. Gray, S. MacDonell, and P. Sallis, "Software forensics:
is accurate. Association rule mining is yet another Extending authorship analysis techniques to computer programs," 1997.
technique used for crime detection having moderate [4] O. De Vel, A. Anderson, M. Corney, and G. Mohay, "Mining
timeliness and data dependency with accurate results. e-mail content for author identification forensics," ACM Sigmod Record,
Sequential pattern mining is used as a technique. vol. 30, pp. 55-64, 2001.
According to the comparison it has moderate time [5] G. Wang, H. Chen, and H. Atabakhsh, "Automatically
consumption with huge amount of data dependency but detecting deceptive criminal identities," Communications of the ACM,
vol. 47, pp. 70-76, 2004.
gives accuracy in results. The technique named deviation
[6] J. Hosseinkhani, S. Ibrahim, S. Chuprat, and J. H. Naniz, "Web
detection has less time consumption moderate data Crime Mining by Means of Data Mining Techniques," Research Journal
dependency and accuracy is optimal. Similarly, of Applied Sciences, Engineering and Technology, vol. 7, pp. 2027-2032,
classification has less timeliness, and data is dependent on 2014.
complete training dataset and accuracy depends on [7] P. Thongtae and S. Srisuk, "An analysis of data mining
applications in crime domain," in Computer and Information Technology
availability of data. String comparator has more time Workshops, 2008. CIT Workshops 2008. IEEE 8th International
consumption, moderate data dependency with accurate Conference on, 2008, pp. 122-126.
results and social network analysis technique has moderate [8] H. Chen, W. Chung, Y. Qin, M. Chau, J. J. Xu, G. Wang, et
time consumption and data dependency with unidentified al., "Crime data mining: an overview and case studies," in Proceedings of
accuracy level. the 2003 annual national conference on Digital government research,
2003, pp. 1-5.
[9] V. Pinheiro, V. Furtado, T. Pequeno, and D. Nogueira,
V. CONCLUSION "Natural language processing based on semantic inferentialism for
We have discussed the use of crime data mining using extracting crime information from text," in Intelligence and Security
Informatics (ISI), 2010 IEEE International Conference on, 2010, pp. 19-
different techniques like clustering, association rules, 24.
sequential patterns and others. We identified eight [10] U. Fayyad and R. Uthurusamy, "Evolving data into mining
techniques along with their strengths and weakness. Our solutions for insights," Communications of the ACM, vol. 45, pp. 28-31,
contribution here was to provide comparative analysis of 2002.
these techniques. This analysis is helpful as a researcher [11] S. V. Nath, "Crime pattern detection using data mining," in
guide for studying and identifying crime data mining Web Intelligence and Intelligent Agent Technology Workshops, 2006.
WI-IAT 2006 Workshops. 2006 IEEE/WIC/ACM International
techniques. There are some limitations of each technique Conference on, 2006, pp. 41-44.
including computational efforts, cost, structured data and [12] U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth, "From data
rules. On the basis of three attributed criteria: timeliness, mining to knowledge discovery in databases," AI magazine, vol. 17, p.
data dependency and accuracy, we have analyzed each 37, 1996.
technique. Based on the strength and weakness identified, [13] J. Hosseinkhani, M. Koochakzaei, S. Keikhaee, and J. H.
it is concluded that each technique is specific to crime Naniz, "Detecting suspicion information on the Web using crime data
mining techniques," International Journal of Advanced Computer Science
detection data mining scenario and has significant and Information Technology, vol. 3, pp. 32-41, 2014.
attributes. One of the facts occurred during this research is [14] H. Chen, W. Chung, J. J. Xu, G. Wang, Y. Qin, and M. Chau,
that crime data they require data mining experts and data "Crime data mining: a general framework and some examples,"
analysts equipped with sufficient knowledge of data Computer, vol. 37, pp. 50-56, 2004.
mining and they need to collaborate with detectives in early [15] H. Hassani, X. Huang, E. S. Silva, and M. Ghodsi, "A review
of data mining applications in crime," Statistical Analysis and Data
phases of crime detection. Mining: The ASA Data Science Journal, vol. 9, pp. 139-154, 2016.

5
 University of Sindh Journal of Information and Communication Technology (USJICT), Vol.2(1), pg.: 1-6

[16] H. Kargupta, K. Liu, and J. Ryan, "Privacy sensitive [41] E. A. Wan, "Neural network classification: A Bayesian
distributed data mining from multi-party data," in International interpretation," IEEE Transactions on Neural Networks, vol. 1, pp. 303-
Conference on Intelligence and Security Informatics, 2003, pp. 336-342. 305, 1990.
[17] J. Han, J. Pei, and M. Kamber, Data mining: concepts and [42] B. Widrow, D. E. Rumelhart, and M. A. Lehr, "Neural
techniques: Elsevier, 2011. networks: applications in industry, business and science,"
[18] G. Gupta, Introduction to data mining with case studies: PHI Communications of the ACM, vol. 37, pp. 93-106, 1994.
Learning Pvt. Ltd., 2014. [43] J. Mena, Investigative data mining for security and criminal
[19] M. Chau, J. J. Xu, and H. Chen, "Extracting meaningful detection: Butterworth-Heinemann, 2003.
entities from police narrative reports," in Proceedings of the 2002 annual [44] A. M. Fard and M. Ester, "Collaborative mining in multiple
national conference on Digital government research, 2002, pp. 1-5. social networks data for criminal group discovery," in Computational
[20] S. Baluja, V. O. Mittal, and R. Sukthankar, "Applying Science and Engineering, 2009. CSE'09. International Conference on,
Machine Learning for High‐Performance Named‐Entity Extraction," 2009, pp. 582-587.
Computational Intelligence, vol. 16, pp. 586-595, 2000. [45] M. K. Sparrow, "The application of network analysis to
[21] A. Borthwick, J. Sterling, E. Agichtein, and R. Grishman, criminal intelligence: An assessment of the prospects," Social networks,
"Exploiting diverse knowledge sources via maximum entropy in named vol. 13, pp. 251-274, 1991.
entity recognition," in Proc. of the Sixth Workshop on Very Large [46] A. Iriberri and G. Leroy, "Natural language processing and e-
Corpora, 1998. government: extracting reusable crime report information," in
[22] S. Miller, M. Crystal, H. Fox, L. Ramshaw, R. Schwartz, R. Information Reuse and Integration, 2007. IRI 2007. IEEE International
Stone, et al., "Algorithms that learn to extract information: Bbn: Tipster Conference on, 2007, pp. 221-226.
phase iii," in Proceedings of a workshop on held at Baltimore, Maryland: [47] K. Chan and J. Liebowitz, "The synergy of social network
October 13-15, 1998, 1998, pp. 75-89. analysis and knowledge mapping: a case study," International journal of
[24] I. H. Witten, Z. Bray, M. Mahoui, and W. J. Teahan, "Using management and decision making, vol. 7, pp. 19-35, 2005.
language models for generic entity extraction," in Proceedings of the [48] Maartin B., et. al., Performance Evaluation of a Natural Language
ICML Workshop on Text Mining, 1999. Processing approach applied in White Collar Crime, Springer adfa,
[25] R. V. Hauck, H. Atabakhsb, P. Ongvasith, H. Gupta, and H. p 1., Berlin, 2011
Chen, "Using Coplink to analyze criminal-justice data," Computer, vol.
35, pp. 30-37, 2002.
[26] R. T. Ng and J. Han, "E cient and E ective Clustering Methods
for Spatial Data Mining," in Proc. of, 1994, pp. 144-155.
[27] H. Yun, D. Ha, B. Hwang, and K. H. Ryu, "Mining association
rules on significant rare data using relative support," Journal of Systems
and Software, vol. 67, pp. 181-191, 2003.
[28] P.-N. Tan, Introduction to data mining: Pearson Education
India, 2006.
[29] J. Ayres, J. Flannick, J. Gehrke, and T. Yiu, "Sequential
pattern mining using a bitmap representation," in Proceedings of the
eighth ACM SIGKDD international conference on Knowledge discovery
and data mining, 2002, pp. 429-435.
[30] C. C. Aggarwal and P. S. Yu, "Outlier detection for high
dimensional data," in ACM Sigmod Record, 2001, pp. 37-46.
[31] A. Arning, R. Agrawal, and P. Raghavan, "A Linear Method
for Deviation Detection in Large Databases," in KDD, 1996, pp. 164-169.
[32] C. J. Stone, "Classification and regression trees," Wadsworth
International Group, vol. 8, pp. 452-456, 1984.
[33] J. R. Quinlan, C4. 5: programs for machine learning: Elsevier,
2014.
[34] J. R. Quinlan, "Induction of decision trees," Machine learning,
vol. 1, pp. 81-106, 1986.
[35] C. Cortes and V. Vapnik, "Support-vector networks," Machine
learning, vol. 20, pp. 273-297, 1995.
[36] P. Langley, W. Iba, and K. Thompson, "An analysis of
Bayesian classifiers," in Aaai, 1992, pp. 223-228.
[37] M. D. Richard and R. P. Lippmann, "Neural network
classifiers estimate Bayesian a posteriori probabilities," Neural
computation, vol. 3, pp. 461-483, 1991.
[38] G. P. Zhang, "Neural networks for classification: a survey,"
IEEE Transactions on Systems, Man, and Cybernetics, Part C
(Applications and Reviews), vol. 30, pp. 451-462, 2000.
[39] H. Gish, "A probabilistic approach to the understanding and
training of neural network classifiers," in Acoustics, Speech, and Signal
Processing, 1990. ICASSP-90., 1990 International Conference on, 1990,
pp. 1361-1364.
[40] P. A. Shoemaker, "A note on least-squares learning procedures
and classification by neural network models," IEEE Transactions on
Neural Networks, vol. 2, pp. 158-160, 1991.