Giles Oatley

Crimes analysis software: ‘pins in maps’, clustering and Bayes net prediction

The OVER Project was a collaboration between West Midlands Police, UK, the Centre for Adaptive Systems, and Psychology Division, from the University of Sunderland. The Project was developed primarily to assist the Police with the high volume crime, burglary from dwelling houses. A developed software system enables the trending of historical data, the testing of ‘short term’ hunches, and the development of ‘medium’ and long term’ strategies to burglary and crime reduction, based upon victim, offender, location and details of victimisations. The software utilises mapping and visualisation tools and is capable of a range of sophisticated predictions, tying together statistical techniques with theories from forensic psychology and criminology. The statistical methods employed (including multi-dimensional scaling, binary logistic regression) and ‘data-mining’ technologies (including neural networks) are used to investigate the impact of the types of evidence available and to determine the causality in this domain. The final predictions on the likelihood of burglary are calculated by combining all of the varying sources of evidence into a Bayesian belief network. This network is embedded in the developed software system, which also performs data cleansing and data transformation for presentation to the developed algorithms. It is important that derived statistics from the software and predictions are interpretable by the intended users of the decision support system, namely Police sector managers, and this paper includes some of the design decisions based upon the forensic psychology and criminology literature, including the graphical representation of geographic data and presentation of results of analyses.

Decision support systems for police: Lessons from the application of data mining techniques to soft forensic evidence

The paper sets out the challenges facing the Police in respect of the detection and prevention of the volume crime of burglary. A discussion of data mining and decision support technologies that have the potential to address these issues is undertaken and illustrated with reference the authors’ work with three Police Services. The focus is upon the use of “soft” forensic evidence which refers to modus operandi and the temporal and geographical features of the crime, rather than “hard” evidence such as DNA or fingerprint evidence. Three objectives underpin this paper. First, given the continuing expansion of forensic computing and its role in the emergent discipline of Crime Science, it is timely to present a review of existing methodologies and research. Second, it is important to extract some practical lessons concerning the application of computer science within this forensic domain. Finally, from the lessons to date, a set of conclusions will be advanced, including the need for multidisciplinary input to guide further developments in the design of such systems. The objectives are achieved by first considering the task performed by the intended systems users. The discussion proceeds by identifying the portions of these tasks for which automation would be both beneficial and feasible. The knowledge discovery from databases process is then described, starting with an examination of the data that police collect and the reasons for storing it. The discussion progresses to the development of crime matching and predictive knowledge which are operationalised in decision support software. The paper concludes by arguing that computer science technologies which can support criminal investigations are wide ranging and include geographical information systems displays, clustering and link analysis algorithms and the more complex use of data mining technology for profiling crimes or offenders and matching and predicting crimes. We also argue that knowledge from disciplines such as forensic psychology, criminology and statistics are essential to the efficient design of operationally valid systems.

Polynomial-Fuzzy Decision Tree Structures for Classifying Medical Data

Decision tree induction has been studied extensively in machine learning as a solution for classification problems. The way the linear decision trees partition the search space is found to be comprehensible and hence appealing to data modelers. Comprehensibility is an important aspect of models used in medical data mining as it determines model credibility and even acceptability. In the practical sense though, inordinately long decision trees compounded by replication problems detracts from comprehensibility. This demerit can be partially attributed to their rigid structure that is unable to handle complex non-linear or/and continuous data. To address this issue we introduce a novel hybrid multivariate decision tree composed of polynomial, fuzzy and decision tree structures. The polynomial nature of these multivariate trees enable them to perform well in non-linear territory while the fuzzy members are used to squash continuous variables. By trading-off comprehensibility and performance using a multi-objective genetic programming optimization algorithm, we can induce polynomial-fuzzy decision trees (PFDT) that are smaller, more compact and of better performance than their linear decision tree (LDT) counterparts. In this paper we discuss the structural differences between PFDT and LDT (C4.5) and compare the size and performance of their models using medical data.

Matching Crimes Using Burglars' Modus Operandi: A Test of Three Models

’Hard’ forensic evidence (eg DNA) may be the best means of linking crimes, but it is often absent at burglary crime scenes. Modus operandi information is always present to some degree, but little is known of its significance in matching burglaries. This paper evaluates the ability of three algorithms to match a target crime to the actual offender within a database of 966 offences. The first (RCPA) uses only MO information, the second (RPAL) only temporal and geographic data and a third (COMBIN) is a combination of the two. A score of one indicates a perfect match between the target crime and the case selected by the algorithm. The lowest possible rank is 965 showing that 965 cases were selected before the target offence. The RPAL and COMBIN each achieve a perfect match for 24 per cent of the crimes and succeed in matching over half of the crimes at a score of 10 or less. For prolific offenders, using MO information alone is better than temporal and geographic data, although the best performance is achieved when in combination. Behavioural, spatial and temporal information is collected by many Police Services. The value and means of utilising such data in linking crimes is clearly demonstrated.

Matching and Predicting Crimes

Our central aim is the development of decision support systems based on appropriate technology for such purposes as profiling single and series of crimes or offenders, and matching and predicting crimes.

Polynomial-fuzzy decision tree structures for classifying medical data

Decision tree induction has been studied extensively in machine learning as a solution for classification problems. The way the linear decision trees partition the search space is found to be comprehensible and hence appealing to data modelers. Comprehensibility is an important aspect of models used in medical data mining as it determines model credibility and even acceptability. In the practical sense though, inordinately long decision trees compounded by replication problems detracts from comprehensibility. This demerit can be partially attributed to their rigid structure that is unable to handle complex non-linear or/and continuous data. To address this issue we introduce a novel hybrid multivariate decision tree composed of polynomial, fuzzy and decision tree structures. The polynomial nature of these multivariate trees enable them to perform well in non-linear territory while the fuzzy members are used to squash continuous variables. By trading-off comprehensibility and performance using a multi-objective genetic programming optimization algorithm, we can induce polynomial-fuzzy decision trees (PFDT) that are smaller, more compact and of better performance than their linear decision tree (LDT) counterparts. In this paper we discuss the structural differences between PFDT and LDT (C4.5) and compare the size and performance of their models using medical data.

ONAR: An Ontologies-based Service Oriented Application Integration Framework

The evolving technologies of Semantic Web and Web services are providing new means for application integration frameworks. The need for semantically enriched information exchange over the flexible environment of the internet provides a valuable enhancement to traditional methods and technologies for Enterprise Application Integration. However the utilization of the Semantic Web and Service Oriented Architecture (SOA) is not as straightforward as it appears and has specific limitations and inefficiencies due to the fact that is was not originally designed for that purpose. This paper presents a methodology that aims at the exploitation of these two technologies and the definition of an ontologies based enterprise application integration framework (ONAR).

Applying the Concept of Revictimization: Using Burglars' Behaviour to Predict Houses at Risk of Future Victimization:

A good predictor of a domestic burglary is whether the property suffered a prior victimization. Using officially reported burglaries, most houses appear to be victimized once and most repeat victims suffer twice only. Defining high-risk properties by waiting for the second burglary has its operational limitations. A police database of burglaries over 45 months is examined to explore whether the modus operandi distinguishes houses burgled once only, from those suffering a revictimization. The use of force, searching behaviour, type of property, place of entry, place of exit, alarm activation and use of a bogus official method of entry are discriminating features. Comparing non-repeats with ‘quick’ Repeats (ie within 365 days), searching behaviour, type of property, entry method and a bogus official strategy are discriminating features. Survival analyses on the latter group reveals that ramming and removing glass are significantly associated with being revictimized sooner rather than later. Conversely, exit via a window indicates a longer period to revictimization. The findings demonstrate the value of crime scene information held by the police and, when guided by an appropriate criminological or operational framework, the benefits of more substantive analyses to prevention and detection initiatives.

R U :-) or :-( ? Character- vs. Word-Gram Feature Selection for Sentiment Classification of OSN Corpora

Binary sentiment classification, or sentiment analysis, is the task of computing the sentiment of a document, i.e. whether it contains broadly positive or negative opinions. The topic is well-studied, and the intuitive approach of using words as classification features is the basis of most techniques documented in the literature. The alternative character n-gram language model has been applied successfully to a range of NLP tasks, but its effectiveness at sentiment classification seems to be under-investigated, and results are mixed. We present an investigation of the application of the character n-gram model to text classification of corpora from online social networks, the first such documented study, where text is known to be rich in so-called unnatural language, also introducing a novel corpus of Facebook photo comments. Despite hoping that the flexibility of the character n-gram approach would be well-suited to unnatural language phenomenon, we find little improvement over the baseline algorithms employing the word n-gram language model.

A Case-Based Reasoning Tool For Vibration Analysis

This paper describes the development of a case-based reasoning (CBR) tool for vibration analysis, the Vibration Case Library (VCL). The system is to help practicing engineers access similar cases while attempting to diagnose actual and potential faults on machines. Of especial interest is the novel calculation of the similarity metric in the complex domain of vibration analysis. This is achieved by means of optimisation of weights based upon analysis of retrieval accuracy (rank ordered lists), using a variant of the Kendal Tau coefficient. Representation is of complex three-dimensional objects and their environments, and is orientated towards high precision retrieval of cases.