A Dissertation Submitted in Partial Fulfilment of the Requirements for the Degree of
Doctor of Philosophy in Information and Communication Science and Engineering of the
Nelson Mandela African Institution of Science and Technology
The spate of crimes in Tanzania, as in many other countries, has been on the increase in the last
few years. The successes recorded by criminals have been attributed by improper mechanisms
for crime detection, prevention and control. Proactive measures are needed to preempt further
crimes. Frequent pattern mining stand to aid in finding emerging patterns, series, and trends
in the crime data. This will eventually help Tanzania Police Force and other law enforcement
agencies to understand crime trends and predict or forecast future occurrences and thus improve
preventive measures against crimes. FP-Growth is the most effective and most widely used algorithm
for frequent pattern mining and association rules generation. Unfortunately, studies
have shown two main weaknesses associated with this algorithm when used in the growing volumes
of crime data. Inability of the algorithm to scale-up well with the growing volumes data is
the first weakness. Second weakness of FP-Growth is associated with the nature of crime data.
Crime data consists of items (different crimes) that vary greatly in terms of frequencies of occurrence.
Some of crimes (e.g. robbery) happen so frequently and thus frequently appear in the
dataset while other crimes (e.g. killing of people with albinism, in the case of Tanzania) happens
seasonally and therefore rarely found in the dataset. Classical FP-Growth algorithm extract
frequent patterns by using single user-defined minimum support. This is the main source of the
algorithm’s challenge, especially when used in crime datasets. To tackle the challenges, this
study have proposed a multiple minimum support FP-Growth algorithm that scans the dataset
and automatically assigns minimum support values to each crime item basing on how frequently
it has appeared in the dataset. The proposed solution is based on the Shannon Entropy in which
an algorithm for obtaining multiple item support values was developed. In connection, the study
developed a working prototype that was based on the proposed approach to help in recording
crime data and mine crime patterns from data. The proposed approach was evaluated against
classical FP-Growth as well as CFPGrowth algorithm on the varying sizes crime data. Evaluation
results showed that the proposed approach in this research is more efficient in mining
frequent crime patterns, and more effective in terms of run-time and memory use. Basing on
the results found, the study recommends for further experimentation of the proposed approach
on streaming data and on distributed environments, among other recommendations as stipulated
in the dissertation.