Temporal data mining : algorithms, language and system for temporal association rules.
Studies on data mining are being pursued in many different research areas, such as
Machine Learning, Statistics, and Databases. The work presented in this thesis is based
on the database perspective of data mining. The main focuses are on the temporal
aspects of data mining problems, especially association rule discovery, and issues on
the integration of data mining and database systems.
Firstly, a theoretical framework for temporal data mining is proposed in this thesis.
Within this framework, not only potential patterns but also temporal features
associated with the patterns are expected to be discovered. Calendar time expressions
are suggested to represent temporal features and the minimum frequency of patterns is
introduced as a new threshold in the model of temporal data mining. The framework
also emphasises the necessary components to support temporal data mining tasks.
As a specialisation of the proposed framework, the problem of mining temporal
association rules is investigated. The methodology adopted in this thesis is eventually
discovering potential temporal rules by alternatively using special search techniques
for various restricted problems in an interactive and iterative process. Three forms of
interesting mining tasks for temporal association rules with certain constraints are
identified. These tasks are the discovery of valid time periods of association rules, the
discovery of periodicities of association rules, and the discovery of association rules
with temporal features. The search techniques and algorithms for those individual
tasks are developed and presented in this thesis.
Finally, an integrated query and mining system (IQMS) is presented in this thesis,
covering the description of an interactive query and mining interface (IQMI) supplied
by the IQMS system, the presentation of an SQL-like temporal mining language
(TML) with the ability to express various data mining tasks for temporal association
rules, and the suggestion of an IQMI-based interactive data mining process. The
implementation of this system demonstrates an alternative approach for the integration
of the DBMS and data mining functions.