Adaptive algorithms for real-world transactional data mining.

Abstract

The accurate identification of the right customer to target with the right product at the right time, through the right channel, to satisfy the customer’s evolving needs, is a key performance driver and enhancer for businesses. Data mining is an analytic process designed to explore usually large amounts of data (typically business or market related) in search of consistent patterns and/or systematic relationships between variables for the purpose of generating explanatory/predictive data models from the detected patterns. It provides an effective and established mechanism for accurate identification and classification of customers. Data models derived from the data mining process can aid in effectively recognizing the status and preference of customers - individually and as a group. Such data models can be incorporated into the business market segmentation, customer targeting and channelling decisions with the goal of maximizing the total customer lifetime profit. However, due to costs, privacy and/or data protection reasons, the customer data available for data mining is often restricted to verified and validated data,(in most cases,only the business owned transactional data is available). Transactional data is a valuable resource for generating such data models. Transactional data can be electronically collected and readily made available for data mining in large quantity at minimum extra cost. Transactional data is however, inherently sparse and skewed. These inherent characteristics of transactional data give rise to the poor performance of data models built using customer data based on transactional data. Data models for identifying, describing, and classifying customers, constructed using evolving transactional data thus need to effectively handle the inherent sparseness and skewness of evolving transactional data in order to be efficient and accurate. Using real-world transactional data, this thesis presents the findings and results from the investigation of data mining algorithms for analysing, describing, identifying and classifying customers with evolving needs. In particular, methods for handling the issues of scalability, uncertainty and adaptation whilst mining evolving transactional data are analysed and presented. A novel application of a new framework for integrating transactional data binning and classification techniques is presented alongside an effective prototype selection algorithm for efficient transactional data model building. A new change mining architecture for monitoring, detecting and visualizing the change in customer behaviour using transactional data is proposed and discussed as an effective means for analysing and understanding the change in customer buying behaviour over time. Finally, the challenging problem of discerning between the change in the customer profile (which may necessitate the effective change of the customer’s label) and the change in performance of the model(s) (which may necessitate changing or adapting the model(s)) is introduced and discussed by way of a novel flexible and efficient architecture for classifier model adaptation and customer profiles class relabeling.