Katarzyna A. Tarnowska

From Knowledge Discovery to Customer Attrition

This article presents a novel approach to handle customer attrition problem with knowledge discovery methods. The data mining is performed on the customer feedback data, which was labelled by means of temporal transactional invoice data in terms of customer activity. The problem was raised within industry-academia collaboration project at University of North Carolina at Charlotte by one of the companies from the heavy equipment repair industry. They expressed interest in gaining better insight into this problem, already having their own active CRM program implemented. The goal and motivation within this topic is to determine whether there are markers in the sales trends that might suggest a customer is getting ready to defect. Observing the behavior of customers who left a company, one might be able to identify customers who may leave as well.

Tinnitus Treatment as a Problem Area

This chapter presents the decision problem area which will be supported with a recommender system technology, that is, tinnitus diagnosis and treatment. It will introduce the problem of tinnitus and next, the successful method of treatment applied by doctor P. Jastreboff. At the end of this chapter major results from the treatment will be showed, along with possible new challenges, which can be handled with the help of information technology.

Knowledge Discovery Approach for Recommendation

This chapter presents concepts of action rules, proposed by Ras and Wieczorkowska in 2000 [RW00] and meta-actions, as a proposed approach for building a rule-based (knowledge-based) recommender system for tinnitus treatment, and motivation for using such methods. It also presents theoretical foundations and algorithms for automatic action rules extraction, as methods for domain knowledge discovery.

Recommender Solutions Overview

This chapter aims at providing an overview of RS technology, describing different types of RS, with emphasis on choosing the right approach for the system supporting tinnitus treatment and justifying particular choice. Current generation of recommendation methods is presented in division to four main categories: collaborative, content-based, knowledge-based, hybrid.

Final Conclusions and Future Work

This book presented a process of analysis, design and prototype implementation of RECTIN recommender system, as a solution to the problem of supporting tinnitus treatment based on Tinnitus Retraining Therapy in a medical facility. Proposed approach in supporting physicians’ diagnosis and treatment decisions addresses scarcity of expert knowledge, time restrictions in today’s medical practice and the need for more efficient evaluation of different treatment methods. Such system can provide accurate support at any time, with full consideration of individual patient profiles, including: demographics, medical history, and tinnitus background.

RECTIN System Design

RECTIN system (shortcut for RECommender for TINnitus) is a prototyping method proposed within this work to verify the hypothesis of possibility to apply information technology in supporting physicians, dealing with tinnitus patients, in the diagnosis and treatment. This chapter describes major steps in the system design: analysis with main use cases for the system, deployment architecture, with detailed description of each component and implementation project, including transactional database and application. Also, knowledge engineering approach is presented, along with detailed description of raw dataset of tinnitus patients and visits, which was made available to the authors. This section also introduces approach taken to data preprocessing so that to make it useful for creating a knowledge base, on which data mining can be performed.

Experiment 1: Classifiers

Following the dataset preprocessing, the next step in implementing RECTIN is classification module development. The classification module will use a model built on historical patients’ data, in order to support physicians in suggesting optimal treatment approach for new patients. Categorization is rather easy and relatively broad. However, a specific approach within each category varies. Before implementing this module, it is necessary to extract new, useful features and conduct experiments in order to obtain the most accurate classifier on the prepared dataset. It is assumed to reiterate the step of feature development in order to obtain the best combination of feature extraction/selection method and the prediction method. This involves the calibration and tuning of prediction methods, as well as comparing them and evaluating in terms of accuracy, F-score and confusion matrix.

Experiment 4: Treatment Rules Enhancement

Experiments on action rules, described in the previous section, did not consider temporal dependencies between patient’s visits (that is, at what relative point in timeline particular actions were taken). On the other hand, it would be effective to search for temporal dependencies between particular treatment actions and their observable results in the form of changed score denoting tinnitus severity. New approach should allow to assess treatment action effectiveness in temporal terms and consider their sequence. For example, some actions might take effect after some time elapse and not be effective in the short-term.

Experiment 2: Diagnostic Rules

Association rule tasks were defined in order to discover common patterns in patients’ visits dataset. Before defining data mining task on rule discovery, it is necessary to formulate an analytical problem in the first place. The main examined associations of interest would be factors affecting patient’s category. Such discovered association rules can be regarded as decision rules, supporting classifier developed in the previous step.

Experiment 3: Treatment Rules

Action rules should help in choosing treatment actions in the course of Tinnitus Retraining Therapy in subsequent visits. In order to understand the process of treatment and formulate appropriate data mining tasks in LISp-Miner, it is necessary to identify treatments actions taken by the doctor to improve tinnitus/hyperacusis patient’s condition.