Degree Type

Dissertation

Date of Award

2018

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

Major

Computer Engineering

First Advisor

Arun K. Somani

Abstract

In this work we give solutions to two key knowledge discovery problems for the Transactional Data model: Cluster analysis and Itemset mining. By knowledge discovery in context of these two problems, we specifically mean novel and useful ways of extracting clusters and itemsets from transactional data. Transactional Data model is widely used in a variety of applications. In cluster analysis the goal is to find clusters of similar transactions in the data with the collective properties of each cluster being unique. We propose the first clustering algorithm for transactional data which uses the latest model definition. All previously proposed algorithms did not use the important utility information in the data. Our novel technique effectively solves this problem. We also propose two new cluster validation metrics based on the criterion of high utility patterns. When comparing our technique with competing algorithms, we miss much fewer high utility patterns of importance than them.

Itemset mining is the problem of searching for repeating patterns of high importance in the data. We show that the current model for itemset mining leads to information loss. It ignores the presence of clusters in the data. We propose a new itemset mining model which incorporates the cluster structure information. This allows the model to make predictions for future itemsets. We show that our model makes accurate predictions successfully, by discovering 30-40% future itemsets in most experiments on two benchmark datasets with negligible inaccuracies. There are no other present itemset prediction models, so accurate prediction is an accomplishment of ours.

We provide further theoretical improvements in our model by making it capable of giving predictions for specific future windows by using time series forecasting. We also perform a detailed analysis of various clustering algorithms and study the effect of the Big Data phenomenon on them. This inspired us to further refine our model based on a classification problem design. This addition allows the mining of itemsets based on maximizing a customizable objective function made of different prediction metrics. The final framework design proposed by us is the first of its kind to make itemset predictions by using the cluster structure. It is capable of adapting the predictions to a specific future window and customizes the mining process to any specified prediction criterion. We create an implementation of the framework on a Web analytics data set, and notice that it successfully makes optimal prediction configuration choices with a high accuracy of "0.895".

Copyright Owner

Piyush Lakhawat

Language

en

File Format

application/pdf

File Size

127 pages

Share

COinS