| Abstract | User profiling has attracted an enormous number of technological methods and applications. With the increasing amount of products and services, user profiling has created opportunities to catch the attention of the user as well as achieving high user satisfaction. To provide the user what she/he wants, when and how, depends largely on understanding them. The user profile is the representation of the user and holds the information about the user. These profiles are the outcome of the user profiling.
Personalization is the adaptation of the services to meet the user’s needs and expectations. Therefore, the knowledge about the user leads to a personalized user experience. In user profiling applications the major challenge is to build and handle user profiles. In the literature there are two main user profiling methods, collaborative and the content-based. Apart from these traditional profiling methods, a number of classification and clustering algorithms have been used to classify user related information to create user profiles. However, the profiling, achieved through these works, is lacking in terms of accuracy. This is because, all information within the profile has the same influence during the profiling even though some are irrelevant user information.
In this thesis, a primary aim is to provide an insight into the concept of user profiling. For this purpose a comprehensive background study of the literature was conducted and summarized in this thesis. Furthermore, existing user profiling methods as well as the classification and clustering algorithms were investigated. Being one of the objectives of this study, the use of these algorithms for user profiling was examined. A number of classification and clustering algorithms, such as Bayesian Networks (BN) and Decision Trees (DTs) have been simulated using user profiles and their classification accuracy performances were evaluated. Additionally, a novel clustering algorithm for the user profiling, namely Multi-Dimensional Clustering (MDC), has been proposed.
The MDC is a modified version of the Instance Based Learner (IBL) algorithm. In IBL every feature has an equal effect on the classification regardless of their relevance. MDC differs from the IBL by assigning weights to feature values to distinguish the effect of the features on clustering. Existing feature weighing methods, for instance Cross Category Feature (CCF), has also been investigated. In this thesis, three feature value weighting methods have been proposed for the MDC. These methods are; MDC weight method by Cross Clustering (MDC-CC), MDC weight method by Balanced Clustering (MDC-BC) and MDC weight method by changing the Lower-limit to Zero (MDC-LZ). All of these weighted MDC algorithms have been tested and evaluated. Additional simulations were carried out with existing weighted and non-weighted IBL algorithms (i.e. K-Star and Locally Weighted Learning (LWL)) in order to demonstrate the performance of the proposed methods. Furthermore, a real life scenario is implemented to show how the MDC can be used for the user profiling to improve personalized service provisioning in mobile environments.
The experiments presented in this thesis were conducted by using user profile datasets that reflect the user’s personal information, preferences and interests. he simulations with existing classification and clustering algorithms (e.g. Bayesian Networks (BN), Naïve Bayesian (NB), Lazy learning of Bayesian Rules (LBR), Iterative Dichotomister 3 (Id3)) were performed on the WEKA (version 3.5.7) machine learning platform. WEKA serves as a workbench to work with a collection of popular learning schemes implemented in JAVA. In addition, the MDC-CC, MDC-BC and MDC-LZ have been implemented on NetBeans IDE 6.1 Beta as a JAVA application and MATLAB. Finally, the real life scenario is implemented as a Java Mobile Application (Java ME) on NetBeans IDE 7.1. All simulation results were evaluated based on the error rate and accuracy. |
|---|