M. Tech. Electrical Engineering. / Telecommunication databases contain billions of records and are among the largest in the world, reaching around 30 terabits (30 trillion bits). Data mining is a proven solution for analysing such large volumes of data where traditional methods of turning data into knowledge are impractical. However, the increasing size (scalability), complexity (complex data types) and high dimensionality of telecommunication databases pose a significant challenge for conventional data mining approaches. In this dissertation, a matrix decomposition method (Singular Value Decomposition or SVD) is used to improve data mining for subscriber classification in mobile cellular networks. Using a large real mobile network dataset, the performance of a standard data mining approach (for clustering analysis) is evaluated when it is used with, and without, matrix decomposition. The proposed approach decreases the computational cost, for a given size of data (in terms of number of rows and columns). We also demonstrate improvement of the quality of clusters, yielding the following improvements in clustering assessment indices: 2.45% in Jaccard score, 3.5% in purity, and 1.35% in efficiency. Subscribers with different behaviours in the network are classified on the basis of various features; SVD analysis on their voice, text message, and data usage patterns are also performed. The proposed data mining model can be used for business intelligence activities such as customer segmentation, traffic modelling and social network analysis.

published_or_final_version / Mathematics / Master / Master of Philosophy

One of the major tasks in Data Mining is classification. The growing of Decision Tree from data is a very efficient technique for learning classifiers. The selection of an attribute used to split the data set at each Decision Tree node is fundamental to properly classify objects; a good selection will improve the accuracy of the classification. In this paper, we study the behavior of the Decision Trees induced with 14 attribute selection measures over three data sets taken from UCI Machine Learning Repository.

Artificial Intelligence Lab, Department of MIS, University of Arizona / Crime and police report information is rapidly migrating from paper records to automated records management databases. Most mid and large sized police agencies have such systems that provide access to information by their own personnel, but lack any efficient manner by which to provide that information to other agencies. Criminals show no regard for jurisdictional boundaries and in fact take advantage of the lack of communication across jurisdictions. Federal standards initiatives such as the National Incident Based Reporting System (NIBRS, US Department of Justice 1998), are attempting to provide reporting standards to police agencies to facilitate future reporting and information sharing among agencies as these electronic reporting systems become more widespread. We integrated platform-independence, stability, scalability, and an intuitive graphical user interface to develop the COPLINK system, which is currently being deployed at Tucson Police Department (TPD). User evaluations of the application allowed us to study the impact of COPLINK on law enforcement personnel as well as to identify requirements for improving the system and extending the project. We are currently in the process of extending the functionality of COPLINK in several areas. These include textual analysis, collaboration, visualization and geo-mapping.

Artificial Intelligence Lab, Department of MIS, University of Arizona / Spiders are the software agents that search engines use to collect content for their databases. We investigated algorithms to improve the performance of vertical search engine spiders. The investigation addressed three approaches: a breadth-first graph-traversal algorithm with no heuristics to refine the search process, a best-first traversal algorithm that used a hyperlink-analysis heuristic, and a spreading-activation algorithm based on modeling the Web as a neural network.

Preprint - For final version, see Nicholson, S. (2006). The basis for bibliomining: Frameworks for bringing together usage-based data mining and bibliometrics through data warehousing in digital library services. Information Processing and Management 42(3), 785-804. Over the past few years, data mining has moved from corporations to other organizations. This paper looks at the integration of data mining in digital library services. First, bibliomining, or the combination of bibliometrics and data mining techniques to understand library services, is defined and the concept explored. Second, the conceptual frameworks for bibliomining from the viewpoint of the library decision-maker and the library researcher are presented and compared. Finally, a research agenda to resolve many of the common bibliomining issues and to move the field forward in a mindful manner is developed. The result is not only a roadmap for understanding the integration of data mining in digital library services, but also a template for other cross-discipline data mining researchers to follow for systematic exploration in their own subject domains.

Artificial Intelligence Lab, Department of MIS, University of Arizona / This paper discusses several data mining algorithms and techniques that we have developed at the University of Arizona Artificial Intelligence Lab.We have implemented these algorithms and techniques into several prototypes, one of which focuses on medical information developed in cooperation with the National Cancer Institute (NCI) and the University of Illinois at Urbana-Champaign.We propose an architecture for medical knowledge information systems that will permit data mining across several medical information sources and discuss a suite of data mining tools that we are developing to assist NCI in improving public access to and use of their existing vast cancer information collections.

A large part of the Web, today, consists of online platforms that allow their users to generate digital content. They include online social networks, multimedia-sharing websites, blogging platforms, and online discussion boards, to name a few examples. Users of those platforms generate content in the form of digital items (e.g. documents, images, or videos), inspect content generated by others, and, finally, interact with each other (e.g. by commenting on each other's generated items). For the social process of information exchange they enable, such platforms are customarily referred to as `social media'. Activity on social media is largely spontaneous and uncoordinated, but it is not random; users choose the discussions they engage in and who they interact with, and their choices and actions reflect what they find important. In this thesis, we define and quantify notions of importance for items, users, and social connections between users, and, based on those definitions, propose efficient algorithms to detect important instances of social media activity. Our description of the algorithms is accompanied with experimental studies that showcase their performance on real datasets in terms of efficiency and effectiveness.

A large part of the Web, today, consists of online platforms that allow their users to generate digital content. They include online social networks, multimedia-sharing websites, blogging platforms, and online discussion boards, to name a few examples. Users of those platforms generate content in the form of digital items (e.g. documents, images, or videos), inspect content generated by others, and, finally, interact with each other (e.g. by commenting on each other's generated items). For the social process of information exchange they enable, such platforms are customarily referred to as `social media'. Activity on social media is largely spontaneous and uncoordinated, but it is not random; users choose the discussions they engage in and who they interact with, and their choices and actions reflect what they find important. In this thesis, we define and quantify notions of importance for items, users, and social connections between users, and, based on those definitions, propose efficient algorithms to detect important instances of social media activity. Our description of the algorithms is accompanied with experimental studies that showcase their performance on real datasets in terms of efficiency and effectiveness.

No description available.