Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Nasraoui, Olfa

Committee Co-Chair (if applicable)

Altiparmak, Nihat

Committee Member

Altiparmak, Nihat

Committee Member

Frigui, Hichem

Committee Member

Sheridan, Mary P.

Committee Member

Yampolskiy, Roman

Author's Keywords

stream clustering; topic modeling


Recent advances and widespread usage of online web services and social media platforms, coupled with ubiquitous low cost devices, mobile technologies, and increasing capacity of lower cost storage, has led to a proliferation of Big data, ranging from, news, e-commerce clickstreams, and online business transactions to continuous event logs and social media expressions. These large amounts of online data, often referred to as data streams, because they get generated at extremely high throughputs or velocity, can make conventional and classical data analytics methodologies obsolete. For these reasons, the issues of management and analysis of data streams have been researched extensively in recent years. The special case of social media Big Data brings additional challenges, particularly because of the unstructured nature of the data, specifically free text. One classical approach to mine text data has been Topic Modeling. Topic Models are statistical models that can be used for discovering the abstract ``topics'' that may occur in a corpus of documents. Topic models have emerged as a powerful technique in machine learning and data science, providing a great balance between simplicity and complexity. They also provide sophisticated insight without the need for real natural language understanding. However they have not been designed to cope with the type of text data that is abundant on social media platforms, but rather for traditional medium size corpora consisting of longer documents, adhering to a specific language and typically spanning a stable set of topics. Unlike traditional document corpora, social media messages tend to be very short, sparse, noisy, and do not adhere to a standard vocabulary, linguistic patterns, or stable topic distributions. They are also generated at high velocity that impose high demands on topic modeling; and their evolving or dynamic nature, makes any set of results from topic modeling quickly become stale in the face of changes in the textual content and topics discussed within social media streams. In this dissertation, we propose an integrated topic modeling framework built on top of an existing stream-clustering framework called Stream-Dashboard, which can extract, isolate, and track topics over any given time period. In this new framework, Stream Dashboard first clusters the data stream points into homogeneous groups. Then data from each group is ushered to the topic modeling framework which extracts finer topics from the group. The proposed framework tracks the evolution of the clusters over time to detect milestones corresponding to changes in topic evolution, and to trigger an adaptation of the learned groups and topics at each milestone. The proposed approach to topic modeling is different from a generic Topic Modeling approach because it works in a compartmentalized fashion, where the input document stream is split into distinct compartments, and Topic Modeling is applied on each compartment separately. Furthermore, we propose extensions to existing topic modeling and stream clustering methods, including: an adaptive query reformulation approach to help focus on the topic discovery with time; a topic modeling extension with adaptive hyper-parameter and with infinite vocabulary; an adaptive stream clustering algorithm incorporating the automated estimation of dynamic, cluster-specific temporal scales for adaptive forgetting to help facilitate clustering in a fast evolving data stream. Our experimental results show that the proposed adaptive forgetting clustering algorithm can mine better quality clusters; that our proposed compartmentalized framework is able to mine topics of better quality compared to competitive baselines; and that the proposed framework can automatically adapt to focus on changing topics using the proposed query reformulation strategy.