Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Computer Science


Charles X. Ling


Web search engines have been adopted by most universities for searching webpages in their own domains. Basically, a user sends keywords to the search engine and the search engine returns a flat ranked list of webpages. However, in university search, user queries are usually related to topics. Simple keyword queries are often insufficient to express topics as keywords. On the other hand, most E-commerce sites allow users to browse and search products in various hierarchies. It would be ideal if hierarchical browsing and keyword search can be seamlessly combined for university search engines. The main difficulty is to automatically classify and rank a massive number of webpages into the topic hierarchies for universities.

In this thesis, we use machine learning and data mining techniques to build a novel hybrid search engine with integrated hierarchies for universities, called SEEU (Search Engine with hiErarchy for Universities).

Firstly, we study the problem of effective hierarchical webpage classification. We develop a parallel webpage classification system based on Support Vector Machines. With extensive experiments on the well-known ODP (Open Directory Project) dataset, we empirically demonstrate that our hierarchical classification system is very effective and outperforms the traditional flat classification approaches significantly.

Secondly, we study the problem of integrating hierarchical classification into the ranking system of keywords-based search engines. We propose a novel ranking framework, called ERIC (Enhanced Ranking by hIerarchical Classification), for search engines with hierarchies. Experimental results on four large-scale TREC (Text REtrieval Conference) web search datasets show that our ranking system with hierarchical classification outperforms the traditional flat keywords-based search methods significantly.

Thirdly, we propose a novel active learning framework to improve the performance of hierarchical classification, which is important for ranking webpages in hierarchies. From our experiments on the benchmark text datasets, we find that our active learning framework can achieve good classification performance yet save a considerable number of labeling effort compared with the state-of-the-art active learning methods for hierarchical text classification.

Fourthly, based on the proposed classification and ranking methods, we present a novel hierarchical classification framework for mining academic topics from university webpages. We build an academic topic hierarchy based on the commonly accepted Wikipedia academic disciplines. Based on this hierarchy, we train a hierarchical classifier and apply it to mine academic topics. According to our comprehensive analysis, the academic topics mined by our method are reasonable and consistent with the real-world topic distribution in universities.

Finally, we combine all the proposed techniques together and implement the SEEU search engine. According to two usability studies conducted in the ECE and the CS departments at our university, SEEU is favored by the majority of participants.

To conclude, the main contribution of this thesis is a novel search engine, called SEEU, for universities. We discuss the challenges toward building SEEU and propose effective machine learning and data mining methods to tackle them. With extensive experiments on well-known benchmark datasets and real-world university webpage datasets, we demonstrate that our system is very effective. In addition, two usability studies of SEEU in our university show that SEEU has a great promise for university search.