Full text
| |
| Author(s): |
Ricardo Marcondes Marcacini
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2014-10-14 |
| Examining board members: |
Solange Oliveira Rezende;
Heloisa de Arruda Camargo;
Alexandre Gonçalves Evsukoff;
Eduardo Raul Hruschka;
Luiz Satoru Ochi
|
| Advisor: | Solange Oliveira Rezende; Eduardo Raul Hruschka |
| Abstract | |
Hierarchical text clustering methods are very useful to analyze the implicit knowledge in textual collections, enabling the organization of textual documents into clusters and subclusters to facilitate the knowledge browsing at various levels of granularity. Such methods are classified as unsupervised machine learning, since the clustering models are obtained only by observing regularities of textual data without human supervision. Traditional clustering methods assume that the text collection is represented only by the technical information, i.e., words and phrases extracted directly from the texts. On the other hand, in many text clustering tasks there is an additional and valuable knowledge about the problem domain, usually extracted by an advanced process with support of the domain experts. Due to the high cost of obtaining such expert knowledge, this additional information is defined as privileged and is usually available to represent only a subset of the textual documents. Recently, a new machine learning paradigm called LUPI (Learning Using Privileged Information) was proposed by Vapnik to incorporate privileged information into supervised learning methods. In this thesis, the LUPI paradigm was extended to unsupervised learning setting, in particular for hierarchical text clustering. We propose and evaluate approaches to deal with different challenges for clustering tasks, involving the extraction and structuring of privileged information and using this additional information to refine or correct clustering models. The proposed approaches were effective in (i) consensus clustering, allowing to combine different clustering solutions and textual representations; (ii) metric learning, in which more robust proximity measures are obtained from privileged information; and (iii) model selection, in which the privileged information is exploited to identify the relevant structures of hierarchical clustering. All the approaches presented in this thesis were investigated in an incremental clustering scenario, allowing its use in practical applications that require computational efficiency as well as deal with high frequency of publication of new textual knowledge. (AU) | |
| FAPESP's process: | 10/20564-8 - Unsupervised multi-view learning in incremental hierarchical clustering of text documents |
| Grantee: | Ricardo Marcondes Marcacini |
| Support Opportunities: | Scholarships in Brazil - Doctorate |