Structure-Based Supervised Term Weighting and Regularization for Text Classification

Text documents have rich information that can be useful for different tasks. How to utilise the rich information in texts effectively and efficiently for tasks such as text classification is still an active research topic. One approach is to weight the terms in a text document based on their relevance to the classification task at hand. Another approach is to utilise structural information in a text document to regularize learning so that the learned model is more accurate. An important question is, can we combine the two approaches to achieve better performance? This paper presents a novel method for utilising the rich information in texts. We use supervised term weighting, which utilises the class information in a set of pre-classified training documents, thus the resulting term weighting is class specific. We also use structured regularization, which incorporates structural information into the learning process. A graph is built for each class from the pre-classified training documents and structural information in the graphs is used to calculate the supervised term weights and to define the groups for structured regularization. Experimental results for six text classification tasks show the increase in text classification accuracy with the utilisation of structural information in text for both weighting and regularization. Using graph-based text representation for supervised term weighting and structured regularization can build a compact model with considerable improvement in the performance of text classification.