The SVM With Uneven Margins and Chinese Document Categorization

The SVM With Uneven Margins And Chinese Document Categorisation Yaoyong Li� Department of Computer Science�John Shawe-Taylor Department of Computer Science Royal Holloway, University of London Royal Holloway, University of London Egham, Surrey, TW20 OEX, UK� yaoyong@cs.rhul.ac.uk�Egham, Surrey, TW20 OEX, UK john©cs.rhul.ac.uk Abstract We propose and study a new variant of the SVM — the SVM with uneven margins, tailored for document categorisation problems (i.e. problems where classes are highly unbalanced). Our experiments showed that the new algorithm significantly outper- formed the SVM with respect to the document categorisation for small categories. Furthermore, we report the results of the SVM as well as our new algorithm on the Reuters Chinese corpus for document categorisation, which we believe is the first result on this new Chinese corpus. 1 Introduction Document Categorisation (DC), the problem of assigning documents to predefined categories, is an active research area in information retrieval and machine learning. For one category, the DC problem is actually a binary classification problem by classifying a document to the category or not. Many machine learning algorithm have been applied to the DC problem, using a training set of categorised documents to obtain a classifier for one category and then judging the relevance of a document to the category by feeding the document into the classifier. The support vector machine (SVM) is a well known learning algorithm for linear classifier and has achieved state of the art results for many classification problem, including document categorisation (see Joachims (1998), Yang and Liu (1999)). However, it has been noticed the the performance of the SVM for small category (i.e. category with a small number of relevant documents in collection) was quite poor. Small category in DC problem corresponds to the clas- sification problem with uneven datasets, where the numbers of positive and negative examples are very different. Several kinds of adjustments have been made to the SVM to deal with the uneven datasets. The algorithm we present in this paper, the SVM with uneven margin, is the latest one. People who are concerned with Chinese information retrieval might notice that, whereas ex- tensive studies have been done to English document categorisation, relatively few results have been reported on Chinese document categorisation (He et al. (2003)). This is mainly because of a lack of Chinese document collection designed particularly for document categorisation. For- tunately, a multilingual corpus RCV2 created recently by Reuters contains a Chinese document collection. The Chinese corpus was manually categorised in the similar way and with the simi- lar high quality with the Reuters English corpora such as Reuters-21578 and the RCV1 corpus. So the Reuters Chinese collection is comparable to the two commonly used English document collections with respect to DC problem. Regarding big difference between English and Chinese, it is interesting to see whether the good methods such as the SVM for English DC problem are able to achieve similar performances for Chinese as for English. 216 In Section 2 we overview the previous works about adapting the SVM towards uneven …

Content truncated to 500 words