Using boosting mechanism to refine the threshold of VSM-based similarity in text classification

The vector space model (VSM)-based similarity classifier is the simplest text categorization method. It has a high classification speed, but with low accuracy. The main reason is that the similarity threshold used by the similarity classifier is decided empirically, but not mathematically. This paper introduces a boosting-based mechanism to adaptively compute out relatively accurate similarity threshold over specific dataset. This method constructs better similarity-based classification rules by combining the similarity thresholds generated by the constituent classifiers of boosting. It greedily minimizes the error rates on training documents; therefore the similarity classifier with thus computed threshold should also have low error rates.