Fast categorisation of large document collections

As the volume of data stored online increases, careful management of large document collections becomes increasingly important. Categorisation is one important document management technique. It has been effectively employed in the Web, where links to documents are maintained in topic or interest areas in, for example, the manually-categorised Yahoo!(1) hierarchy. The drawback of manual categorisation is that it is practical only on small numbers of documents, it is not scalable, and relies on the subjective judgement of human assessors. Automatic categorisation has been shown to be an accurate alternative to manual categorisation. In automatic categorisation, documents are processed and automatically assigned to pre-defined categories that represent an interest or topic area. We propose and investigate heuristics for fast categorisation of large collections of documents that are focused on selecting a minimal set of representative features from uncategorised documents. We show that these new heuristics are accurate-in some cases more accurate than the baseline techniques-and also permit more than three fold reductions in processing time for categorising large collections.