A generic multi-level framework for building term-weighting schemes in text classification

Abstract

Abstract Term weighting is essential for text classification tasks, and thus various supervised term-weighting (STW) methods have been designed and presented in recent years, such as TF (term frequency)-IG (information gain), TF-MI (mutual information), TF-RF (relevance frequency), and TF-IDF (inverse document frequency)-ICSDF (inverse class space density frequency). Unlike other schemes, TF-IDF-ICSDF considers not only the local factor (i.e. TF) and the category factor (i.e. ICSDF) but also the global factor (i.e. IDF) in the weighting process. Hence, a natural question is whether IDF is really useful for improving the classification performance of STW schemes. To explore this issue, a generic multi-level framework composed of term-level, text-level, and category-level is first established, which corresponds to local factor, global factor, and category factor, respectively. Based on the generic multi-level framework, a new two-level STW method, TF-ICSDF, can be generated by removing the IDF from the TF-IDF-ICSDF scheme. Conversely, we also integrated the IDF with other two-level STW schemes (e.g. TF-IG, TF-MI, TF-RF) to obtain several three-level STW schemes. We verified the general classification performance of our proposed STW schemes on three open benchmark datasets. The results manifest that performance can usually be boosted if IDF is incorporated into the STW schemes, indicating that weighting terms utilizing the IDF factor could provide better text representation. Therefore, the generic multi-level framework and STW schemes we proposed are effective.