A Framework for Collaborative Document Classification with GA-SVM
Snehashish Chakraverty, Ujwal Pandey, P. Vijaya Bhargava Phani Dutt · Zenodo (CERN European Organization for Nuclear Research) · 2016
Text Classification has been addressed by purely statistical approaches that utilize the frequency of occurrence of significant terms as well as by tapping a range of semantic features conveyed by the text. Both approaches have proved their strengths, yet each has its own limitations when applied to corpuses with different sizes and expressive styles. This raises two interesting problems- given a corpus, how to automate the process of (i) finding an optimum blend of statistical and contextual contributions for the most appropriate classification, and (ii) determining the relative importance of different kinds of contextual features that are employed? In this paper, we address these issues by developing a Collaborative Document Classification (CDC) system that adapts according to a given corpus, the weighted contributions of statistical features, an array of lexical-semantic features derived from the WordNet ontology and categorical-semantic features obtained from the hierarchical organization of Wikipedia category pages. Given the complexity of this multivariate problem, it is judicious to seek approximate solutions using metaheuristics. We employ a GA that embeds a multi-class SVM classifier into its fitness function evaluator to cull out an optimal mix of statistical and semantic features as tailored to a given corpus. We experimented on small as well as large data sets derived from three sources: the 20 Newsgroup corpus, the Reuters 21578 corpus and a Creative corpus that we handcrafted by collecting news articles from the Times of India news portal. Results indicate that the DC system was able to balance between statistical and context approaches and also beefed up the contributions of the most relevant semantic features for each corpus to achieve a high classification accuracy ranging from 88% to 100% with an average of 95.55%. The results highlight the significance of a collaborative DC approach that taps the power of ontological databases and can adapt to varying corpora seamlessly. The final population output by the GA contains a set of non-inferior solutions that give trade-off possibilities between recall and precision.