Text Classification using Improved IWO-HAN

Abstract

In this paper, a novel Improved Invasive Weed Optimization-based Hierarchical Attention Network (IIWO-HAN) is developed to achieve text classification. As for huge datasets, automatic labelling is required to get useful insights. Thus, the text classification becomes very important. IIWO has the quality of randomness and imitating compatibility of weeds colony and HAN makes use of levelled document structure. Hence, these two are integrated together to achieve text classification. Then, the proposed method has been analysed based on five popular parameters namely, Accuracy, Precision, TPR, TNR and FNR. For this purpose, three datasets: Reuters dataset, 20-Newsgroup dataset and Self-created dataset have been utilized where Reuters dataset, 20- Newsgroup dataset are standard datasets and the Self-created dataset consists of 5000 documents comprising of abstracts taken from various reputed journals. Further, the proposed methodology has been compared with an existing Improved Sine Cosine Algorithm (ISCA). It is found that IIWO-HAN results in 81.3%, 86.5%, 84.1%, 89.4% and 12.4% for TPR, TNR, Accuracy, Precision and FNR respectively and achieves better performance as compared to existing ISCA which gives 78.1%, 79.4%, 79.4%, 75.6%, 17.8% as TPR, TNR, Accuracy, Precision and FNR respectively for Reuters dataset. For 20- Newsgroup dataset, IIWO- HAN provides the TPR, TNR, Accuracy, Precision and FNR of 87.5%, 88.4%, 87.1%, 91.4% and 12.4% respectively and 84.1%, 86.5%, 89.4%, 92.4% and 15.8% as TPR, TNR, Accuracy, Precision and FNR respectively for Self-created dataset. It is found that the Precision of 92.47% and Accuracy of 89.4% is achieved using self-created dataset at training data value of 90% which is clearly better than Precision of 77.6% and Accuracy of 82.3% for existing ISCA.