Deep Learning for Indonesia Standard Industrial Classification
Mahendri Dwicahyo, Budi Yuniarto · 2020
Indonesia Standard Industrial Classification (Klasi-fikasi Baku Lapangan Usaha Indonesia or KBLI) is a classification reference developed by Statistics Indonesia (BPS) to classify any industry according to activity it carries out. Due to a large number of activities covered in KBLI, classification is usually performed by the experts using the industry activity description. Therefore, the time necessary to classify depends on the number of experts. That means as the data size grows, the process becomes much slower. One way to lessen the amount of knowledge needed is to recommend activities similar to the textual description provided. To realize that, we build a deep learning model of Gated Recurrent Unit, fastText word vectors, and label smoothing regularization (LSR) loss. We used 2016 BPS Economic Census data and BPS official KBLI guidebook as train and test data. Our results show model is able to recommend activities matching the input. Model capability to learn general traits of each activity can be seen from the similarity of the top recommendations. This study highlights the possibility of deep learning assisted classification, allowing the process to be carried out without much knowledge by leveraging existing data.