A novel approach in the combination of Adam optimizer with cyclic learning rate for text classification by using convolution neural network Available to Purchase

Data Explosion is a more prevalent problem across the world because of digitalization. It is estimated that the volume of data may expand to 175 zettabytes by 2025. Text data are growing exponentially in the form of web pages, emails, tweets, etc. on the web. Accommodating, accessing, manipulating, analyzing, retrieving, and reporting text data is hard. Text mining techniques make it quite compatible to perform the tasks. The objective of text mining includes collecting the text from multiple heterogeneous sources, cleansing the text data, and preparing it into a suitable format for analyzing it. It discovers the hidden patterns in the texts and extracts knowledge from them. Text classification is one such technique that predicts the class labels of the unseen data based on learning the data during training phase. The proposed system employs glove; a word embedding model for representing the features into vectors and performing the classification of text by using a deep convolution neural network. It also explores optimal hyper-parameters viz. learning rate, epochs, batch size, loss function, etc. during back propagation of the deep Convolution Neural Network to minimize the error in predicting the class labels of unseen data. We employ three different datasets viz. Fake News detection, IMDB movie review corpus, and Depression Detection dataset for experimental analysis. It uses the cyclic learning rate in combination with the Adam optimizer; a method to yield the optimal range of the learning rate. It provides competitive results regarding precision, recall, accuracy, and F-measure in text classification. The adaptive learning method namely Adam with a cyclic learning rate outperforms when compared to the classical Stochastic Gradient Descent method and Adam’s constant learning rate method.