DEEP LEARNING-BASED SENTIMENT ANALYSIS OF AMAZON KINDLE STORE REVIEWS USING A HYBRID 3-CONVOLUTIONAL CNN AND 2-GRU LAYERS RNN MODEL

Abstract

Sentiment Analysis is a part of Natural Language Processing (NLP) where we try to train a machine in such a way that it generates the ability to define the overall opinion about certain context such as negative, neutral, or positive. Data is taken and various pre processing steps are applied and tagging of data is done to define its orientation then this minimalized data is converted into vector space as machine understand numbers not text using sentiment score or weightage to each word or frequency count methods. Then various machine learning algorithms are applied, and results are evaluated. In this work we have taken a different data set and has classified the reviews as positive and negative easily which can help company to see negative reviewed products without reading too much reviews. Focus of this work is negative rated products. Data is first preprocessed by removing the null values and feature is extracted using vectorizer. Model is trained with balanced data, 80% of the data is used to train the model and 20% data is used to test. Further the test accuracy is obtained using different classifiers. To capture the complex relationships among users and comments, we employ GNN to model the graph structure inherent in the dataset. By leveraging the graph representation, our approach achieves improved sentiment classification accuracy compared to traditional methods. The results demonstrate the effectiveness of GNN in capturing nuanced sentiment patterns within Weibo comments, offering valuable insights for understanding public sentiment on social media platforms. v For dealing with imbalanced data the Support Vector Machine (SVM) algorithm is combined with Particle Swarm Optimization (PSO) and other oversampling methods to create a hybrid strategy. The data was obtained with the assistance of Jeeran, a well-known Arabic assessment social network. To correct the feature weights, a PSO strategy is used, and four distinct oversampling methods—Synthetic Minority Oversampling Technique (SMOTE), SVM-SMOTE, Adaptive Synthetic Sampling (ADASYN), and borderline SMOTE—are utilized to correct the dataset's imbalance and produce an optimized dataset. In terms of accuracy, F-measure, G-mean, and area under the curve (AUC), the proposed PSO-SVM method performs better than previous classification algorithms for a variety of dataset versions. Further we tried deep learning models. In this particular CNN model, we utilize an embedding layer with a dimension of 64, followed by three convolutional layers. For the RNN model, we are opting for a simple architecture. It consists of an embedding layer, two layers of GRU, followed by two dense layers, and ultimately the output layer. To optimize performance, we employ the CuDNNGRU instead of the regular GRU, as it offers significantly faster computation, potentially over ten times faster. The maximum accuracy score achieved on the testing data set is 95.02% with the highest F1 score of 95.04% on negative review predictions. Highest accuracy is achieved by the hybrid model.