ANALYSIS OF E-COMMERCE PRODUCT REVIEW FOR TRUST ASSESSMENT USING ENSEMBLED DEEP LEARNING MODEL

In the realm of e-commerce, online reviews plays a vital role in influencing consumer decisions.. Especially there is huge demand in computing the trust of the online reviews to prevent the consumer from misguiding and major impacts of the products of the similar businesses. However many researcher has initiated their trust assessment approaches in online product review using machine learning and deep learning based Artificial intelligence architecture. Despite of several benefits, those approaches faces major challenges in terms of the processing the linguistic features and semantic features. In this paper, a new ensemble deep learning model with sentiment analysis has been designed to evaluate trust in product reviews from the Flipkart platform on providing consumer with reliable deeper insights of the product quality and user experiences. Initially preprocessing of the dataset is carried out to obtain reliable word vector from product reviews containing text data through stop word removal , special character removal , stemming and tokenization processes. Next, N-gram model is used to extract the contextual and behaviour features through linguistic and semantic approaches on token of the word vector. Obtained contextual feature and behaviour word vector is encoded using transformer model. Encoded Word vector is projected to BERT analysis for sentiment extraction. Encoded word vector with sentiment is projected to ensemble deep learning models termed as EDL-RNN+CNN+BIGRU. It composed of Recurrent Neural Networks (RNN), Improved Pooling based Convolutional Neural Networks (IPCNN) and Bidirectional Gated Recurrent Units (BiGRU) model. Finally the complementary strengths of each model is leveraged as it is more robust and effective in capturing the nuanced features of online reviews for sentiment analysis compared to using any single architecture alone. Weighted averaging is a common approach employed to ensemble learning to interface the predictions of individual architectures while assigning different weights to each model's prediction based on its performance on computing the trust on the processing sentiment incorporated encoded feature vector in the layer of the model. Experimental and performance analysis of the proposed ensemble model is carried out using Flipkart dataset which is extracted using the online platform such as kaggle. Performance analysis of the model using cross fold validation of test data is performed to compute the accuracy and effectiveness of the model . On evaluation, model produces 94.2% of accuracy, 94% of precision and 93% of recall which found to be outperforming on comparing it against state of art approaches.