Product Helpfulness Detection With Novel Transformer Based BERT Embedding and Class Probability Features

Abstract

Nowadays global market products are readily accessible worldwide, and a vast array of reviews across numerous platforms are posted daily in several categories, making it challenging for customers to stay informed about their product interests. To make informed decisions regarding product quality, users require access to reviews and ratings. Owners and managers must analyze customer ratings and the underlying emotional content of reviews to enhance the product’s quality, cost, customer service, and environmental impact. The primary aim of our proposed research is to accurately predict product helpfulness through customer reviews using the Large Language Model (LLM), thereby assisting customers in saving time and money. We employed a benchmark dataset, the Amazon Fine Food Reviews, to develop numerous advanced machine-learning techniques. We introduced a novel transformer approach BERF (BERT Random Forest) for feature engineering to enhance the value of user evaluations for Amazon’s gourmet food products. The BERF method utilizes BERT embeddings and class probability features derived from product helpfulness online reviews textual data. We have balanced the dataset using the Synthetic Minority Over-sampling TEchnique (SMOTE) approach. Our comprehensive study results demonstrated that the Light Gradient Boosting Machine (LGBM) strategy outperformed existing state-of-the-art approaches, achieving an accuracy of 98%. The performance of each method is confirmed using a k-fold method and further improved through hyperparameter optimization. Our innovative study employing a transformer model has significantly enhanced the utility of customer reviews, substantially reducing online product scams and preventing wasted time and money.

Keywords

• Product helpfulness
• large language model (LLM)
• machine learning
• deep learning
• text mining
• BERT