Intent-IQ: customer’s reviews intent recognition using random forest algorithm

Customer review is a written evaluation provided by the consumer. The evaluation will cover the product, services provided, delivery service and experience. Customer intent is the root cause or purpose that drives a customer's behaviors or actions when they interact with the business, website, or product. In this research, we are focusing on customer intent through their reviews. The main problem is with the large-scale manufacturer relies on consumer reviews, ratings, and opinions regarding the quality and design of the products and the manual analysis of these ratings is time-consuming, hence diminishing effectiveness. The problem of information overload in online review platforms has substantially affected many buyers' abilities to properly evaluate the quality of the item or organizations while making decisions about purchases. In order to overcome this problem, a classification model for intent recognition is developed. Dataset from Kaggle which contains English reviews from Shopee is downloaded to be used for the modelling process. Data annotation using semi-supervised learning which is self-training technique need to be accomplished as the dataset is unlabeled. Two machine learning model is chosen to build the classification models which are Random Forest (RF) algorithm and Multinomial Naïve Bayes (MNB) algorithm. Intent-IQ is a web application system which allows users to input Shopee product link and it leads to the intent classification, where the reviews can be classified into its intent categories such as praise, complaint and suggestion. As for the result, the dataset that has gone through data annotation using self-training technique with SVM model is used for further analysis as it achieved 90.0% accuracy and F1-score. The RF classification model achieved accuracy of 81.94%, while MNB classification model with 71.38%. Therefore, the RF classification model is chosen to be integrated into the Intent-IQ system. The validation and functionality testing conducted on the system reflects that the system works as expected. The usability test achieved 97.3% SUS score which falls in the range of excellent ratings. The future recommendation for this project is to build a classification model for Malay reviews and explore other algorithms which could gain better accuracy by capturing more contextual meanings in the reviews such as BERT.