AppGuider: Feature Comparison System using Neural Network with FastText and Aspect-based Sentiment Analysis on Play Store User Reviews

Abstract

Nowadays, there's a rapid growth in the number of apps downloaded from the app stores. People nowadays use apps for even the most simple daily tasks. In this situation, people always tend to search for new apps for the new tasks they come across in daily life. User reviews have a high impact on the app downloads. When analysing user reviews, it's important to consider the aspect that has been discussed in reviews. In mobile app reviews, the discussed aspect is mostly a functionality or feature of the mobile app. Therefore, it's crucial to make use of this important data in a way that helps app seekers to easily find the best-suited app for their requirements and also helps app developers to identify their weak features that need to be improved. This research was conducted to provide a strategy that visualizes user review summaries in a form that is relevant to the end user with the intention of achieving a model that is not only lightweight but also highly accurate and effective in terms of its performance. The AppGuider system was implemented, mainly with two models for sentiment analysis and aspect extraction. The sentiment classification model was developed with a deep learning approach that included a two-layer neural network, while the aspect extraction model was built with an unsupervised machine learning approach using the LdaMulticore algorithm. FastApi was used for data visualization in Frontend. User reviews were vectorized with FastText prior to input into the model. The accuracy of the sentiment classification model is 91%, with an 85.97% f1 score, an 85.93% recall, and an 86.05% precision. The FastText model outperformed the Stanford CoreNLP library in the performance test. The integrated system was evaluated by 25 user reviews that were entered manually and sentiment classification model scored 92% while the aspect extraction model scored of 76% accuracy.