Contradiction in text review and apps rating: prediction using textual features and transfer learning

Area of Science:

• Natural Language Processing
• Machine Learning
• Software Engineering

Background:

• Mobile app stores like Google Play are key platforms for app distribution and user feedback.
• User reviews offer valuable insights but can be subject to bias, creating discrepancies with numerical ratings.
• Existing methods struggle to reconcile numerical ratings with the rich information in user reviews.

Purpose of the Study:

• To develop a transfer learning model for predicting authentic numerical ratings of mobile apps based on user reviews.
• To address the issue of biased numerical ratings by leveraging the content of user-provided text reviews.
• To evaluate the effectiveness of the proposed ELMo model against other transfer learning and machine learning approaches.

Main Methods:

• A transfer learning approach using the ELMo model, based on word vector representations, was employed.
• A dataset of app reviews and ratings was collected from the Google Play Store across 14 app categories.
• TextBlob analysis was used to segregate biased and unbiased user ratings to establish ground truth for evaluation.

Main Results:

• The ELMo model demonstrated a high potential for predicting authentic numerical ratings by analyzing user reviews.
• The proposed ELMo model outperformed three other transfer learning and five machine learning models in accuracy.
• The study successfully identified and addressed discrepancies between numerical ratings and user-reported experiences.

Conclusions:

• Transfer learning, specifically using the ELMo model, is effective in predicting accurate mobile app ratings from user reviews.
• Analyzing user reviews provides a more authentic measure of app quality than numerical ratings alone.
• This approach offers a valuable tool for developers and users to better understand app performance and user satisfaction.