Performance evaluation of machine learning models on large dataset of android applications reviews

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Science

Background:

• Mobile applications are increasingly popular, generating vast amounts of user reviews.
• Analyzing these reviews is crucial for app improvement and development but presents challenges due to their volume and complexity.
• Categorizing reviews aids users in selecting appropriate applications.

Purpose of the Study:

• To propose a framework for sentiment analysis of mobile app reviews across eight diverse categories.
• To evaluate the effectiveness of machine learning models combined with TF-IDF for feature extraction in app review analysis.

Main Methods:

• Scrapped a dataset of 251,661 user reviews using Regular Expression and Beautiful Soup.
• Employed Term Frequency-Inverse Document Frequency (TF-IDF) for feature extraction.
• Utilized various machine learning models, including Support Vector Machine, and evaluated performance with preprocessing and statistical features.

Main Results:

• Combining statistical features with TF-IDF significantly improved model performance.
• The Support Vector Machine model demonstrated the highest accuracy in sentiment analysis.
• The study provides a large, diverse, and balanced dataset for further research.

Conclusions:

• The proposed framework effectively analyzes sentiment in mobile app reviews.
• The combination of TF-IDF and statistical features, along with SVM, offers a robust approach for sentiment analysis.
• The findings can guide researchers in selecting optimal models for app review analysis and contribute to the development of better mobile applications.