Cluster-Based Predictive Modeling of User Ratings for Physical Activity Apps Using Mobile App Rating Scale (MARS)...

Area of Science:

• Mobile Health
• Machine Learning
• Data Science

Background:

• Mobile health apps require tools to evaluate quality before deployment.
• The Mobile App Rating Scale (MARS) assesses apps but has limited predictive power for user satisfaction.
• Predictive modeling for app quality is an emerging need.

Purpose of the Study:

• To predict user ratings for physical activity apps using MARS dimensions and machine learning.
• To forecast app ratings before production using k-means clustering and ML models.
• To identify key drivers of user satisfaction in mobile health apps.

Main Methods:

• Analyzed 155 MARS-rated physical activity apps, splitting data into training and testing sets.
• Applied k-means clustering to identify app clusters, followed by training 5 ML models.
• Evaluated model performance using accuracy, mean absolute error, and R², with validation on external datasets.

Main Results:

• Clustering revealed distinct app types: feature-rich (cluster 1) and simpler (cluster 2).
• A combined model (support vector regression + k-nearest neighbors) achieved 88.64% accuracy, outperforming unclustered models.
• Clustering improved prediction accuracy and generalization to external datasets.

Conclusions:

• The combined clustering and modeling approach enhances prediction accuracy for app user ratings.
• This method transforms MARS into a predictive tool, aiding app development and quality assessment.
• The approach offers a scalable and transparent method for forecasting user ratings, especially in early development stages.