Modeling Interdependent and Periodic Real-World Action Sequences.

Area of Science:

• Computational science
• Behavioral science
• Data science

Background:

• Mobile health apps collect extensive user data on activities like exercise, sleep, and diet.
• Accurate prediction of human actions is crucial for personalized health recommendations and app functionality.
• Existing models struggle with the complexity of human behavior, including its temporal dynamics, interdependencies, and periodicity.

Purpose of the Study:

• To develop a novel statistical model for predicting time-varying, interdependent, and periodic action sequences.
• To address limitations in previous approaches that focused on item consumption rather than broader behaviors.
• To enhance the accuracy of predicting future user actions and their timing in mobile health contexts.

Main Methods:

• Introduced TIPAS (Time-varying, Interdependent, and Periodic Action Sequences), a novel statistical model.
• Utilized personalized, multivariate temporal point processes with a mixture of Gaussian intensities.
• Incorporated Hawkes process-based self-excitations to capture short-term and long-term action interdependencies and periodicities.

Main Results:

• TIPAS significantly improved action and timing predictions over existing methods by up to 156% and 37%, respectively.
• Performance gains were substantial for rare and periodic actions like walking and biking, exceeding baselines by up to 256%.
• The model was evaluated on two large-scale datasets with 12 million actions from 20,000 users over 17 months.

Conclusions:

• Explicitly modeling dependencies and periodicities in real-world behavior enhances prediction accuracy.
• The TIPAS model offers improved capabilities for understanding and predicting human actions.
• Findings have implications for human behavior modeling, mobile app personalization, and targeted health interventions.