Abstract:In this paper we describe NudgeRank, an innovative digital algorithmic nudging system designed to foster positive health behaviors on a population-wide scale. Utilizing a novel combination of Graph Neural Networks augmented with an extensible Knowledge Graph, this Recommender System is operational in production, delivering personalized and context-aware nudges to over 1.1 million care recipients daily. This enterprise deployment marks one of the largest AI-driven health behavior change initiatives, accommodating diverse health conditions and wearable devices. Rigorous evaluation reveals statistically significant improvements in health outcomes, including a 6.17% increase in daily steps and 7.61% more exercise minutes. Moreover, user engagement and program enrollment surged, with a 13.1% open rate compared to baseline systems' 4%. Demonstrating scalability and reliability, NudgeRank operates efficiently on commodity compute resources while maintaining automation and observability standards essential for production systems.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to use the recommender system (Recommender System, RecSys) to promote healthy behavior change in a large - scale population. Specifically, the paper introduces NudgeRank™, a digital algorithm - based recommender system based on graph neural networks (Graph Neural Network, GNN) and knowledge graphs (Knowledge Graph), aiming to encourage positive healthy behaviors through personalized and context - aware nudges. The following are the specific problems that the paper attempts to solve:
1. **Health behavior change in a large - scale population**:
- The paper explores how to effectively promote health behavior change in a large - scale population through personalized and context - aware nudges. Traditional health intervention methods often lack personalization and context - awareness, and it is difficult to produce continuous effects in a wide range of populations.
2. **Generation of personalized and context - aware nudges**:
- How to design a recommender system that can generate personalized and context - aware nudges, which need to be customized according to users' personal characteristics, health status, and behaviors, in order to improve user participation and the effectiveness of behavior change.
3. **Technical challenges**:
- Implementing a personalized health - nudge recommender system that can operate in a large - scale population requires solving multiple technical challenges, including the system's scalability, performance, reliability, and compliance. The system needs to be able to generate personalized recommendations for millions of users in real - time or on a regular basis, and be able to handle the cold - start problem of new users and new health goals.
4. **Evaluation and optimization**:
- How to evaluate and optimize the effectiveness of the recommender system, including user participation, the effect of health behavior change, and overall system performance. Through actual deployment and evaluation, the paper shows the significant effect of NudgeRank™ in improving user participation and health behavior change.
### Solutions to specific problems
1. **Combination of graph neural networks (GNN) and knowledge graphs (Knowledge Graph)**:
- The paper proposes a recommender system framework based on GNN and knowledge graphs. GNN is used to capture the complex relationships between users, nudges, and their attributes, while the knowledge graph is used to store and update relevant information about users and nudges, thereby achieving highly personalized recommendations.
2. **Dynamic knowledge graph**:
- The knowledge graph is updated dynamically. As new user data is collected, the nodes and edges in the graph are continuously updated to ensure that the recommendations always reflect the user's current state and behavior.
3. **Solution to the cold - start problem**:
- By adding new users and new nudges to the knowledge graph and randomly initializing node embeddings using the Xavier initialization method, the cold - start problem of new users and new nudges is solved. In addition, a diversity mechanism is introduced. By randomly sampling some candidate nudges, the chance of new nudges being recommended is increased, thereby collecting more user feedback data.
4. **Business rules and constraints**:
- Multiple business rules and constraints are configured in the system, such as negative - feedback filters and nudge - budget filters, to ensure the positivity of the user experience and avoid nudge fatigue.
5. **Performance and resilience**:
- The system is built based on Kubernetes, adopts a distributed architecture, and improves the system's performance and resilience through parallel processing and an automatic retry mechanism, ensuring stable operation in a large - scale user group.
Through these methods, NudgeRank™ has been successfully deployed in the Healthy 365 mobile application of the Health Promotion Board (HPB) in Singapore, sending personalized health nudges to more than 1.1 million active users every day, significantly improving user participation and the effect of health behavior change.