Collective-Intelligence Recommender Systems: Advancing Computer Tailoring for Health Behavior Change Into the 21st Century

Rajani S. Sadasivam, Sarah L. Cutrona, Rebecca L. Kinney, Benjamin M. Marlin, Kathleen M. Mazor, Stephenie C. Lemon, Thomas K. HoustonJournal of Medical Internet Research 2016 · DOI

TL;DR

This paper explores how "collective-intelligence recommender systems," similar to those used by Netflix and Amazon, can revolutionize computer-tailored health communication (CTHC). By moving beyond pre-programmed rules to data-driven personalization, these systems promise more effective health behavior change interventions, though significant challenges remain in their implementation.

What Problem Does This Paper Address?

Traditional Computer-Tailored Health Communication (CTHC) systems rely on health behavior theories and expert-defined rules to personalize messages for individuals. While effective, this approach has limitations: it's resource-intensive to create complex rule sets, can struggle to capture the full nuance of individual behavior, and adapts slowly to real-time changes. The authors argue that we may be reaching the natural limits of this rule-based approach and that a new frontier is needed to advance CTHC into the 21st century. They propose exploring collective-intelligence recommender systems as a solution.

Methods

The authors conducted a focused literature review and interviewed eight subject matter experts. The literature review focused on identifying limitations of current CTHC, advantages of recommender systems, and challenges in their implementation for health promotion. Expert interviews aimed to gather insights on these same themes, using open-ended questions to probe their experiences and perspectives. The findings from both sources were synthesized to form the basis of the paper's conclusions and proposed research agenda.

Datasets

Not discussed in the paper. The paper focuses on the conceptual framework and challenges of applying recommender systems to CTHC, rather than presenting empirical results from a specific dataset.

Key Insights and Ideas

  • The "Natural Limits" of Rule-Based CTHC: The paper highlights that while theory-driven CTHC has proven effective, the complexity of creating and managing rule-based systems for highly personalized interventions becomes a significant bottleneck. This limits the number of tailoring variables that can be practically incorporated.
  • Recommender Systems as a Paradigm Shift: The core idea is to leverage machine learning algorithms, already successful in commercial applications, to dynamically tailor health messages. Instead of pre-defined "if-then" rules, these systems learn from user interactions (implicit and explicit feedback) to personalize content.
  • Beyond Designer Intent: Recommender systems shift the focus from what study designers predict users will respond to, to what users actually respond to based on their observed behavior. This allows for more nuanced and potentially more effective tailoring.
  • Hybrid Approaches as the Sweet Spot: The paper suggests that hybrid recommender systems, which combine theory-based approaches with empirical data-driven methods, hold the most promise for advancing CTHC. This allows for leveraging existing health behavior theories while benefiting from the adaptive power of machine learning.
  • Significant Implementation Hurdles: Despite the potential, the authors identify critical challenges, including the "cold start" problem (lack of initial data), steep intervention attrition rates, and the potential for unintended consequences where tailored messages might inadvertently reinforce undesirable behaviors.

Key Results

The paper does not present quantitative results from a specific study. Instead, it synthesizes findings from a literature review and expert interviews to:

  • Identify limitations of current CTHC: These include the difficulty of implementing high-level tailoring due to rule complexity, potential limitations in individual relevance due to theoretical constraints, and slow adaptation to real-time user behavior.
  • Outline advantages of recommender systems: These include the ability to tailor based on observed behavior, adapt to a near-infinite number of variables, and learn and adapt in real-time.
  • Detail challenges in incorporating recommender systems: Key challenges include the limited availability of collective-intelligence data at the start of interventions, the impact of small sample sizes and short study timelines, steep intervention attrition rates, and potential unintended consequences of data-driven tailoring.

Limitations and Future Work

Limitations:

  • The research on incorporating recommender systems into CTHC is in its early stages, with limited published work.
  • The paper acknowledges that its views are limited by the current state of research and aims to initiate a conversation.

Future Work:

The authors propose a research agenda focused on:

  • Understanding Collective-Intelligence Data: Research is needed to determine what constitutes appropriate feedback for health messages, how to collect it effectively (e.g., single vs. multiple dimensions of feedback), and how to overcome the "cold start" problem by leveraging existing data or social networks.
  • Selecting the Appropriate Recommender Approach: Studies are needed to compare the performance of content-based, collaborative filtering, and hybrid recommender systems in CTHC interventions, considering factors like cost, engagement, and behavior change.
  • Understanding the Impact of Collective-Intelligence Data: Comparative effectiveness studies are required to evaluate the benefits of recommender systems over traditional CTHC in terms of engagement and behavior change.
  • Addressing Interdisciplinary Gaps: The authors emphasize the need for changes in training and funding models to foster interdisciplinary research between behavioral science and computing, promoting better communication and understanding of challenges across fields.

Summary updated

Abstract

We promote discussion of ways to move CTHC into the 21st century by incorporation of recommender systems.
BibTeX
@article{sadasivam2016collective,
  author = {Rajani S. Sadasivam and Sarah L. Cutrona and Rebecca L. Kinney and Benjamin M. Marlin and Kathleen M. Mazor and Stephenie C. Lemon and Thomas K. Houston},
  title = {{Collective-Intelligence Recommender Systems: Advancing Computer Tailoring for Health Behavior Change Into the 21st Century}},
  journal = {Journal of Medical Internet Research},
  year = {2016},
  doi = {10.2196/jmir.4448},
}
Figure