About Me
CV
Teaching
Research & Talks

Research & Talks

My research focuses on response-adaptive experimentation and personalization, with applications in education, physical activity, and mental health. I build bandit algorithms for finite-horizon and data-sparse settings, and develop inference methods for adaptively collected data.

Research Areas

  • Response-adaptive experimentation (contextual bandits; Thompson sampling variants)
  • Statistical inference for adaptively collected data (testing and uncertainty)
  • Personalization in real-world behavioral and educational platforms
  • LLM-assisted messaging systems within adaptive experimentation pipelines

Selected Research Projects

Adaptive Text Message-Based System for Mental Health Interventions (Project Leader; Apr 2023 – Present)

  • Led development of an adaptive text-based dialogue system with collaborators at Microsoft, Northwestern University, and Mental Health America
  • Supported treatment configuration and model management in the MOOCLet platform
  • Reached 5,500+ users in North America

Personalized Physical Activity Recommendations with Bandits and LLMs (Project Leader; Aug 2023 – Present)

  • Designed a personalized messaging system combining bandit algorithms with LLM-generated messages
  • Built survey + feedback pipelines to learn user preferences and perceived personalization
  • Focused on behavior change themes such as Behavioral Activation, Gratitude, Help Seeking, and Self-Compassion

WAPTS: Weighted Allocation Probability Adjusted Thompson Sampling (Principal Author; Sept 2023 – Present)

  • Developed WAPTS to improve personalization in finite-horizon, high-dimensional, and data-sparse adaptive experiments
  • Applied WAPTS to adaptive surveys and learner-sourcing contexts and evaluated performance empirically

Allocation Probability Test for Inference (Principal Author; Oct 2022 – Present)

  • Developed an allocation-probability-based hypothesis test for adaptively randomized experiments
  • Studied stability-controlled and continuous variants to improve performance and computation

Balancing Reward and Statistical Power in Adaptive Experiments (Co-author; Jul 2022 – Present)

  • Worked on algorithms that trade off reward and statistical goals in adaptive experimentation
  • Applied in messaging-based intervention settings

Publications & Working Papers (Selected)

  • Song, Haochen, Li, T., Franklin, M., Deliu, N., & Williams, J. (2024). Hypothesis Testing in Adaptively Randomized Experiments: Using the Allocation Probabilities for Inference. JSM 2024 Proceedings.
  • Musabirov, I., Reza, M., Song, H., et al. (2025). Platform-based Adaptive Experimental Research in Education: Lessons Learned from Digital Learning Challenge. Accepted to LAK 2025.
  • Song, Haochen, Musabirov, I., Bhattacharjee, A., et al. (2025). Adaptive Experiments Under High-Dimensional and Data Sparse Settings: Applications for Educational Platforms. arXiv:2501.03999
  • Song, Haochen*, Hofer, D.*, et al. (2025). Investigating Relationship Between Physical Activity Level with Personalized Behavior Change Message: Connecting Contextual Bandit with Large Language Models. KDD 2025 Workshop (Accepted).
  • Li, T., Nogas, J., Song, H., et al. (2021). Algorithms for adaptive experiments that balance statistical analysis and reward. arXiv:2112.08507

Talks & Conference Presentations (Selected)

  • CODE @ MIT (2024): WAPTS: A Weighted Allocation Probability Adjusted Thompson Sampling Algorithm for High-Dimensional and Sparse Experiment Settings.
  • JSM (2024): Hypothesis Testing in Adaptively Randomized Experiments: Using the Allocation Probabilities for Inference. (Theory and Foundations of Bayesian Statistics session)
  • Learning @ Scale Workshop (2024): WAPTS: An Allocation Probability Adjusted Thompson Sampling Algorithm for Learner Sourcing.
  • CODE @ MIT (2023): WAPTS: A Weighted Allocation Probability Adjusted Thompson Sampling Algorithm for High-Dimensional and Sparse Experiment Settings.