Bayesian Hierarchical Pointing Models

Document Title

Bayesian Hierarchical Pointing Models

Document Information

• Subject Area: Human-Computer Interaction, Bayesian Modeling, Movement Time Prediction
• Keywords: Fitts' Law, Bayesian Modeling, Hierarchical Models, Human-Computer Interaction, Movement Time Prediction, Information Criteria, Parameter Estimation, Data Sparsity, Weakly Informative Priors, Pointing Tasks

Research Background and Problem

• Identified Problems or Challenges: Traditional pointing models, such as those based on Fitts' Law, exhibit limitations in parameter estimation and predicting specific user behavior. When data is sparse or noisy, standard "fully pooled models" and "independent models" fail to effectively model and predict.

• Importance of the Research: Pointing models in Human-Computer Interaction (HCI) are crucial for interface design, optimization, and evaluation, aiding in the development of better user operation models. However, current models struggle with inter-user variability or unknown users.

• Motivation and Related Work: This study leverages the characteristics of Bayesian hierarchical models, including "partial pooling" and noise filtering, to provide a systematic solution for pointing tasks. While Bayesian hierarchical models have been widely applied in cognitive science, political science, and machine learning, their application in modeling interactive behavior remains underexplored. This study aims to address this research gap.

Solution

• Proposed Method or Solution: The authors propose a Bayesian hierarchical pointing model, extending Fitts' Law into a Bayesian framework. Movement time is modeled as a hierarchical structure with group and individual levels, and Bayesian methods are employed for inference.

• Innovations:

  • Extending Fitts' Law to model the distribution of movement time, rather than merely predicting the mean.
  • Introducing three Bayesian models (hierarchical, independent, and pooled) and systematically comparing their performance.
  • Utilizing "partial pooling" and inference capabilities to more accurately estimate model parameters and behavior distributions in sparse data scenarios.
  • Exploring the impact of different types of priors (uninformative, highly informative, weakly informative) on model prediction performance.

• Implementation Steps and Key Techniques:

  1. Model Definition: Using exGaussian distribution to model movement time and extending it to a hierarchical model structure.
  2. Parameter Estimation for Three Models:
     • Pooled Model: All users share a fixed parameter.
     • Independent Model: Each user has a unique set of parameters.
     • Hierarchical Model: Captures inter-user variability through group and individual-level parameters.
  3. Prior Specification:
     • Uninformative Prior: Using uniform distributions.
     • Highly Informative Prior: Based on mean and variance from previous studies.
     • Weakly Informative Prior: Combining prior study means with broader variance ranges.
  4. Model Implementation and Evaluation: Training and inference using PyStan, and evaluating model performance through information criteria (AIC, DIC, WAIC), correlation coefficient R², and root mean square error (RMSE).

Research Findings

• Specific Outcomes:

  • The Bayesian hierarchical model outperformed traditional pooled and independent models in predicting movement time distributions and mean values, especially in sparse training data scenarios.
  • Weakly informative priors showed slight performance improvements, while highly informative priors were limited when prior data mismatched the current task.

• Advantages:

  1. The "partial pooling" capability of hierarchical models utilizes group-level information to filter individual data noise.
  2. The model more accurately predicts unknown user data, addressing issues with pooled and independent models in handling outlier users.
  3. Suitable for environments with limited data, reducing the need for extensive user training data in interactive systems.

• Experimental or Evaluation Results:

  • Full data fitting tests showed that hierarchical and independent models had significantly better AIC, DIC, and WAIC values compared to the pooled model.
  • Leave-one-user-out cross-validation demonstrated that the hierarchical model predicted unknown users' movement time distributions more accurately than the pooled model.
  • Small data training tests validated the "partial pooling" capability of the hierarchical model, which provided better predictions than the other two models even when training data was limited to 5%-20%.

• Limitations and Future Directions:

  • The hierarchical model's performance is highly dependent on the quality of highly informative priors; mismatched prior data may lead to performance degradation.
  • Future research should extend hierarchical model construction to other tasks (e.g., guiding and crossing tasks).
  • Investigating effective ways to integrate multi-domain research results as weakly informative priors to further enhance model generalization.

This study demonstrates the application of Bayesian hierarchical models in pointing tasks, establishing them not only as an effective approach but also as a more systematic modeling practice.