A Novel Mixture Model for Characterizing Human Aiming Performance Data

Yanxi Li, Derek S. Young, Julien Gori, Olivier Rioul

TL;DR

This paper introduces a new "flare regression model" that uses a mixture of Gaussian and exponential distributions to better capture the variability in human aiming performance data, especially when users aren't strictly optimizing for speed. They also developed an efficient ECM algorithm for fitting this model, which outperforms existing methods in simulations and real-world data analysis.

What Problem Does This Paper Address?

Fitts' Law is a well-established model for predicting human movement time, particularly in human-computer interaction. It typically assumes a Gaussian distribution for errors, which works well for controlled experiments where users are instructed to be as fast and precise as possible. However, when analyzing data from "in the wild" scenarios (like everyday computer use), this assumption breaks down. Real-world data often shows a positive skew, meaning movement times are frequently longer than predicted by a simple Gaussian model. This skew arises because users aren't always trying to minimize their task completion time; they might be distracted, interrupted, or simply not as focused. While the Exponentially Modified Gaussian (EMG) model can handle some of this skew, it doesn't fully capture situations where users are clearly not trying to be fast. The authors aim to develop a model that can identify and characterize these non-optimal performance regions, going beyond just fitting a skewed distribution.

Methods

The core of the paper is the introduction of a "flare regression model." This model assumes that the errors in predicting movement time come from a mixture of two distributions:

  1. A Gaussian component: Represents typical, efficient movements.
  2. An Exponential component: Represents movements where the user is not optimizing for speed, or potentially includes outliers.

This mixture structure allows the model to be more flexible and to identify distinct behavioral patterns. To fit this novel model, the authors developed an Expectation-Conditional-Maximization (ECM) algorithm. This algorithm is an iterative process that refines parameter estimates. A key advantage highlighted is the guaranteed convergence of their ECM algorithm to a local maximum of the likelihood function, which is not always the case for other methods used for similar models. They also developed a block-relaxation algorithm for the existing EMG regression model, noting its computational challenges and lack of guaranteed convergence.

Datasets

The primary dataset used for analysis is a "real-world" dataset of human aiming performance collected from 24 users over several months, involving over 2 million movements. This dataset was originally produced by Chapuis et al. (2007). The authors also generated synthetic datasets for extensive simulation studies to evaluate the performance and robustness of their proposed flare regression model and ECM algorithm.

Key Insights and Ideas

  • "In the wild" data is fundamentally different: The assumption of Gaussian errors, common in controlled studies, is often inappropriate for real-world human-computer interaction data.
  • Not all slow movements are the same: The exponential component in their mixture model helps distinguish between typical, fast movements and slower, less optimized ones, offering more insight than a single skewed distribution.
  • Model-based clustering is powerful: The flare regression model, through its mixture components, naturally lends itself to clustering observations into "Gaussian" (efficient) and "exponential" (less efficient/outlier) groups, which can help identify problematic data points or distinct user behaviors.
  • ECM algorithm offers computational advantages: The developed ECM algorithm for the flare model is more robust and guarantees convergence, unlike some existing methods for similar models.

Key Results

  • Superior Performance: The flare regression model consistently outperformed the EMG regression model and other candidate models (classic linear regression, two-component mixture of linear regressions) in terms of lower Bayesian Information Criterion (BIC) values across various data truncation thresholds and simulation settings.
  • Effective Clustering: When applied to the real-world aiming data, the flare regression model successfully classified observations into Gaussian and exponential components. For a sample of 200 generated observations, the model correctly identified 68 "exponential" observations and misclassified only 6 "Gaussian" observations as exponential.
  • Computational Efficiency: The ECM algorithm for the flare regression model was significantly faster than the block-relaxation algorithm for the EMG regression model, especially for larger datasets. Mean running time for ECM was ~8.85 seconds, compared to ~193.7 seconds for block-relaxation.
  • Parameter Estimation: The flare regression model produced parameter estimates that were generally more accurate and within typical ranges for controlled studies, especially for the slope parameter, which is considered more informative for movement difficulty.

Limitations and Future Work

  • Model Identifiability: The authors note that, like other mixture models, their flare regression model can suffer from identifiability issues, particularly when the mixture components are not well-separated or when the proportion of data from one component is very small. This was observed in simulation setting M12.
  • Outlier Handling: While the flare model can help identify potential outliers, the paper uses a simple truncation approach (removing data above a certain time threshold) to handle extreme values. More sophisticated outlier detection and removal methods could be explored.
  • Subject Heterogeneity: The current model treats all data points similarly. The authors suggest that incorporating random effects to account for subject-to-subject variability in performance could lead to even more informative models.
  • Alternative Distributions: While the exponential component was chosen for its relevance to non-optimal performance, other skewed distributions could be explored for the second component of the mixture model.

Summary updated

Abstract

Fitts' law is often employed as a predictive model for human movement, especially in the field of human-computer interaction. Models with an assumed Gaussian error structure are usually adequate when applied to data collected from controlled studies. However, observational data (often referred to as data gathered "in the wild") typically display noticeable positive skewness relative to a mean trend as users do not routinely try to minimize their task completion time. As such, the exponentially-modified Gaussian (EMG) regression model has been applied to aimed movements data. However, it is also of interest to reasonably characterize those regions where a user likely was not trying to minimize their task completion time. In this paper, we propose a novel model with a two-component mixture structure -- one Gaussian and one exponential -- on the errors to identify such a region. An expectation-conditional-maximization (ECM) algorithm is developed for estimation of such a model and some properties of the algorithm are established. The efficacy of the proposed model, as well as its ability to inform model-based clustering, are addressed in this work through extensive simulations and an insightful analysis of a human aiming performance study.
BibTeX
@inproceedings{li2022novel,
  author = {Yanxi Li and Derek S. Young and Julien Gori and Olivier Rioul},
  title = {{A Novel Mixture Model for Characterizing Human Aiming Performance Data}},
  booktitle = {Statistical},
  year = {2022},
}
Figure