Modeling individual differences using Dirichlet processes

Danielle Navarro, Thomas L. Griffiths, Mark Steyvers, Michael LeePsyArXiv 2019 · DOI

TL;DR

This paper introduces a flexible Bayesian framework for modeling individual differences by assuming subjects can belong to an infinite number of latent groups. This "infinite groups model," powered by Dirichlet processes, allows for a richer representation of variation than traditional methods, automatically determining the number of groups needed and avoiding complex model selection.

What Problem Does This Paper Address?

Traditional cognitive modeling often averages data across subjects, masking important individual differences. While fitting models to each subject individually can capture these differences, it risks overfitting and doesn't leverage similarities between subjects. Existing "group models" assume a fixed, finite number of groups, requiring difficult model selection to determine the optimal number. This paper addresses the challenge of simultaneously modeling both similarities and differences across individuals in a flexible and computationally tractable way, without pre-specifying the number of underlying groups.

Methods

The core of the paper is the infinite groups model, a Bayesian hierarchical model that uses the Dirichlet process as a prior for the distribution of individual differences. This approach allows for a potentially infinite number of latent groups, with the number of observed groups growing naturally with the data. The model is built upon the idea of a Chinese Restaurant Process (CRP) for assigning individuals to groups and a stick-breaking process for defining the weights of these groups. This framework allows for flexible parameter distributions and avoids the need for explicit model selection for the number of groups. The authors also detail the mathematical derivations for applying this model to discrete data, including a Gibbs sampling scheme for inference.

Datasets

The paper demonstrates the infinite groups model on three real-world datasets:

  • Category learning data: From McKinley and Nosofsky (1995), involving 10 subjects learning to discriminate between probabilistic categories.
  • Publication habits data: Collected from the websites of psychology departments at six institutions, analyzing the publication lists of 125 academics.
  • Web browsing data: From MSNBC.com, tracking the page-viewing behavior of 1000 users across 17 categories.

Key Insights and Ideas

  • Infinite groups as a flexible prior: The Dirichlet process provides a principled way to model individual differences without pre-specifying the number of groups, allowing the data to reveal the underlying structure.
  • Automatic model order selection: The infinite groups model inherently handles the problem of determining the number of groups, which is a significant advantage over finite group models.
  • Bridging stochastic and group models: The infinite groups model offers a way to capture both the continuous variation within groups (like stochastic parameter models) and the distinct, multimodal differences between groups.
  • Psychological interpretability: The model's ability to identify latent groups and their associated characteristics (e.g., publication patterns, browsing behavior) provides rich insights into how individuals differ.

Key Results

The paper demonstrates the effectiveness of the infinite groups model across its applications:

  • Category Learning: The model successfully identified distinct groups of subjects with different learning strategies, aligning with qualitative observations of their performance.
  • Publication Habits: The analysis revealed between 12 and 17 likely groups of psychologists, with some groups showing clear interpretations in terms of sub-fields of psychology.
  • Web Browsing: The model identified approximately 40 groups among 1000 users, distinguishing between focused search behavior and more exploratory browsing patterns.

Limitations and Future Work

  • Continuous distributions: The authors acknowledge that while the infinite groups model can capture multimodal distributions, it might be too restrictive for phenomena requiring continuous, unimodal distributions within groups. Future work could involve combining the infinite groups model with continuous distributions.
  • Structured relationships between groups: The current model treats groups independently. Future research could explore models that allow for structured relationships between groups, such as hierarchical Dirichlet processes.
  • Idiosyncratic strategies: The paper suggests exploring models where individuals draw on a common set of strategies but combine them in unique ways.
  • Generalization of priors: The authors mention extending the framework to other stick-breaking priors and nonparametric priors for continuous distributions (e.g., Pólya trees).

Summary updated

Abstract

We introduce a Bayesian framework for modeling individual differences, in which subjects are assumed to belong to one of a potentially infinite number of groups. In this model, the groups observed in any particular data set are not viewed as a fixed set that fully explains the variation between individuals, but rather as representatives of a latent, arbitrarily rich structure. As more people are seen, and more details about the individual differences are revealed, the number of inferred groups is allowed to grow. We use the Dirichlet process—a distribution widely used in nonparametric Bayesian statistics—to define a prior for the model, allowing us to learn flexible parameter distributions without overfitting the data, or requiring the complex computations typically required for determining the dimensionality of a model. As an initial demonstration of the approach, we present three applications that analyze the individual differences in category learning, choice of publication outlets, and web browsing behavior.
BibTeX
@misc{navarro2019modeling,
  author = {Danielle Navarro and Thomas L. Griffiths and Mark Steyvers and Michael Lee},
  title = {{Modeling individual differences using Dirichlet processes}},
  year = {2019},
  doi = {10.31234/osf.io/bc3ya},
  howpublished = {PsyArXiv},
}
Figure