Combining causal and similarity-based reasoning

Charles Kemp, Patrick Shafto, Allison Berke, Joshua B. TenenbaumNeurIPS 2006 · OpenReview

Introduction

When we encounter a new object, we often try to predict its unobserved properties. If you learn a specific duck has "Gene X," how likely is it that a goose has "Enzyme Y"?

Traditionally, cognitive scientists have modeled this inductive reasoning in two separate ways:

  1. Causal Reasoning: Focuses on the relationships between properties (e.g., Gene X causes Enzyme Y). These models often assume that different objects (rows in a data matrix) are independent.
  2. Similarity-Based Reasoning: Focuses on the relationships between objects (e.g., ducks are similar to geese). These models often assume that different properties (columns in a data matrix) are independent.

The gap in these traditional approaches is that human reasoning is rarely so compartmentalized. We know that identical twins (high similarity) are more likely to share a disease (causal outcome) than two strangers, even if both smoke. Existing models struggle to combine these two sources of information, often failing when an inference requires navigating both a causal chain and a taxonomic tree simultaneously.

This paper introduces a Bayesian framework that unifies these two pillars into a single generative model, successfully predicting human inferences about biological properties that neither causal nor similarity-based models can explain in isolation.

The Problem: Bridging Two Paradigms

To understand the challenge, imagine a matrix where rows are objects (like animals) and columns are properties (like genes or enzymes).

Visualizing the integration of causal and similarity-based reasoning to predict missing object properties.

Standard causal models can predict missing values in a row based on other values in that same row, but they ignore the relationships between rows. Conversely, similarity models predict missing values in a column based on other values in that same column, but they ignore how columns interact. The authors' goal is to build a model that can "fill in the blanks" by leveraging both the vertical (similarity) and horizontal (causal) structures simultaneously.

Methods: The Combined Model

The authors propose a Combined Model represented as a hierarchical Bayesian framework.

  • Object Structure ($S_o$): They represent object similarity using a tree-structured graphical model, similar to a phylogenetic tree. Properties are assumed to "mutate" or spread across this tree.
  • Feature Structure ($S_f$): They represent causal relationships between properties using a directed graphical model (a causal Bayes net).
  • The Integration: To merge them, the researchers convert the probabilistic causal model into a deterministic model ($S_f^D$) by introducing hidden variables. For example, a "transmission" variable determines if a cause actually triggers an effect.
  • The Generative Process: The core innovation is assuming that the "root" variables of the causal network (the primary causes and the hidden transmission mechanisms) are themselves distributed across the objects according to the similarity tree ($S_o$).

This allows the model to perform inference over the combined graph, effectively using similarity to estimate the likelihood of shared, unobserved causal mechanisms.

Experiments and Key Insights

The researchers tested their model against behavioral data from controlled experiments with adult participants. Participants were trained on pseudo-biological causal chains (e.g., Enzyme A $\rightarrow$ Enzyme B $\rightarrow$ Enzyme C) and asked to make predictions about four animals (mouse, rat, squirrel, sheep).

A key concept explored was the Root-Variables Hypothesis. A surprising finding was that even when the "root cause" of a trait is known for all animals, humans still use similarity to predict the outcome. This suggests we assume there are always unobserved mediating factors that follow taxonomic patterns. Similarity acts as a proxy for the unknown: if two twins both get cancer despite only one smoking, we infer there are hidden genetic or environmental factors they share due to their similarity.

Key Results and Model Performance

The researchers compared human ratings against three computational models: the Combined Model, a Causal-only model, and a Similarity-only model.

Comparison of human judgments against the Combined, Causal-only, and Similarity-only models.
  • Superior Correlation: The Combined Model achieved very high correlations with human judgment ($r$ values between 0.91 and 0.99 across various tasks).
  • Failure of Isolated Models: The Causal-only model failed because it couldn't distinguish between a rat and a sheep if the causal evidence was the same. The Similarity-only model failed because it couldn't account for how one enzyme might trigger another.
  • Qualitative Match: In specific "chain conditions," the causal model predicted that a rat and a sheep were equally likely to have a downstream enzyme. However, humans (and the combined model) correctly identified that the rat was less likely because it was "closer" to a known negative result in the similarity tree.

Limitations and Future Work

While highly successful at predicting human behavior in these tasks, the model has limitations. It uses a simplified tree structure for similarity, whereas real human similarity judgments can be multidimensional or based on different types of categories (e.g., "desert animals" vs. "mammals"). Additionally, as the number of objects and properties grows, computing the joint distribution becomes increasingly complex.

Future work could expand this framework to handle interventions (e.g., predicting what happens if an animal is given a drug to inhibit an enzyme, rather than just passively observing). Ultimately, causal and similarity-based reasoning are just two types of knowledge; integrating scripts, analogies, or more complex intuitive theories into the same probabilistic language remains an exciting frontier for cognitive science.

Summary updated

Abstract

Everyday inductive reasoning draws on many kinds of knowledge, including knowledge about relationships between properties and knowledge about relationships between objects. Previous accounts of inductive reasoning generally focus on just one kind of knowledge: models of causal reasoning often focus on relationships between properties, and models of similarity-based reasoning often focus on similarity relationships between objects. We present a Bayesian model of inductive reasoning that incorporates both kinds of knowledge, and show that it accounts well for human inferences about the properties of biological species.
BibTeX
@inproceedings{kemp2006combining,
  author = {Charles Kemp and Patrick Shafto and Allison Berke and Joshua B. Tenenbaum},
  title = {{Combining causal and similarity-based reasoning}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2006},
  url = {https://openreview.net/forum?id=r1m51vWu-r},
}
Figure