Predicting Preference Reversals via Gaussian Process Uncertainty Aversion

Rikiya Takahashi, Tetsuro MorimuraAISTATS 2015 · OpenReview

Abstract

Modeling of a product or service’s attractiveness as a function of its own attributes (e.g., price and quality) is one of the foundations in econometric forecasts. Standard economic models assume perfect rationality, yet humans frequently change their minds based on the "context" of available choices. This paper introduces a new cognitive choice model, GPUA (Gaussian Process Uncertainty Aversion), which explains these irrationalities as a form of uncertainty aversion. By leveraging Gaussian Process Regression, the model captures complex psychological nuances while remaining computationally efficient through convex optimization.

Introduction: The Challenge of Context Effects

In classical econometrics, models like the "Logit" model rely on the Independence from Irrelevant Alternatives (IIA) principle. This principle suggests that the relative preference between two options should not change when a third, irrelevant option is introduced. However, experimental psychology has documented three major "context effects" that consistently violate IIA:

  1. Similarity Effect: A new option "steals" market share more from similar items than dissimilar ones.
  2. Attraction Effect: Adding a "decoy" (an option clearly inferior to one existing choice) increases the popularity of the superior choice.
  3. Compromise Effect: People tend to avoid extremes and pick the "middle" option in a set.
The Attraction and Compromise effects demonstrate how introducing new options can shift consumer preference toward specific existing choices, violating standard economic principles.

Existing models that account for these effects are often computationally difficult (non-convex) or require extensive manual feature engineering. The authors aim to create a model that captures these psychological nuances while remaining easy to train.

Methods: Gaussian Process Uncertainty Aversion

The authors propose the GPUA model, which treats the human decision-maker as a Bayesian agent experiencing a mental conflict between two systems: a rational assessment of objective attributes and a subjective prior shaped by the current choice set.

The core mechanism is uncertainty aversion. When a decision-maker evaluates a set of options, they use similarities between them to "interpolate" value. If an option is an outlier (isolated from others), the Gaussian Process assigns it a high posterior variance (uncertainty). The model assumes humans are uncertainty-averse and "shrink" the value of these uncertain, isolated options toward a mean of zero. This mathematical shrinkage naturally produces the attraction and compromise effects without needing complex, hand-coded rules.

The GPUA model's final evaluation shows how the introduction of a decoy (A-) increases the perceived value of the target (A) relative to a competitor (D) by reducing uncertainty.

To ensure scalability, the model utilizes Random Fourier Features to approximate the kernel. Crucially, the training objective is globally convex, allowing it to be solved reliably and quickly using standard Newton-Raphson methods.

Results and Datasets

The model was evaluated on three datasets:

  1. PC Dataset (n=1,088): Choices between portable computers with varying CPU speeds and memory.
  2. SP Dataset (n=972): Choices between audio speakers based on power and price.
  3. SwissMetro (n=10,719): A large-scale real-world dataset regarding transportation preferences in Switzerland.

GPUA outperformed standard linear and non-parametric Logit models across all datasets in terms of both log-likelihood and classification accuracy. In the PC and SP datasets, GPUA successfully predicted the preference for "middle" options (the compromise effect), whereas standard models failed because they couldn't "see" the context of the choice set. On the large SwissMetro dataset, GPUA was the second-best performer, nearly matching the highly complex Non-parametric Mixed Logit (NpMix) but with a much simpler and more interpretable framework.

Discussion: Irrationality as Bayesian Inference

The paper's most elegant insight is that "irrational" context effects aren't necessarily failures of logic, but a byproduct of how humans handle uncertainty when comparing items. By framing the problem as a GPR-based shrinkage task, the authors turned a traditionally difficult optimization problem into a convex one, making it practical for machine learning pipelines.

Future work will focus on testing the model on higher-dimensional choice tasks and engineering more detailed covariance functions to gain deeper economic insights into which specific attributes drive irrational decisions. Potential applications include improving travel planning or product recommendation engines where context-dependent "preference reversals" are common.

Summary updated

Abstract

Modeling of a product or service’s attractiveness as a function of its own attributes (e.g., price and quality) is one of the foundations in econometric forecasts, which have been provided with an ...
BibTeX
@inproceedings{takahashi2015predicting,
  author = {Rikiya Takahashi and Tetsuro Morimura},
  title = {{Predicting Preference Reversals via Gaussian Process Uncertainty Aversion}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2015},
  url = {https://openreview.net/forum?id=F_q0vjCZ1Qx},
}
Figure