Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ConjointNet: Enhancing Conjoint Analysis for Preference Prediction with Representation Learning

Yanxia Zhang, Francine Chen, Shabnam Hakimi, Totte Harinen, Alex Filipowicz, Yan-Ying Chen, Rumen Iliev, Nikos Arechiga, Kalani Murakami, Kent Lyons, Charlene Wu, Matt Klenk

TL;DR

The paper targets the limitations of linear conjoint analysis in capturing non-linear attribute interactions. It introduces ConjointNet, comprising a semi-supervised autoencoder-based architecture and a residual ConjointNet that combines linear and non-linear components to model complex preferences. Across the Moral Machine and Car Preference datasets, ConjointNet consistently outperforms traditional conjoint models in accuracy (and AUC on MM), demonstrating the value of representation learning for preference prediction. The approach enables end-to-end learning from raw choice data, reduces reliance on hand-crafted features, and opens avenues for integrating additional modalities and visualizing the learned interactions.

Abstract

Understanding consumer preferences is essential to product design and predicting market response to these new products. Choice-based conjoint analysis is widely used to model user preferences using their choices in surveys. However, traditional conjoint estimation techniques assume simple linear models. This assumption may lead to limited predictability and inaccurate estimation of product attribute contributions, especially on data that has underlying non-linear relationships. In this work, we employ representation learning to efficiently alleviate this issue. We propose ConjointNet, which is composed of two novel neural architectures, to predict user preferences. We demonstrate that the proposed ConjointNet models outperform traditional conjoint estimate techniques on two preference datasets by over 5%, and offer insights into non-linear feature interactions.

ConjointNet: Enhancing Conjoint Analysis for Preference Prediction with Representation Learning

TL;DR

The paper targets the limitations of linear conjoint analysis in capturing non-linear attribute interactions. It introduces ConjointNet, comprising a semi-supervised autoencoder-based architecture and a residual ConjointNet that combines linear and non-linear components to model complex preferences. Across the Moral Machine and Car Preference datasets, ConjointNet consistently outperforms traditional conjoint models in accuracy (and AUC on MM), demonstrating the value of representation learning for preference prediction. The approach enables end-to-end learning from raw choice data, reduces reliance on hand-crafted features, and opens avenues for integrating additional modalities and visualizing the learned interactions.

Abstract

Understanding consumer preferences is essential to product design and predicting market response to these new products. Choice-based conjoint analysis is widely used to model user preferences using their choices in surveys. However, traditional conjoint estimation techniques assume simple linear models. This assumption may lead to limited predictability and inaccurate estimation of product attribute contributions, especially on data that has underlying non-linear relationships. In this work, we employ representation learning to efficiently alleviate this issue. We propose ConjointNet, which is composed of two novel neural architectures, to predict user preferences. We demonstrate that the proposed ConjointNet models outperform traditional conjoint estimate techniques on two preference datasets by over 5%, and offer insights into non-linear feature interactions.

Paper Structure

This paper contains 20 sections, 3 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Discrete Choice Models
  4. Feature Selection
  5. Representation Learning
  6. Choice-based Conjoint Analysis
  7. ConjointNet: Conjoint Analysis with Representation Learning
  8. ConjointNet with Semi-supervised Learning
  9. Autoencoders for Representation Learning
  10. Choice Classification Network
  11. Residual ConjointNet
  12. Experiments
  13. Data Preprocessing
  14. The Moral Machine (MM) Dataset
  15. Car Preference Dataset
  16. ...and 5 more sections

Figures (5)

  • Figure 1: Proposed SSL ConjointNet Architecture on Choice Problems
  • Figure 2: The proposed Residual ConjointNet jointly enforce linear and non-linear components.
  • Figure 3: We show the reconstruction results of an unseen sample (left) from the MM dataset with a VAE (center) and a plain AE (right). The attributes are converted with a one-hot encoding where the corresponding level between 0 and 5 is denoted as 1 otherwise 0.
  • Figure 4: Training and testing accuracy on the car preference dataset with 16 hidden nodes.
  • Figure 5: Training and testing accuracy on the car preference dataset with 64 hidden nodes.