Personalized bundle recommendation using preference elicitation and the Choquet integral

Erich Robbi, Marco Bronzini, Paolo Viappiani, Andrea PasseriniFrontiers Artif. Intell 2024 · DOI · OpenReview

Introduction

Recommending a "bundle"—a set of related items—is significantly more complex than recommending a single product. Traditional recommendation systems often calculate a bundle's value by simply summing the scores of its individual attributes. However, this linear approach fails to capture the non-linear synergies common in real-world scenarios. For example, in sustainable e-commerce, two products are significantly more eco-friendly if they share a warehouse, reducing total shipping trips. This paper introduces a novel framework that utilizes the Choquet integral to model these complex interactions and an interactive preference elicitation strategy to learn user priorities without requiring massive historical datasets.

The Proposed Framework

The authors propose a multi-stage process for generating personalized bundle recommendations. The system first identifies candidate products and generates potential subsets (bundles). These subsets are then evaluated using the Choquet integral, which can be configured with either fixed criteria or personalized criteria derived from user interaction.

General framework for bundle recommendation

Capturing Synergies with the Choquet Integral

Unlike a standard weighted sum, the Choquet integral is a non-linear aggregation function. It utilizes "capacities" to assign weights to coalitions of attributes. This allows the model to recognize that certain attributes are more valuable when they appear together—essentially modeling the idea that the whole can be greater than the sum of its parts.

By representing the Choquet integral through a Möbius transform, the system can also identify which specific groups of attributes contributed most to a recommendation. This provides a clear path for generating human-readable explanations, enhancing the transparency of the recommendation process.

Interactive Preference Elicitation

To address the "cold-start" problem—where a new user's specific values are unknown—the system engages the user in a short query phase. It asks the user to choose between hypothetical bundles, using the Minimax Regret (MMR) criterion to select the most informative questions. This process iteratively reduces the "worst-case loss" of making a sub-optimal recommendation.

Diagram of the adopted elicitation process

As shown in the diagram, the system finds the parameters that minimize the Worst-case Minimax Regret (WMMR), queries the user, and updates the weight space ($W$) until an optimal recommendation can be made with high confidence.

Evaluation and Key Results

The researchers evaluated their approach using real-world data from a local food e-commerce platform in Northern Italy and five synthetic user personas (e.g., "Net-zero," "Healthy," "Ethical").

  • Superior Sustainability: The Choquet-based approach consistently recommended bundles with higher environmental scores than linear baselines, as it could prioritize synergies like shared warehouse locations.
  • Efficiency of Interaction: The study found that a handful of queries (typically between 2 and 11) was sufficient to reach zero regret, meaning the system successfully identified the mathematically optimal bundle for the user's hidden preferences.
  • Personalization Accuracy: The system successfully pivoted its recommendations based on persona. For the "Healthy" persona, it prioritized low-sodium and sugar-free combinations, while for the "Net-zero" persona, it prioritized bundles that minimized shipping footprints.

Limitations and Future Work

While effective, the framework faces computational challenges; calculating the Choquet integral and minimax regret involves looking at $2^N$ possible subsets of attributes, which may become a bottleneck as the number of features grows. Additionally, while 11 queries is a small number for research, it may still cause user fatigue in a casual shopping context. Future work will explore hybrid models that combine this elicitation with collaborative filtering and conduct real-world human studies to validate the intuitiveness of the interactive queries.

Summary updated

Abstract

p>Bundle recommendation aims to generate bundles of associated products that users tend to consume as a whole under certain circumstances. Modeling the bundle utility for users is a non-trivial task, as it requires to account for the potential interdependencies between bundle attributes. To address this challenge, we introduce a new preference-based approach for bundle recommendation exploiting the Choquet integral. This allows us to formalize preferences for coalitions of environmental-related attributes, thus recommending product bundles accounting for synergies among product attributes. An experimental evaluation of a dataset of local food products in Northern Italy shows how the Choquet integral allows the natural formalization of a sensible notion of environmental friendliness and that standard approaches based on weighted sums of attributes end up recommending bundles with lower environmental friendliness even if weights are explicitly learned to maximize it. We further show how preference elicitation strategies can be leveraged to acquire weights of the Choquet integral from user feedback in terms of preferences over candidate bundles, and show how a handful of queries allow to recommend optimal bundles for a diverse set of user prototypes.</p>
BibTeX
@article{robbi2024personalized,
  author = {Erich Robbi and Marco Bronzini and Paolo Viappiani and Andrea Passerini},
  title = {{Personalized bundle recommendation using preference elicitation and the Choquet integral}},
  journal = {Frontiers Artif. Intell.},
  year = {2024},
  doi = {10.3389/frai.2024.1346684},
}
Figure