Cross-domain Correspondences for Explainable Recommendations

Aaron Stockdill, Daniel Raggi, Mateja Jamnik, Grecia Garcia Garcia, Holly E. A. Sutherland, Peter C.-h. Cheng, Advait Sarkar

Introduction

Modern recommender systems primarily rely on statistical correlations, often operating as "black boxes" that suggest items based on user similarity (e.g., "Users who liked X also liked Y"). While effective for prediction, these systems struggle to provide transparent logic, particularly when recommending items across disparate domains—such as suggesting a book based on a mathematical interest. This paper introduces a formal framework for "correspondences," which are logical links between domains that enable AI to provide human-readable, analogical justifications for its recommendations.

The rep2rep Framework

The authors propose the rep2rep framework to formalize analogies as a structured set of correspondences. The process begins by breaking down every domain or representation into discrete properties. Each property is defined by a kind (e.g., a token), a value (e.g., the number 1), and a set of attributes (e.g., "is a number").

Example of property formalization for algebraic and dot-based domains.

By standardizing how data is represented across different domains, the system can begin to identify functional similarities between seemingly unrelated objects.

Defining and Discovering Correspondences

A correspondence is formally defined as a triple $\langle p_1, p_2, s \rangle$, where $p_1$ and $p_2$ are properties in different domains and $s$ represents the "strength" of the link. This strength is calculated probabilistically, measuring how much the presence of $p_1$ increases the likelihood of $p_2$, adjusted for baseline frequencies.

To scale the system without manual mapping, the authors introduce five logic-based discovery rules:

  • Identity: Identical properties correspond perfectly.
  • Reversal: If a link exists from A to B, a reciprocal link likely exists from B to A.
  • Composition: If A corresponds to B and B to C, the system infers a link between A and C.
  • Attribute/Value Rules: Shared attributes (e.g., "is a character") allow the system to infer correspondences between the items themselves.
Mapping correspondences between properties across different representation domains.

Key Insights: Analogies as Roles

The core innovation of this framework is treating analogies as functional roles. Correspondences link elements that "fill the same role" within their respective systems. For instance, the algebraic summation symbol $\sum$ might correspond to the visual action of "stacking" in a dot diagram.

This approach supports two modes of explanation:

  1. Descriptive: Explaining why two existing items are similar (e.g., "This film is like this book because the protagonists share similar narrative roles").
  2. Constructive: Guiding a user to build a new solution by translating concepts from a familiar domain to an unfamiliar one.

Results and Case Studies

The framework was tested on two illustrative datasets: mathematical representations (algebra vs. dot diagrams) and media recommendations (films vs. books). In the math example, the system successfully mapped the concept of summation to visual dot stacking by using the Attribute Rule to chain logic from simple seeds (e.g., 1 = one dot). In the media example, the system deduced a correspondence between the film character Rick Deckard (Blade Runner) and the book character Elijah Baley (The Caves of Steel) based on their shared roles and attributes.

Limitations and Future Work

While the logical rules are powerful, they can occasionally generate "nonsense" or unsound correspondences if properties share generic attributes. Furthermore, the current model relies on expert-defined prior probabilities. Future research will focus on automating the calculation of these priors from large datasets and developing natural language generation techniques to translate these logical links into conversational explanations for end-users.

Summary updated

BibTeX
@inproceedings{stockdill2020cross,
  author = {Aaron Stockdill and Daniel Raggi and Mateja Jamnik and Grecia Garcia Garcia and Holly E. A. Sutherland and Peter C.-H. Cheng and Advait Sarkar},
  title = {{Cross-domain Correspondences for Explainable Recommendations}},
  booktitle = {ExSS-ATEC@IUI},
  year = {2020},
}
Figure