# PRINCE: Provider-side Interpretability with Counterfactual Explanations in Recommender Systems

**Authors:** Azin Ghazimatin, Oana Balalau, Rishiraj Saha Roy, Gerhard Weikum  
**arXiv:** [1911.08378](https://arxiv.org/abs/1911.08378)  
**Date:** 2019-11-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1911.08378v1/

## Key points

- Interpretable explanations are crucial for establishing user trust in recommender systems and other machine learning models
- PRINCE is a provider-side mechanism that generates tangible explanations for end-users
- PRINCE identifies the minimal set of actions required to alter a recommendation using an optimal algorithm and random walks over dynamic graphs
- Experimental results show that PRINCE offers more concise explanations compared to intuitive baselines
- PRINCE produces scrutable, actionable, and succinct explanations by leveraging counterfactual evidence, a user's own actions, and minimal sets respectively

## Summary

, , , , 

Interpretable explanations are crucial for establishing user trust in recommender systems and other machine learning models. However, previous approaches have fallen short in providing true explanations or addressing privacy concerns. To address these limitations, this study introduces PRINCE: a provider-side mechanism that generates tangible explanations for end-users. These explanations consist of minimal actions taken by the user that would result in a different recommendation being made if removed. PRINCE utilizes an optimal algorithm to navigate through a search space using random walks over dynamic graphs to identify the minimal set of actions required to alter a recommendation. Experimental results from real-world datasets demonstrate that PRINCE offers more concise explanations compared to intuitive baselines. A crowdsourced user-study also confirms the practicality and effectiveness of action-based explanations provided by PRINCE. The key strengths of PRINCE lie in its ability to produce scrutable, actionable, and succinct explanations by leveraging counterfactual evidence, a user's own actions, and minimal sets respectively. By focusing on individual user actions' impact on recommendations, PRINCE enhances transparency and trustworthiness in machine learning models while addressing the need for interpretable explanations in complex systems like recommender algorithms.