# A new system-wide diversity measure for recommendations with efficient algorithms

**Authors:** Arda Antikacioglu, Tanvi Bajpai, R. Ravi  
**arXiv:** [1812.03030](https://arxiv.org/abs/1812.03030)  
**Date:** 2018-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.03030v1/

## Key points

- Authors: Arda Antikacioglu, Tanvi Bajpai, R. Ravi
- Introduce two novel system-wide diversity metrics for recommender systems
- Metrics aim to diversify item categories and broaden user exposure while maintaining high recommendation quality
- Frame problem as subgraph selection dilemma in bipartite graph of potential recommendations between users and items
- Solutions achievable in polynomial time for disjoint item categories and user types
- NP-completeness established for non-disjoint categories and user types
- Propose efficient approximation algorithms leveraging submodularity property of objective function
- Experiments conducted using MovieLens-1m and Netflix datasets to validate algorithms' efficacy
- Algorithms tailored for defined objective perform well based on sales diversity metrics and intent-aware diversity metrics
- Composite diversity metrics enhance recommendation systems' overall performance and inclusivity

## Summary

In their paper titled "A new system-wide diversity measure for recommendations with efficient algorithms," authors Arda Antikacioglu, Tanvi Bajpai, and R. Ravi delve into the realm of recommender systems and introduce two novel system-wide diversity metrics. These metrics aim to address the challenges of diversifying item categories and broadening user exposure while maintaining high recommendation quality standards. The authors frame this problem as a subgraph selection dilemma within the bipartite graph of potential recommendations between users and items. They demonstrate that in scenarios where item categories and user types are disjoint, solutions can be achieved in polynomial time through a reduction to a minimum cost flow problem. However, when dealing with non-disjoint categories and user types, they establish the NP-completeness of the objective and propose efficient approximation algorithms leveraging the submodularity property of the objective function. To validate their proposed algorithms' efficacy, experiments were conducted using datasets from MovieLens-1m and Netflix. The results indicate that algorithms tailored for their defined objective not only perform well based on sales diversity metrics but also exhibit promising outcomes concerning some intent-aware diversity metrics. This empirical evidence underscores the credibility and utility of their newly introduced composite diversity metrics in enhancing recommendation systems' overall performance and inclusivity.