[Paper Review] Using Collaborative Filtering to Recommend Champions in League of Legends

D. Tiffany, Dylan S. Yu+2 more|arXiv (Cornell University)|Jun 17, 2020

Digital Games and Media14 references4 citations

TL;DR

This paper proposes a collaborative filtering-based champion recommendation system for League of Legends using Singular Value Decomposition (SVD) on Champion Mastery Points (CMP) as a proxy for player preference. The system outperforms random recommendations in a preliminary user study, demonstrating that CMP effectively captures player enjoyment and enables personalized, enjoyable champion suggestions.

ABSTRACT

League of Legends (LoL), one of the most widely played computer games in the world, has over 140 playable characters known as champions that have highly varying play styles. However, there is not much work on providing champion recommendations to a player in LoL. In this paper, we propose that a recommendation system based on a collaborative filtering approach using singular value decomposition provides champion recommendations that players enjoy. We discuss the implementation behind our recommendation system and also evaluate the practicality of our system using a preliminary user study. Our results indicate that players significantly preferred recommendations from our system over random recommendations.

Motivation & Objective

• To address the lack of personalized champion recommendation systems in League of Legends beyond win-rate optimization.
• To explore whether Champion Mastery Points (CMP) can serve as a reliable proxy for player preference and enjoyment.
• To design and evaluate a collaborative filtering system that recommends champions based on user engagement rather than win-rate synergy.
• To validate the practicality and user preference of SVD-based recommendations over random recommendations through a preliminary user study.

Proposed method

• The system uses an unbiased SVD algorithm to model user-champion interactions in a low-rank latent factor space.
• Champion Mastery Points (CMP) are used as implicit feedback to represent player preference and engagement with champions.
• The SVD model is trained on a dataset of 311,727 player-champion pairs, with parameters tuned via grid search to minimize root mean square error.
• The model maps both players and champions into a shared latent space where recommendations are generated as inner products of user and champion vectors.
• Alternative algorithms like Slope One and SVD++ were tested but discarded due to high popularity bias or excessive runtime.
• A user study compared SVD-based recommendations against random recommendations to assess user preference.

Experimental results

Research questions

• RQ1 Can Champion Mastery Points (CMP) effectively represent a player's preference for champions in League of Legends?
• RQ2 Does a collaborative filtering system using SVD on CMP data produce recommendations that players prefer over random recommendations?
• RQ3 How do alternative recommendation algorithms like Slope One and SVD++ compare in terms of accuracy and practicality for this use case?
• RQ4 To what extent do metagame shifts and champion reworks affect the reliability of CMP-based recommendations?

Key findings

• Players significantly preferred recommendations from the SVD-based system over random recommendations in the preliminary user study.
• The optimal SVD model parameters were found to be epoch=20, λ=0.4, and γ=0.0005, which minimized root mean square error and improved theoretical accuracy.
• Slope One and SVD++ models were outperformed due to high popularity bias and impractically long training times, respectively.
• CMP was validated as a useful and practical measure for capturing player champion preference, even in the absence of explicit ratings.
• The system effectively mitigates popularity bias by leveraging latent factor decomposition, enabling recommendations for less popular champions.
• Limitations include potential bias from champion reworks and evolving metagames, as well as a non-representative user sample in the initial study.