Prefer2SD: A Human-in-the-Loop Approach to Balancing Similarity and Diversity in In-Game Friend Recommendations

TL;DR

Balancing similarity and diversity in in-game friend recommendations is challenging due to evolving player preferences. The authors present Prefer2SD, a two-step human-in-the-loop workflow with a visual analytics interface that lets algorithm experts mediate the SD ratio for representative players (Step 1) and propagate these ratios to the broader group via label propagation and active learning (Step 2). The approach is validated through a within-subjects study (N=12), a case study, and expert interviews, showing improved SD balance and recommendation quality over a baseline, along with reduced cognitive load. The work demonstrates a practical, scalable framework for human-in-the-loop customization of multi-modal recommendations in dynamic online games, with potential applicability to other personalized recommendation domains.

Abstract

In-game friend recommendations significantly impact player retention and sustained engagement in online games. Balancing similarity and diversity in recommendations is crucial for fostering stronger social bonds across diverse player groups. However, automated recommendation systems struggle to achieve this balance, especially as player preferences evolve over time. To tackle this challenge, we introduce Prefer2SD (derived from Preference to Similarity and Diversity), an iterative, human-in-the-loop approach designed to optimize the similarity-diversity (SD) ratio in friend recommendations. Developed in collaboration with a local game company, Prefer2D leverages a visual analytics system to help experts explore, analyze, and adjust friend recommendations dynamically, incorporating players' shifting preferences. The system employs interactive visualizations that enable experts to fine-tune the balance between similarity and diversity for distinct player groups. We demonstrate the efficacy of Prefer2SD through a within-subjects study (N=12), a case study, and expert interviews, showcasing its ability to enhance in-game friend recommendations and offering insights for the broader field of personalized recommendation systems.

Figures (14)

• Figure 1: Workflow of multi-stage recommendation systems: Illustrated is an example of multi-channel matching for candidate generation. (A) Multi-channel classifies various items into specific channels, forming an extensive recommendation pool. (B) Candidate Generation efficiently narrows down items from millions to hundreds using models designed for large data handling. (C-D) The Ranking stage then sorts items by predicted user relevance, delivering a top@N list through sophisticated models.
• Figure 2: Framework of Prefer2SD. It introduces a two-step human-in-the-loop workflow. Step 1: Achieve an optimal SD ratio among representative players in a specific group by mediating their preference ratios. Step 2: Extend the preference ratios from Step 1 across the entire player cohort to achieve the desired group-level SD ratio.
• Figure 3: Framework of Prefer2SD supported by the Backend and the Frontend. (A-C) & (F-I) Backend and Frontend sections corresponding to Step 1. (D-E) & (J) Backend and Frontend sections for Step 2.
• Figure 4: The Prefer2SD interface incorporates six main views, each serving a distinct purpose. Color-coded labels enhance navigation: blue indicates Step 1, purple represents Step 2, and black highlights guidance. (A) The Guidance View provides an overview of progress and directs subsequent actions. (B) The Preference Projection View enables selection of specific player groups. (C) The Representative Player Selection View assists in choosing a representative player. (D) The Preference Ratio Adjustment View supports iterative mediation of preference ratios for ideal SD ratios. (E) The Propagation with Active Learning View helps identify uncertain players post-propagation. (F) The Result Comparison View summarizes recommendation results across iteration.
• Figure 5: Sampling and Fusion Design. (A) Sampling: a scatter plot of candidates created by radial projection. Concentric circles divide the candidates into four classes.