Content recommendation tasks increasingly use Graph Neural Networks, but it remains challenging for machine learning experts to assess the quality of their outputs. Visualization systems for GNNs that could support this interrogation are few. Moreover, those that do exist focus primarily on exposing GNN architectures for tuning and prediction tasks and do not address the challenges of recommendation tasks. We developed RekomGNN, a visual analytics system that supports ML experts in exploring GNN recommendations across several dimensions and making annotations about their quality. RekomGNN straddles the design space between Neural Network and recommender system visualization to arrive at a set of encoding and interaction choices for recommendation tasks. We found that RekomGNN helps experts make qualitative assessments of the GNN's results, which they can use for model refinement. Overall, our contributions and findings add to the growing understanding of visualizing GNNs for increasingly complex tasks.

翻译：内容推荐任务日益广泛采用图神经网络，但机器学习专家在评估其输出质量时仍面临挑战。目前支持此类检验的图神经网络可视化系统较为稀缺。现有系统主要聚焦于揭示用于调参和预测任务的GNN架构，未能解决推荐任务的特殊问题。我们开发了RekomGNN——一个支持机器学习专家从多维度探索GNN推荐结果并对其质量进行标注的可视分析系统。该系统跨越神经网络与推荐系统可视化的设计空间，形成面向推荐任务的编码与交互方案。研究发现，RekomGNN能够帮助专家对GNN结果进行定性评估，并据此优化模型。总体而言，我们的贡献与发现为理解针对日益复杂任务的GNN可视化方法提供了新的视角。