Visual Analysis and Interactive Comparison for Heterogeneous Information Network Embedding Model

Abstract

<p indent="0mm">To understand the differences between various heterogeneous information networks (HINs) embedding models, and to solve intricate and hidden problems in model evaluation, the comparative analysis method should include the following steps. First, the evaluation indicators and tasks are made consistent for all models. Then the parameters and structural features are extracted from the model training process and the complete and non-averaged evaluation results are retained for visualization. Based on the model parameters and characteristics data, we design and implement a visual analysis tool named HINCompare, which includes an overview of the distribution of basic evaluation indicators, a comparison view of recommended results, and a feature view of the local topology structure for model embedding. The tool allows developers to explore the common patterns of different feature aggregation methods of different models and the differences in their architectures. In addition, HINCompare shows user’s preferences for movie types and years with heat maps, which are combined with the contextual information of the recommended results for further analysis and evaluation. The system provides insights into black-box problems of models and increases interpretability by supplying information about the sources of the results. We conduct the preliminary evaluation study with Douban movie data to verify the effectiveness of the system.