Observation Is Reality? A Graph Diffusion-Based Approach for Service Tags Recommendation

Shuang Yu, Qingfeng Li, Mingyi Liu, Zhongjie Wang

Abstract: Accurate service tags recommendation plays a crucial role in classifying, searching, managing, composing, and expanding services. However, many service tags recommendation studies fail to consider real-world scenarios, greatly limiting their performance and capability of handling complex situations. First, the simplification of service tags recommendation to single-tag classification or clustering overlooks the complexity and diversity of crossover services, as well as the intricate interactions between services or their tags. Second, inadequate or ambiguous descriptions of many services result in insufficient information for accurate recommendations. Third, the observation is not always reality due to the presence of unseen data or noise. To address these issues, a new graph diffusion-based graph neural network framework is proposed for multi-tags recommendation, named SpiderTags. It considers both the textual description of services and explicit relationships between services or their tags to enhance performance. Moreover, considering that the observed explicit graph may not be reality and not optimal for downstream tasks, SpiderTags introduces a graph diffusion mechanism to search for a more optimal graph for downstream tasks. A series of experiments conducted on the real-world ProgrammableWeb dataset demonstrate the effectiveness of SpiderTags in service tags recommendation task. Our code is available on https://github.com/gplinked/SpiderTags.