Latent Graph Structure Learning for Service-Oriented Complex Network: Task Definition, Data Simulation, and Applications in Service Ecosystem

In service-oriented systems, complex architectures frequently exhibit latent patterns that can be effectively represented by graph-based structures. For instance, an industrial ecosystem consists of interconnected industrial chains rather than isolated interactions among individual enterprises. Uncovering these latent patterns is critical for understanding the operational mechanisms of service systems and formulating optimization strategies. Currently, this challenge has not received adequate attention. The ambiguity in problem definitions and application scenarios, coupled with the scarcity of available datasets, remains a primary obstacle to in-depth research. In this paper, we define this problem as the Latent Graph Learning (LGL) task and propose a formal definition along with a data simulator to mitigate these challenges. We also explore potential application scenarios to demonstrate practical implications. Our experiments evaluate the performance of existing models on the LGL task, revealing that current methods exhibit poor performance when applied directly. Finally, we discuss future directions and potential advancements for the LGL task. The code can be found at https://github.com/ssnowyu/Latent-Graphs-Learning.