Self-Supervised Prototype Representation Learning for Event-Based Corporate Profiling

Event-based corporate profiling aims to assess the evolving operational status of corresponding from its event sequence. Existing studies on have partially addressed problem via (i) case-by-case empirical analysis by leveraging traditional financial methods, or (ii) automatic profile inference reformulating into a supervised learning task. However, both approaches heavily rely domain knowledge and are labor-intensive. More importantly, task-specific nature prevents obtained profiles being applied diversified downstream applications. To this end, in paper, we propose Self-Supervised Prototype Representation Learning (SePaL) framework for dynamic profiling. By exploiting topological information an graph exploring self-supervised techniques, SePaL can obtain unified representations that robust noises be easily fine-tuned benefit various down-stream applications with only few annotated data. Specifically, first infer initial cluster distribution noise-resistant prototypes based latent events. Then, construct four permutation-invariant self-supervision signals guide representation prototype. In terms applications, exploit learned time-evolving stock price spike prediction default risk evaluation. Experimental results two real-world datasets demonstrate effectiveness these