Reproduce “Prompt-augmented Temporal Point Process for Streaming Event Sequence” on OpenResearch

About this paper

Siqiao Xue, Yan Wang, Zhixuan Chu, Xiaoming Shi, Caigao Jiang, Hongyan Hao, Gangwei Jiang, Xiaoyun Feng and 2 more

Researchers at Ant Group developed PromptTPP, a framework that integrates neural Temporal Point Processes with a continuous-time retrieval prompt pool, enabling models to continually learn from streaming event data without rehearsal buffers or explicit task identities. It achieved state-of-the-art performance on three real-world datasets, demonstrating superior next-event type and time prediction while mitigating catastrophic forgetting.

Key insights

• Prompt learning can be effectively adapted for continuous-time event sequence modeling, enabling neural TPPs to continually learn from streaming data without reliance on data buffers or explicit task identities.
• The use of 'temporal prompts' that incorporate continuous-time positional encodings is crucial for capturing dynamic temporal patterns and significantly mitigating catastrophic forgetting in sequential learning.
• A retrieval prompt pool with learnable keys facilitates efficient knowledge storage and dynamic selection, decoupling query learning from prompt learning and enhancing model adaptability to evolving data distributions.

Summary by alphaXiv.