[Paper Review] Prompt-augmented Temporal Point Process for Streaming Event Sequence

Siqiao Xue, Yan Wang+8 more|arXiv (Cornell University)|Oct 8, 2023

Human Pose and Action RecognitionComputer Science3 citations

TL;DR

PromptTPP proposes a novel continual learning framework for streaming event sequences using a continuous-time retrieval prompt pool, enabling efficient, memory-efficient, and task-agnostic learning without rehearsal buffers or task identity. It achieves state-of-the-art performance across three real-world user behavior datasets by jointly optimizing learnable prompts and a base neural temporal point process.

ABSTRACT

Neural Temporal Point Processes (TPPs) are the prevalent paradigm for modeling continuous-time event sequences, such as user activities on the web and financial transactions. In real-world applications, event data is typically received in a \emph{streaming} manner, where the distribution of patterns may shift over time. Additionally, \emph{privacy and memory constraints} are commonly observed in practical scenarios, further compounding the challenges. Therefore, the continuous monitoring of a TPP to learn the streaming event sequence is an important yet under-explored problem. Our work paper addresses this challenge by adopting Continual Learning (CL), which makes the model capable of continuously learning a sequence of tasks without catastrophic forgetting under realistic constraints. Correspondingly, we propose a simple yet effective framework, PromptTPP\footnote{Our code is available at {\small \url{ https://github.com/yanyanSann/PromptTPP}}}, by integrating the base TPP with a continuous-time retrieval prompt pool. The prompts, small learnable parameters, are stored in a memory space and jointly optimized with the base TPP, ensuring that the model learns event streams sequentially without buffering past examples or task-specific attributes. We present a novel and realistic experimental setup for modeling event streams, where PromptTPP consistently achieves state-of-the-art performance across three real user behavior datasets.

Motivation & Objective

• Address the challenge of continual learning for streaming event sequences under real-world constraints like privacy and memory limits.
• Overcome catastrophic forgetting in neural temporal point processes (TPPs) when processing sequential data in real time.
• Develop a task-agnostic, memory-efficient method that avoids rehearsal buffers and does not require task identity at inference.
• Enable effective knowledge transfer and plasticity in TPPs through a learnable prompt-based mechanism for continuous-time event modeling.

Proposed method

• Introduce a continuous-time retrieval prompt pool that stores small, learnable parameters as prompts in a key-value memory space.
• Design a dynamic retrieval mechanism that selects task-relevant prompts based on input event sequences using query-key matching.
• Jointly optimize the base neural TPP and the prompt pool via a generative loss to preserve shared knowledge and adapt to task-specific patterns.
• Implement an asynchronous prompt update mechanism with refresh frequency $C$ to improve training efficiency without sacrificing performance.
• Structure prompts in a key-value shared memory space to decouple query and prompt learning, enhancing retrieval accuracy and model stability.
• Ensure compatibility with any neural TPP architecture by treating the prompt pool as a modular, plug-in component.

Figure 1 : Overview of the classical schemes and PromptTPP framework for streaming event sequences.

Experimental results

Research questions

• RQ1 Can a prompt-augmented framework effectively mitigate catastrophic forgetting in neural temporal point processes under streaming, real-time data conditions?
• RQ2 How does the retrieval-based prompt mechanism compare to traditional rehearsal-based or task-identity-dependent continual learning methods in streaming event modeling?
• RQ3 To what extent does the prompt pool size, prompt length, and selection size influence model performance and generalization in continual event sequence learning?
• RQ4 Can the proposed method achieve state-of-the-art performance while maintaining low memory and computational overhead in realistic deployment scenarios?
• RQ5 How effective is the asynchronous prompt update strategy in accelerating convergence without degrading predictive performance?

Key findings

• PromptTPP achieves state-of-the-art performance on three real-world user behavior datasets (Amazon, Taobao, and another unspecified dataset) under streaming, continual learning settings.
• The model increases total parameters by only 8–12% compared to the base TPP, with minimal impact on training speed, demonstrating high parameter efficiency.
• The asynchronous prompt update with $C=2$ improves convergence speed on the Amazon dataset while maintaining competitive performance, showing scalability and efficiency.
• Removing the key-value prompt design (w/o CtRroPP) leads to a notable performance drop, confirming the importance of structured prompt retrieval for knowledge retention.
• Replacing learnable keys with mean pooling (w/o k-v) causes a moderate performance decline, indicating that learnable keys significantly enhance retrieval accuracy and model stability.
• Increasing the prompt pool size $M$ consistently improves performance, confirming that larger prompt capacity enhances knowledge encoding and transfer.

Figure 2 : Overview of PromptTPP. Up: At training time, PromptTPP selects a subset of temporal prompts from a key-value paired CtRetroPromptPool based on our proposed retrieval mechanism; then it prepends the selected prompts to the event representations; finally it feeds the extended event represen