Prompt-augmented Temporal Point Process for Streaming Event Sequence

Authors: Siqiao Xue, Yan Wang, Zhixuan Chu, Xiaoming Shi, Caigao JIANG, Hongyan Hao, Gangwei Jiang, Xiaoyun Feng, James Zhang, Jun Zhou

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present a novel and realistic experimental setup for modeling event streams, where Prompt TPP consistently achieves state-of-the-art performance across three real user behavior datasets.
Researcher Affiliation Industry Ant Group Hangzhou, China {siqiao.xsq,luli.wy,chuzhixuan.czx}@alibaba-inc.com
Pseudocode Yes Algorithm 1 Prompt TPP at training time of the T -th task. Algorithm 2 Prompt TPP at test time of the T -th task.
Open Source Code Yes Our code is available at https://github.com/yanyan Sann/Prompt TPP
Open Datasets Yes We conduct our real-world experiments on three sequential userbehavior datasets. In each dataset, a sequence is defined as the records pertaining to a single individual. The Taobao (Alibaba, 2018) dataset... The Amazon (Ni, 2018) dataset... The Stack Overflow (Leskovec & Krevl, 2014) dataset...
Dataset Splits Yes The subset in each task is split into training, validation, and test sets with a 70%, 10%, 20% ratio by chronological order.
Hardware Specification Yes All the experiments were conducted on a server with 256G RAM, a 64 logical cores CPU (Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz) and one NVIDIA Tesla P100 GPU for acceleration.
Software Dependencies No All models are implemented using the Py Torch framework (Paszke et al., 2017). For O-TPP, as the authors Yang et al. (2017) have not published the code, we implement it using the tick (Bacry et al., 2017) library. All models are optimized using Adam (Kingma & Ba, 2015).
Experiment Setup Yes For NHP, the main hyperparameters to tune are the hidden dimension D... usually 32, 64, 128... For Att NHP, another important hyperparameter to tune is the number of layers L of the attention structure... usually 1, 2, 3, 4... we set α = 0.1 consistently for both datasets. For the prompts, we set M = 10, N = 4, Lp = 10 for both datasets. For the asynchronous training parameter C, we choose C = 2 for Taobao and Amazon datasets by default. Table 3 contains descriptions that list all of the hyperparameters set throughout our experiments.