Residual TPP: A Unified Lightweight Approach for Event Stream Data Analysis

Authors: Ruoxin Yuan, Guanhua Fang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate that Residual TPP consistently achieves state-of-the-art goodness-of-fit and prediction performance in multiple domains and offers significant computational advantages as well.
Researcher Affiliation Academia 1School of Management, Fudan University, Shanghai, China. Correspondence to: Guanhua Fang <EMAIL>.
Pseudocode Yes Algorithm 1 The Overall Pseudocode of Residual TPP
Open Source Code Yes Our code is publicly available at https://github.com/ruoxinyuan/Residual TPP.
Open Datasets Yes Datasets. We evaluate our method on six real-world benchmark datasets: MIMIC-II (Johnson et al., 2016), Retweet (Zhou et al., 2013), Earthquake (Xue et al., 2024), Stack Overflow (Leskovec & Krevl, 2014), Amazon (Ni, 2018) and Volcano (Xue et al., 2024). The MIMIC-II dataset is available at the public Git Hub repository1, and the others can be accessed via the Easy TPP library2. 1https://github.com/hongyuanmei/neurawkes 2https://github.com/ant-research/ Easy Temporal Point Process
Dataset Splits Yes Each dataset is split into training, validation, and test sets. The validation set is for hyperparameter tuning, and the test set is for model evaluation. Table 7. Statistics of the real-world datasets. The table columns, from left to right, represent the dataset, number of event types, number of events, sequence length, and number of sequences. DATASET # TYPES # EVENTS SEQUENCE LENGTH # SEQUENCES MIN MEAN MAX TRAIN VALID TEST MIMIC-II 75 2,419 2 4 33 527 58 65 Table 8. Statistics of the synthetic datasets. ... DATASET # TYPES # EVENTS # RESIDUALS SEQUENCE LENGTH # SEQUENCES MIN MEAN MAX TRAIN VALID TEST POSSION-BASED 5 33,243 4,253 20 29 100 600 200 200
Hardware Specification No All training was conducted on a CPU.
Software Dependencies No All experiments in this study were implemented using Py Torch (Paszke et al., 2019), trained using the Adam (Kingma, 2014) optimizer, and executed on a CPU. In our experiments, we implement this using Tick3 (Bacry et al., 2017).
Experiment Setup No For simplicity, we model the event stream data using a Hawkes process with a fixed baseline intensity and an exponential decay function to capture the self-excitation property: λ(1) k (t) = µk + X i:ti<t αk,kiβk,kie βk,ki(t ti). All experiments in this study were implemented using Py Torch (Paszke et al., 2019), trained using the Adam (Kingma, 2014) optimizer, and executed on a CPU. In all experiments presented in this paper, we set ρ1 = ρ2 = 1. The parameters a and b are adjusted based on the characteristics of the data from different datasets to ensure that the proportion of residual events remains within an appropriate range.