Khái niệm cốt lõi
The proposed model efficiently models the multivariate temporal point process by directly learning a flexible but well-defined cumulative hazard function, enabling accurate and efficient likelihood evaluation.
Tóm tắt
The paper presents a flexible and well-defined cumulative hazard function (CHF) network and incorporates it into a parameter-efficient multivariate temporal point process (MTPP) model.
Key highlights:
Most existing MTPP models define intensity (or density) functions for each event type, which can be prohibitively expensive when dealing with real-world event sequences with numerous event types.
The proposed model directly models the CHF and the time-dependent type distribution, significantly reducing the model complexity compared to existing methods.
The CHF network is designed to satisfy the mathematical constraints of a valid CHF, addressing the limitations of previous CHF-based methods.
Experimental results on six datasets show that the proposed model achieves state-of-the-art performance on data fitting and event prediction tasks, while having significantly fewer parameters and memory usage than strong competitors.
The paper first introduces the background on temporal point processes and neural point process models. It then presents the details of the proposed model architecture, including the CHF network, the time predictor, and the time-dependent type predictor. The training loss function is also described.
The experimental evaluation compares the proposed model against various state-of-the-art baselines on data fitting, event type prediction, and event time prediction tasks. The results demonstrate the effectiveness of the proposed approach, outperforming the baselines on most datasets. Additionally, the paper analyzes the model complexity in terms of required parameters and memory usage, showing significant reductions compared to the competitors.
The paper also includes ablation studies to investigate the impact of the hyperparameter α in the loss function and the choice of activation function in the CHF network.
Thống kê
The proposed model achieves the lowest negative log-likelihood (per sequence) on 5 out of 6 datasets compared to the baselines.
The proposed model has the lowest mean absolute error for event time prediction on most datasets.
Trích dẫn
"Most existing MTPP models define intensity (or density) functions for each event type, which can be prohibitively expensive when dealing with real-world event sequences with numerous event types."
"The proposed model directly models the CHF and the time-dependent type distribution, significantly reducing the model complexity compared to existing methods."
"Experimental results on six datasets show that the proposed model achieves state-of-the-art performance on data fitting and event prediction tasks, while having significantly fewer parameters and memory usage than strong competitors."