Learning Time Series Associated Event Sequences With Recurrent Point Process Networks.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Real-world sequential data, like event sequences and associated time series, are common but challenging to model.
• Traditional parametric temporal point processes often fail to incorporate concurrent time series data and require task-specific intensity functions.

Purpose of the Study:

• To propose a novel model, recurrent point process networks (RPPNs), that effectively models event sequences and associated time series.
• To improve the accuracy and interpretability of temporal event prediction and relational network mining.

Main Methods:

• Instantiating temporal point process models with temporal recurrent neural networks (RNNs).
• Utilizing two RNNs to model intensity functions: one for event relationships, another for time series updates.
• Incorporating an attention mechanism to uncover interpretable influence strengths among events.

Main Results:

• Demonstrated the superiority of RPPNs on both synthetic and real-world datasets.
• Achieved improved performance in temporal event prediction and relational network mining tasks.
• The attention mechanism provided interpretable insights into event influences.

Conclusions:

• Recurrent point process networks offer a powerful and flexible framework for modeling complex sequential data with associated time series.
• The proposed model overcomes limitations of traditional methods by integrating multiple data sources and providing interpretable results.
• RPPNs advance the state-of-the-art in analyzing event sequences and uncovering underlying network structures.