ITPP: Learning Disentangled Event Dynamics in Marked Temporal Point Processes

TL;DR

ITPP tackles the entanglement problem in marked temporal point processes by enforcing channel-independent encoding of per-type event dynamics and introducing a type-aware inverted self-attention mechanism to explicitly model inter-type correlations. The model uses an ODE-based backbone with extrapolation and jump transitions for each event-type channel and decodes channel-specific intensities to form the joint likelihood, enabling robust learning and better generalization. Empirical results across six real and synthetic datasets show that ITPP achieves superior probabilistic fitting, time/mark prediction, and intensity recovery, with ablations confirming the critical roles of channel independence and inverted self-attention. The approach offers a principled, scalable pathway to disentangle complex multi-type event dynamics in continuous time, with strong practical impact for applications requiring precise event forecasting and uncertainty quantification.

Abstract

Marked Temporal Point Processes (MTPPs) provide a principled framework for modeling asynchronous event sequences by conditioning on the history of past events. However, most existing MTPP models rely on channel-mixing strategies that encode information from different event types into a single, fixed-size latent representation. This entanglement can obscure type-specific dynamics, leading to performance degradation and increased risk of overfitting. In this work, we introduce ITPP, a novel channel-independent architecture for MTPP modeling that decouples event type information using an encoder-decoder framework with an ODE-based backbone. Central to ITPP is a type-aware inverted self-attention mechanism, designed to explicitly model inter-channel correlations among heterogeneous event types. This architecture enhances effectiveness and robustness while reducing overfitting. Comprehensive experiments on multiple real-world and synthetic datasets demonstrate that ITPP consistently outperforms state-of-the-art MTPP models in both predictive accuracy and generalization.

Figures (9)

• Figure 1: An illustration of channel-mixing versus channel-independent architectures applied to user browsing history from the Taobao e-commerce dataset. The channel-independent approach avoids the pattern confusion inherent in channel-mixing.
• Figure 2: The general architecture of the proposed ITPP model. The framework has a channel-independent encoder-decoder architecture, with a type-aware inverted self-attention layer lying in the middle, which explicitly captures the latent correlations between different channels.
• Figure 3: The architecture of channel-independent context encoding. The encoder simulates two types of state transition, namely extrapolation and jump.
• Figure 4: The architecture of type-aware inverted self-attention. The property of type-awareness is achieved by channel-specific mapping functions.
• Figure 5: Visualisation of intensity recovery. The grey areas indicate the mass of recovery loss.