Hierarchical Dirichlet Gaussian Marked Hawkes Process for Narrative Reconstruction in Continuous Time Domain

Abstract

In news and discussions, many articles and posts are provided without their related previous articles or posts. Hence, it is difficult to understand the context from which the articles and posts have occurred. In this paper, we propose the Hierarchical Dirichlet Gaussian Marked Hawkes process (HD-GMHP) for reconstructing the narratives and thread structures of news articles and discussion posts. HD-GMHP unifies three modeling strategies in previous research: temporal characteristics, triggering event relations, and meta information of text in news articles and discussion threads. To show the effectiveness of the model, we perform experiments in narrative reconstruction and thread reconstruction with real world datasets: articles from the New York Times and a corpus of Wikipedia conversations. The experimental results show that HD-GMHP outperforms the baselines of LDA, HDP, and HDHP for both tasks.

Mitigating Performance Saturation in Neural Marked Point Processes

“…Wang et al [28] modulates the intensity function by an additional nonlinear link function, in order to capture the nonlinear effects. Another major development of marked Hawkes processes is Bayesian Hawkes processes [13,25,33]. These models are usually fused with mixture models, especially in the context of natural language processing.…”

“…Both models assume a Dirichlet prior distribution for the parameters, and therefore they are applicable to clustering tasks. More recently, some works [13,25] propose hierarchical Bayesian Hawkes processes to deal with continuous features associated with events. One common drawback in Bayesian Hawkes processes is the poor scalability.…”

“…An attributed event sequence contains not only the timestamps of asynchronously generated events but also event features/attributes. 2 It is naturally generated from databases and event logfiles, and has been applied to various application scenarios and disciplines including financial transactions [2], natural language processing [25], and spatial dependence among trees [22], etc. Existing methods that deal with attributed event sequences are usually based on the statistical toolmarked point processes, which have been used for recommendation [8], network inference [15], fake news mitigation [9], and many other tasks [23,27,33].…”

Tweedie-Hawkes Processes: Interpreting the Phenomena of Outbreaks

“…The other category is more related to our model, where the generation process of features is involved in the model. Typical models also belong to marked point processes, such as (Simma and Jordan 2010) and (Seonwoo, Oh, and Park 2018), or mixture models combined Hawkes process and the generation process of features, such as (Yang and Zha 2013), (He et al 2015) and (Wang et al 2017). These models always assume that all the events share the same distribution of α (or even same α), which is somehow unrealistic.…”

“…Text 1 is a root node. Texts 2, 3 and 4 are descendant of Text 1. and (Seonwoo, Oh, and Park 2018). Due to the omnipresence of the Tweedie distribution, THP greatly enhances Hawkes processes when dealing with text-based cascades in social networks, especially in the following aspects:…”