Topic detection and tracking on heterogeneous information

Long Chen, Huaizhi Zhang, Joemon M. Jose, Haitao Yu, Yashar Moshfeghi, Peter TriantafillouJ. Intell. Inf. Syst. 2018 · DOI · OpenReview

Abstract

Given the proliferation of social media and the abundance of news feeds, a substantial amount of real-time content is distributed through disparate sources, which makes it increasingly difficult to glean and distill useful information. Although combining heterogeneous sources for topic detection has gained attention from several research communities, most of them fail to consider the interaction among different sources and their intertwined temporal dynamics. To address this concern, we studied the dynamics of topics from heterogeneous sources by exploiting both their individual properties (including temporal features) and their inter-relationships. We first implemented a heterogeneous topic model that enables topic–topic correspondence between the sources by iteratively updating its topic–word distribution. To capture temporal dynamics, the topics are then correlated with a time-dependent function that can characterise its social response and popularity over time. We extensively evaluate the proposed approach and compare to the state-of-the-art techniques on heterogeneous collection. Experimental results demonstrate that our approach can significantly outperform the existing ones.

Introduction

In the modern information ecosystem, news often breaks on social media platforms like Twitter before being analyzed in depth by traditional news outlets. This creates a complex, intertwined stream of data where information flows across disparate sources. Existing Topic Detection and Tracking (TDT) systems often struggle with two primary challenges:

  1. Heterogeneity: Most models treat short-form social media posts and long-form news articles as identical objects, ignoring unique metadata such as authorship or social engagement metrics.
  2. Temporal Dynamics: Publication volume alone is a noisy indicator of importance. A spike in tweet volume may represent a delayed social reaction rather than the peak of the event itself.

This paper proposes the Heterogeneous Topic Model with Temporal Dynamics (HTMT) to address these issues by modeling source-specific characteristics and deconvolving social response from underlying topic importance.

Methodology

The Heterogeneous Topic Model (HTM)

The HTM framework avoids a "one-size-fits-all" approach by blending two distinct models. It utilizes the Author Topic Model (ATM) for Twitter to capture user-specific behaviors and standard Latent Dirichlet Allocation (LDA) for news articles. These models are linked via a shared "common-topic distribution" ($\eta$), which acts as a bridge for inter-source communication.

Graphical representation of the Heterogeneous Topic Model (HTM), showing separate plates for Twitter and News linked by a common distribution.

The specific parameters and variables used in this graphical model are defined in the following table:

Table of symbols and descriptions used in the HTMT framework.

Topic Deconvolution

To capture temporal dynamics, the authors treat the observed volume of a topic as a convolution of two functions: the topic importance (the actual event) and the social response (the public's reaction over time). By applying deconvolution, the model can filter out social "noise" to identify the original signal of a topic's significance. The model was fitted using collapsed Gibbs Sampling within an Expectation-Maximization (EM) framework.

Experimental Setup

The model was evaluated using two large-scale datasets from April 2015:

  • Twitter Dataset: 30 million tweets from 786,000 users, filtered for Scotland-related content.
  • News Dataset: 224,272 unique URLs from mainstream media and blogs, processed via Boilerpipe.

Key Results

The HTMT model was compared against several baselines, including Source-LDA. Key findings include:

  • Predictive Performance: HTMT achieved significantly lower perplexity scores. For instance, on the Twitter dataset with 300 topics, HTMT scored approximately 4136, compared to 7269 for Source-LDA.
  • Topic Clarity: The model demonstrated lower Topic Entropy, indicating that the discovered topics were more distinct and less prone to overlap.
  • Qualitative Accuracy: Case studies on local events (e.g., #Scotrail) showed that HTMT produced smoother importance curves that better reflected real-world event timelines compared to models that overfit to raw tweet spikes.

Conclusion and Future Work

The HTMT framework successfully demonstrates that preserving the unique "soul" of different data sources—modeling people-driven social media differently than content-driven news—leads to superior topic tracking. Future research directions include integrating sentiment analysis to track emotional shifts, utilizing entity linking to structured knowledge bases for deeper context, and optimizing the computational complexity for real-time global streams.

Summary updated

Abstract

Given the proliferation of social media and the abundance of news feeds, a substantial amount of real-time content is distributed through disparate sources, which makes it increasingly difficult to glean and distill useful information. Although combining heterogeneous sources for topic detection has gained attention from several research communities, most of them fail to consider the interaction among different sources and their intertwined temporal dynamics. To address this concern, we studied the dynamics of topics from heterogeneous sources by exploiting both their individual properties (including temporal features) and their inter-relationships. We first implemented a heterogeneous topic model that enables topic–topic correspondence between the sources by iteratively updating its topic–word distribution. To capture temporal dynamics, the topics are then correlated with a time-dependent function that can characterise its social response and popularity over time. We extensively evaluate the proposed approach and compare to the state-of-the-art techniques on heterogeneous collection. Experimental results demonstrate that our approach can significantly outperform the existing ones.
BibTeX
@article{chen2018topic,
  author = {Long Chen and Huaizhi Zhang and Joemon M. Jose and Haitao Yu and Yashar Moshfeghi and Peter Triantafillou},
  title = {{Topic detection and tracking on heterogeneous information}},
  journal = {J. Intell. Inf. Syst. 2018},
  year = {2018},
  doi = {10.1007/s10844-017-0487-y},
}
Figure