A Density-based Nonparametric Model for Online Event Discovery from the Social Media Data

Jinjin Guo, Zhiguo GongIJCAI 2017 · DOI

TL;DR

This paper introduces DP-density, a novel nonparametric model for discovering events in real-time from social media data. It uniquely combines density estimation for temporal dynamics with Dirichlet Process for flexible event discovery and Gaussian mixtures for spatial patterns, offering a parameter-free approach to understanding "what, when, and where" events unfold online.

What Problem Does This Paper Address?

Social media platforms like Twitter and Flickr generate a constant stream of geo-tagged data, which is rich with information about real-world events. However, existing event discovery methods often rely on static datasets, limiting their applicability to dynamic, online streams. Furthermore, many online event discovery techniques struggle with determining the optimal number of events (topics) or defining their temporal evolution without manual tuning. This paper addresses these limitations by proposing a model that can automatically infer the number of events and dynamically capture their temporal and spatial characteristics from streaming social media data.

Methods

The core of the DP-density model lies in its integration of several key techniques:

  • Dirichlet Process (DP): Used to handle the nonparametric nature of event discovery. This allows the model to automatically infer the number of events (topics) as new documents arrive, rather than requiring a predefined number. It models the "rich get richer" phenomenon where popular events tend to attract more attention.
  • Density Estimation with Gaussian Kernels: To capture the temporal dynamics of events, the model estimates a density function over time. When a new document arrives, its temporal influence is modeled by summing Gaussian kernels centered around the timestamps of previous documents within a sliding window. This allows the model to understand how event popularity evolves over time.
  • Sliding Window and Dynamic Bandwidth: To manage the computational burden of processing an ever-growing stream, a sliding window approach is used. This focuses the density estimation on recent documents. Additionally, the bandwidth of the Gaussian kernels is dynamically adjusted to maintain a consistent influence range as the stream progresses.
  • Mixture of Gaussians: To incorporate the spatial aspect of events, a mixture of Gaussians is employed. This models how events are distributed across geographical regions, allowing for events that occur in one or multiple locations.
  • Sequential Monte Carlo (SMC): This is the inference algorithm used to efficiently process the streaming data. SMC allows for online computation of the posterior distribution by maintaining a set of weighted particles, each representing a hypothesis about the latent variables (events and regions).

Datasets

The paper evaluates the proposed DP-density model on two real-world datasets:

  • Paris Dataset: Crawled from Flickr, containing over 26,000 photos with textual descriptions from June to August 2010.
  • New York (NY) Dataset: Also crawled from Flickr, with over 40,000 photos and textual descriptions from January to March 2010.

These datasets were preprocessed by splitting Flickr tags into single words and removing irrelevant terms.

Key Insights and Ideas

  • Parameter-Free Online Event Discovery: The most significant insight is the development of a truly parameter-free model for online event discovery that jointly considers temporal and spatial dimensions. This is a major step towards more automated and adaptable event detection systems.
  • Temporal Dynamics are Crucial: The paper highlights that understanding the temporal evolution of event popularity (e.g., how an event gains and loses traction) is key to accurately discovering and characterizing it. The density estimation approach effectively captures this.
  • Geospatial Context Enhances Event Definition: Incorporating location information alongside temporal and textual data helps to better distinguish and define events, as real-world events are inherently tied to specific places.
  • Dirichlet Process for Flexible Event Number: The use of DP is a clever way to avoid pre-specifying the number of events, which is often a difficult and arbitrary choice in online settings. The model naturally adapts to the complexity of the data.

Key Results

  • Effective Event Discovery: The DP-density model successfully extracts interpretable events from both datasets, demonstrating high coherence and low perplexity.
  • Competitive Performance: Compared to baseline methods like RCRP-geo and DCT-geo, DP-density shows comparable or superior performance in terms of perplexity and coherence, especially when compared to methods that require manual epoch settings.
  • Automatic Event Number Inference: The model consistently infers a reasonable number of events, which aligns with the performance of a fixed-event-number variant (DP-density-fixed) when that number is optimized.
  • Efficient Processing: The SMC inference algorithm, combined with the sliding window, allows for efficient online processing of the document stream, with processing time stabilizing after an initial build-up period.

Limitations and Future Work

  • Computational Cost: While SMC is efficient for online processing, the paper acknowledges that the overall computational cost can still be significant, especially with very large datasets or long temporal windows.
  • Sliding Window Size: The choice of the sliding window size (L) is a parameter that influences performance, though the paper suggests it can be set based on the decay of the Gaussian kernel influence.
  • Future Work: The authors suggest exploring more sophisticated density estimation techniques and potentially extending the model to handle more complex event structures or relationships between events. Further investigation into optimizing the dynamic bandwidth adjustment could also be beneficial.

Summary updated

BibTeX
@inproceedings{guo2017density,
  author = {Jinjin Guo and Zhiguo Gong},
  title = {{A Density-based Nonparametric Model for Online Event Discovery from the Social Media Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/240},
}
Figure