Discussion Graphs: Putting Social Media Analysis in Context

Emre Kiciman, Scott Counts, Michael Gamon, Munmun De Choudhury, Bo ThiessonICWSM 2014 · OpenReview

Introduction

Human behavior is deeply tied to context—gender, socioeconomic status, time, and environment. However, computational social media research often struggles to incorporate these "confounding factors" due to the high technical barrier of extracting and tracking these variables across millions of posts. Researchers frequently "average out" these factors, potentially leading to misleading conclusions. This paper introduces Discussion Graphs, a novel data model and accompanying tool (DGT) designed to bridge this gap by providing a standardized framework for contextual co-occurrence analysis.

The Discussion Graph Model

The core of this framework is the Discussion Graph, a formal hyper-graph model. In this model, nodes represent feature-values (such as a specific location, a mood, or a hashtag), while hyper-edges represent the context (e.g., a tweet or a user's visit history) connecting these nodes. Unlike traditional graphs, these edges are annotated with statistical summaries of the original context, such as gender, time, and sentiment.

The Discussion Graph model represents social media interactions as hyper-graphs where edges store rich metadata.

To manipulate these graphs, the authors developed the Discussion Graph Tool (DGT), which utilizes a domain-specific scripting language with five core commands:

  1. LOAD: Ingests the raw social media corpus.
  2. EXTRACT: Applies feature extractors for sentiment, gender, and location.
  3. RELATE BY: Defines the nature of relationships (e.g., co-occurrence in a user's history).
  4. PROJECT TO: Reduces the hyper-graph to a simpler graph of interest (e.g., Activity $\leftrightarrow$ Location) while automatically aggregating metadata into contextual statistics.
  5. OUTPUT TO: Exports data for visualization or further statistical analysis.

Datasets and Evaluation

The framework was evaluated using several large-scale datasets, including a one-month sample of the Twitter Firehose (219k location nodes) and 2.3 million geo-located tweets from New York City collected over three months. Reference data from Wikipedia and search query logs were used to ground locations and activities, providing a robust baseline for the contextual analysis.

Disambiguating Social Context

A key insight from this research is that context is a "first-class citizen." By baking metadata into the edges, researchers can "pivot" their analysis instantly. For instance, the system can distinguish between "pseudo-cliques"—groups of nodes that appear similar in a standard graph but differ significantly in context.

Comparison of demographics and mood distributions between 'New York Tourist' and 'Midtown Worker' cliques.

As shown in Figure 5, the DGT framework allows researchers to distinguish between a "Tourist" clique (characterized by high joviality and non-local users) and a "Midtown Worker" clique (characterized by higher fatigue and local residency), even when their physical locations overlap. This demonstrates that social boundaries, such as "neighborhoods," are not static physical entities but social constructs that shift based on the population and time of day.

Key Results: The Fluidity of Urban Boundaries

The authors performed a case study on Manhattan's neighborhood boundaries, finding significant shifts when conditioning on different factors. Using a pair-counting F-measure (where 1.0 indicates identity), they quantified the impact of various contexts on the resulting "map":

  • Weekend vs. Weekday: 0.47 (Highest impact)
  • Male vs. Female: 0.54
  • Day vs. Night: 0.67 (Least impact, but still significant)

These results confirm that temporal and demographic factors significantly reshape how urban spaces are utilized and perceived, a nuance often lost in aggregate social media studies.

Conclusion and Future Work

While DGT simplifies data processing and conditioning, it currently relies on external tools for high-level graph algorithms like community detection. Future work aims to integrate these methods directly into the pipeline and make the system publicly available to the social science community to spur broader research into contextualized human behavior.

Summary updated

BibTeX
@inproceedings{kiciman2014discussion,
  author = {Emre Kiciman and Scott Counts and Michael Gamon and Munmun De Choudhury and Bo Thiesson},
  title = {{Discussion Graphs: Putting Social Media Analysis in Context}},
  booktitle = {ICWSM},
  year = {2014},
  url = {https://openreview.net/forum?id=o7_vtEBK5J},
}
Figure