TL;DR
This paper introduces a novel Topical Link Model (TLM) for discovering communities in "textual interaction graphs," where interactions are documented by text. TLM leverages Hierarchical Dirichlet Processes (HDP) to model hidden topical structures within these documents, allowing for more nuanced community detection than traditional link-based methods.
What Problem Does This Paper Address?
Traditional community discovery methods often rely solely on the structure of interaction graphs (who interacted with whom). However, in many real-world scenarios, these interactions are accompanied by textual data (e.g., emails, forum posts, research papers) that contain rich semantic information. Existing methods that incorporate text often treat links as simple weighted connections, failing to fully exploit the topical content of the documents that define these interactions. This paper addresses the challenge of discovering communities by not only considering the interaction structure but also by deeply analyzing the topical content of the documents that link entities, aiming to uncover more meaningful and coherent communities.
Methods
The core contribution is the Topical Link Model (TLM). TLM models a "textual interaction graph" by combining two key components:
- Participation Graph Modeling: This part models how users participate in documents. It uses a Hierarchical Dirichlet Process (HDP) to capture the idea that users might engage with documents related to different underlying topics. This is visualized as a graphical model where users are linked to documents they participate in.
- Document Modeling: This component models the content of the documents themselves. It uses another HDP to infer latent topics within each document. The model assumes that words in a document are generated from a mixture of these latent topics.
The TLM unifies these two aspects by treating both the user-document participation and the document-topic generation within a single HDP framework. This allows the model to learn the number of communities automatically (nonparametric) and to infer topical variables associated with the links (documents). The inference is performed using Gibbs sampling.
Datasets
The authors evaluate their model on two datasets:
- PAPER: A corpus of abstracts from 7 research conferences (ACL, ICML, SIGGRAPH, SIGIR, SIGKDD, SIGMOD, WWW) from 2005-2009. Each conference is treated as a community, and the proportion of papers an author published in each conference serves as the "ideal" community distribution.
- NYT: A dataset of news articles from the New York Times that mention at least 3 companies.
Key Insights and Ideas
- Textual context enriches community discovery: Simply looking at who interacted with whom is insufficient. The content of the interactions (documents) provides crucial semantic information for defining more accurate communities.
- Hierarchical Dirichlet Processes are powerful for this task: HDPs are well-suited for modeling hierarchical topic structures and for automatically determining the number of latent topics (and thus communities) without prior specification.
- TLM can uncover semantically coherent communities: By modeling topics within links, TLM can identify communities that are not just based on co-occurrence but on shared topical interests reflected in the documents.
- Nonparametric nature is a significant advantage: The ability to learn the number of communities from the data is a key benefit, avoiding the need for manual tuning of this parameter.
Key Results
- Improved performance over baselines: TLM consistently outperforms two state-of-the-art baseline models (NCut and NetPLSA) on the PAPER dataset, as measured by the Categorical Clustering Distance (CCD). The best CCD for TLM was 1219.65, a significant improvement over NetPLSA's best of 1408.9.
- Semantic interpretability: On the PAPER dataset, TLM successfully identified 6 major communities that correspond well to established research fields (information retrieval, computer graphics, data mining, databases, computational linguistics, machine learning). The top terms extracted by TLM for each community further support this semantic coherence.
- Parameter sensitivity: The paper shows that TLM's performance is sensitive to its hyperparameters (α, αg, β), and optimal values were found through experimentation.
Limitations and Future Work
- Parameter tuning: While TLM is nonparametric in terms of community number, its performance is still influenced by the choice of hyperparameters for the Dirichlet Processes.
- Future directions: The authors suggest exploring other document modeling techniques (e.g., LDA) and applying TLM to larger-scale datasets. They also plan to investigate other document modeling approaches.
Summary updated