TL;DR
This paper introduces a novel method for analyzing the historical evolution of Artificial Intelligence (AI) conferences by quantifying their "closeness" based on author co-publication patterns. By applying Principal Component Analysis (PCA) and clustering to this closeness metric, the authors reveal how AI sub-communities have formed, evolved, and interacted with other computer science fields over time.
What Problem Does This Paper Address?
Defining the boundaries of scientific domains, especially a broad and interdisciplinary field like Artificial Intelligence, is notoriously difficult. It's even harder to track the historical evolution of these communities and understand how different AI sub-fields relate to each other and to broader computer science. Traditional methods of analyzing conference content, like session titles, are often subjective and labor-intensive. This paper aims to provide an objective, data-driven approach to map the landscape of AI conferences and their relationships, offering insights into the structure and dynamics of the AI research community.
Methods
The core of this work lies in the coscinus tool, which quantifies the "closeness" between publication venues (conferences and journals) based on author co-publication habits. The fundamental hypothesis is that researchers tend to specialize, so if many authors publish in two different venues, those venues are likely thematically connected.
Here's a breakdown of their methodology:
- Data Source: The DBLP (Digital Bibliography & Library Project) database, a comprehensive repository of computer science publications, is used as the primary data source.
- Closeness Metric:
- A matrix of author publication counts per venue is constructed.
- A "closeness matrix" is built by counting authors who have published at least
npapers in two different venues. Thisnis called the "Minimal Count" (MC) threshold, used to filter out noise from occasional publications outside a researcher's core area. - This raw count matrix is then normalized to account for differences in publication volume between venues. Three normalization methods are discussed (min, max, avg), with the paper focusing on the "min" normalization. The resulting matrix represents the proportion of authors of one venue who also published in another.
- Dimensionality Reduction & Clustering:
- Principal Component Analysis (PCA): Used to project the high-dimensional closeness matrix (where each dimension represents another conference) into a 2-dimensional space, allowing for visualization of conference relationships.
- K-means Clustering: Applied to group similar conferences together, identifying distinct sub-communities within AI and computer science.
- Temporal Analysis: A novel temporal window selection feature is added to
coscinusto study the historical evolution of conference closeness, allowing researchers to examine relationships over specific time periods (e.g., 5-year windows).
Datasets
- DBLP Database: The primary dataset used, containing millions of computer science publications. The paper specifies using the May 22, 2018 version.
- CORE Ranking: Used to classify publication venues into categories (A*, A, B, C) to focus the analysis on reputable conferences.
The authors also provide links to their adapted coscinus tool and the specific data used for reproducibility.
Key Insights and Ideas
- Objective Community Mapping:
coscinusoffers a data-driven, objective way to understand the structure of research communities, moving beyond subjective interpretations of conference themes. - AI as a Distinct, Yet Interconnected, Field: The analysis reveals a clear "AI region" in the PCA plots, distinct from other computer science areas like "Database + Web." However, AI conferences also show connections to fields like data mining, web, and vision, reflecting the pervasive nature of AI.
- Evolution of AI Sub-communities: The temporal analysis highlights how AI sub-communities have coalesced and shifted over decades. For instance, early AI conferences show connections to theoretical computer science and cryptography, while later periods see stronger links to data mining and web technologies.
- "Machine Learning" as a Stable Cluster: A consistent "machine learning cluster" emerges, comprising conferences like NIPS, ICML, and COLT, indicating a strong and enduring sub-community.
- Articulation Points: Some conferences are found to be at the "articulation" between different sub-communities (e.g., between automated reasoning and core AI groups), suggesting they bridge different research areas.
Key Results
- Visualization of AI Landscape: PCA plots effectively group AI conferences, demonstrating the method's ability to reveal community structures. Figure 1 shows A* conferences clustered, with a clear AI group and a machine learning group.
- Temporal Trends Identified: Table 3 and Table 4 illustrate how conference clusters change over time. For example, in earlier years (1980s), AI conferences were clustered with theoretical computer science and logic, while by 2015, they are more closely associated with data mining and web conferences.
- Stable AI Clusters: Core AI conferences like IJCAI, AAAI, ECAI, KR, and AAMAS consistently appear together in clusters across different time periods and clustering granularities.
Limitations and Future Work
- Noise from Anomalies: While the MC threshold helps, occasional publications outside a researcher's core expertise can still introduce noise. The authors acknowledge that using smaller MC values for temporal analysis (e.g., MC=1 for 5-year windows) might increase this noise.
- DBLP Coverage: The analysis is limited by the completeness and accuracy of the DBLP database.
- Subjectivity in CORE Ranking: The classification of conferences by CORE, while useful, is still a form of external judgment.
- Future Work: The authors suggest that this tool can be combined with other observations (e.g., qualitative analysis of paper topics) for a richer understanding. They also highlight the potential for studying the evolution of other computer science fields and for identifying thematic changes that drive community shifts. The
coscinus.org/aiwebsite is mentioned as a resource for further exploration.
Summary updated