Proximity Matters: Analyzing the Role of Geographical Proximity in Shaping AI Research Collaborations

Technical Summary: Proximity Matters: Analyzing the Role of Geographical Proximity in Shaping AI Research Collaborations

Problem Statement
While the role of geographical proximity in facilitating inter-regional and inter-organizational collaborations is well-documented, its specific effect on forming scientific collaborations at the individual level remains under-addressed. Despite significant advancements in transportation and communication technologies, geography remains a primary determinant of collaboration, often impeding the formation of long-distance partnerships. This study investigates whether geographical distance continues to hinder individual-level scientific collaboration in the rapidly evolving field of Artificial Intelligence (AI) and examines whether network proximity can act as a substitute for geographical proximity to encourage long-distance collaborations.

Methodology
The study utilizes publication data from the AI field spanning 2001 to 2019, sourced from Elsevier's Scopus database. The dataset includes 45,738 publications (articles, conference papers, book chapters, and books) authored by 153,720 researchers across 162 countries. To analyze the data, the authors constructed four distinct scenarios to compare findings at different levels:

1. Canadian AI researchers only.
2. Researchers from Canada and the United States.
3. Researchers from Canada, the US, and Europe.
4. All AI researchers in the database.

Data Construction and Variables:

• Time Windows: To prevent simultaneity bias, the authors employed 2-year and 3-year sliding time windows. Independent variables were calculated based on a 3-year window, while the dependent variable (co-publication) was measured over a subsequent 2-year window.
• Dependent Variable: A binary variable indicating whether two authors had at least one co-publication during the 2-year window.
• Independent Variables:
  • Geographical Proximity: Measured as the "as the crow flies" distance (Haversine formula) between author affiliations.
  • Network Proximity: Estimated using the "Total Expected Number of Bridging Paths" (TENB), which quantifies indirect connections between authors.
  • Institutional Proximity: Binary variables indicating whether authors share the same province or country.
  • Cognitive Proximity: Calculated as the inverse of the correlation between authors' knowledge base vectors, derived using Latent Dirichlet Allocation (LDA) topic modeling on publication titles and abstracts.
  • Regional Contiguity: Binary variables capturing whether authors are located in contiguous provinces or countries.

Analytical Approach:
The study employed a dual-method approach:

1. Conventional Logistic Regression: To estimate coefficients and test hypotheses regarding the linear relationships between proximity dimensions and collaboration likelihood.
2. Machine Learning Classification: Several classifiers (Logistic Regression, Naive Bayes, Nearest Neighbors, SVM, Random Forest, and XGBoost) were trained to predict co-publication. The Extreme Gradient Boosting (XGBoost) model demonstrated the best performance. The SHapley Additive exPlanations (SHAP) technique was used to interpret feature importance and directionality.

Key Results

• Geographical Distance: Across all four scenarios, geographical distance exhibits a significant negative association with the likelihood of scientific collaboration. The logistic regression results confirm that increased distance impedes collaboration, even after controlling for other factors. Machine learning feature importance analysis (SHAP) reinforces this, showing that greater geographical distance correlates with lower predicted probabilities of collaboration.
• Network Proximity: Network proximity (TENB) shows a significant positive effect on collaboration. Authors with more bridging paths in the network are more likely to collaborate.
• Substitutability: A critical finding is the significant positive interaction between geographical distance and network proximity. This indicates that the effect of network proximity on collaboration likelihood increases as geographical distance increases. Consequently, network proximity acts as a substitute for geographical proximity; it is particularly vital for enabling collaborations between geographically distant researchers.
• Other Factors: Cognitive distance (dissimilarity in research topics) negatively impacts collaboration. Institutional proximity (sharing a country) also positively influences collaboration, with the effect being more pronounced in broader, multi-country scenarios.

Key Contributions
The paper makes four specific contributions to the literature:

1. Individual-Level Analysis: Unlike previous studies focusing on regions or organizations, this work analyzes geo-patterns at the individual researcher level, uncovering nuances often masked in aggregate data.
2. Multi-Scalability: The study employs a step-by-step expansion from a national (Canadian) to a global scope, providing a comprehensive view of how proximity dimensions function across different geographical scales.
3. Temporal Scope: By covering AI publications from 2001 to 2019, the study captures a wide range of recent developments in a field that has seen exponential growth since 2000.
4. Methodological Rigor: The integration of conventional econometric techniques with advanced machine learning classifiers and interpretability tools (SHAP) offers a robust validation of the relationships between proximity and collaboration.

Significance and Claims
The authors claim that their findings challenge the assumption that technological advancements in communication have diminished the role of physical distance. Instead, geography remains a strong barrier to individual-level collaboration. However, the study highlights a mechanism for overcoming this barrier: network proximity. The identification of a substitutability pattern suggests that strengthening network connections (e.g., through prior collaborations or shared intermediaries) can compensate for the lack of geographical proximity.

From a policy perspective, the authors suggest that supporting long-distance collaborations is not only beneficial for producing high-quality research but also serves a structural function: it increases the network proximity of the involved researchers, thereby triggering a positive feedback loop of future collaborations. This implies that policies aimed at fostering distant partnerships can effectively ease future knowledge flows and enhance the overall scientific ecosystem, particularly in the AI domain.

Limitations
The authors acknowledge that the study is limited to the AI field and that extending the analysis to other scientific domains is necessary to verify the generalizability of the substitutability pattern. Additionally, the study lacks individual-level control variables (e.g., age, gender), and the impact of the post-pandemic era on collaboration dynamics remains an open question for future research.