Abstract
Quantifying the evolution of scientific fields is a complex task often reliant on single metrics that may provide inconsistent results. This study proposes a robust framework that combines twelve different mathematical dissimilarity measures through Principal Component Analysis (PCA) to create a unified "Measure of Evolution." By analyzing over 274,000 bibliographic records from 1991 to 2019 across four distinct fields—Environmental Studies, Information Science, Medical Informatics, and Religion—the researchers identify a general trend of decelerating scientific evolution. This methodology provides a more robust and stable metric for science of science studies compared to traditional single-metric approaches.
Introduction
The "evolution" of a scientific field is typically tracked through shifts in keywords, citation patterns, or co-authorship networks. However, the choice of metric—such as Cosine similarity or Euclidean distance—can significantly influence the perceived rate of change. Different mathematical formulas often yield contradictory results, making it difficult to establish a stable "evolution map." To address this, we introduce a multi-metric approach that mitigates the quirks of individual statistical formulas by blending them into a single, robust proxy for evolutionary speed.
Methodology
The proposed pipeline for measuring scientific evolution consists of four primary stages:
- Keyword Distribution Extraction: Relative frequencies of author-provided keywords are calculated for each year within a specific field.
- Distribution Alignment: To facilitate year-over-year comparison, keyword lists are "updated" to a common set. Keywords present in one year but absent in another are assigned a frequency of zero. Alignment of keyword distributions between two years to ensure a consistent feature set. Combining dissimilarity measure for the study of evolution in scientific fields, page 7
- Calculation of Dissimilarity Measures: Twelve distinct dissimilarity measures are calculated for successive year pairs. These include standard metrics like Euclidean distance and Cosine similarity, as well as specialized measures like Canberra and Jensen-Shannon. The first six mathematical expressions used to calculate dissimilarity between keyword distributions. Combining dissimilarity measure for the study of evolution in scientific fields, page 7
- Principal Component Analysis (PCA): Given the high linear correlation (often >0.9) between these twelve measures, PCA is employed to extract the first principal component (PC1). PC1 accounts for over 80% of the total variance, serving as the final "evolutionary signal."
Data and Analysis
The study utilized 274,453 records from the Web of Science (1991–2019) across four fields:
- Environmental Studies (ES): 129,248 articles.
- Information Science & Library Science (ILS): 85,072 articles.
- Medical Informatics (MI): 58,358 articles.
- Religion (RE): 9,216 articles.
Results
Analysis revealed that the "speed" of evolution—defined as the dissimilarity between successive years—has generally decreased across all studied fields over the last thirty years. This suggests a shift toward more incremental research or a stronger dependence on immediate prior work.
Despite the general slowdown, field-specific "pulses" were observed:
- Religion (RE) surprisingly ranked first in evolution speed during the 1990s and 2000s, indicating a highly dynamic shift in its vocabulary and focus during that era.
- Medical Informatics (MI) took the lead between 2009 and 2016, becoming the field with the most rapid evolution as digital health and data science exploded.
- Environmental Science (ES) showed the most steady decrease in evolution speed, despite having the highest volume of publications and the fastest growth in the raw number of articles.
Discussion
The authors demonstrated that while these twelve dissimilarity measures operate on different scales, they are strongly linearly correlated. This justifies using PCA to extract a single "evolutionary signal" from the noise. The "slowing down" phenomenon observed across diverse disciplines raises important questions about the nature of modern scientific progress. It may indicate that research activities now have a stronger dependence on the immediate work of previous years or that the "low-hanging fruit" in these fields has already been picked.
Conclusion
By synthesizing multiple dissimilarity measures, this framework provides a more stable and holistic view of scientific change. Future research will focus on applying this methodology to "hard" sciences like Physics or Chemistry and investigating the underlying mathematical reasons for the high correlation between different dissimilarity metrics.
Summary updated