Comparison of the predictive ability of frequency, degree and PAFit when measuring keyword popularity

To perform our experiment, we should answer a vital question in the first place: What is popularity? In the dictionary, popularity is “the quality or state of being popular” (“Popularity.” Merriam-Webster.com), while the definitions of popular include “of or relating to the general public” and “frequently encountered or widely accepted” (“Popular.” Merriam-Webster.com). Therefore, a popular keyword should be related to large number of other keywords and occurring frequently in the ecological journals. These two characters could be well represented by degree and frequency mentioned in the former section.

Popularity of keywords should not only be descriptive but also predictive. In other words, when we say a keyword is popular, it has been popular for some time, and this trend will not disappear in the near future. For instance, if we gain the popularity of keywords in a specific time period, we might be able to predict the growth of the keywords in the following years. Therefore, we split our data into two parts, and tried to use the historical keyword popularity to predict the growth of keywords’ frequency and degree in the coming three years. The experiment procedure was designed as follows: 1. Construct the ecological knowledge network with data from 1988 to 2014, and calculate the frequency, degree and PAFit for every keyword appeared in these 27 years; 2. Construct the ecological knowledge network with data from 1988 to 2017, calculate the frequency and degree for every keyword appeared in the total 30 years; 3. Subtract the frequency of 27 years from frequency of 30 years, and we gain the change (or growth) of frequency in the recent three years (namely 2015–2017). The same is done to the keywords’ degree. Note that keywords emerging in the recent three years but not in the previous 27 years would be excluded from our analysis; 4. Fit a simple linear regression model using frequency, degree and PAFit in the former 27 years to predict the growth of frequency and degree in the following 3 years respectively. Compare the results and see if PAFit yields better predictions.