[2023-08] A model for interpreting the dynamics of indicators in scientific-scholarly papers using knowledge distillation and perturbation
운영자
날짜 2023.09.12
조회수 189
[2023-08] A model for interpreting the dynamics of indicators in scientific-scholarly papers using knowledge distillation and perturbation Ryu, Sungwook Moon Soul Graduate School of Future Strategy, KAIST Yi, Sangyoon Moon Soul Graduate School of Future Strategy, KAIST Abstract Developing techniques for predicting the impact of scientific-scholarly works in early-stage has recently received a lot of attention because of their role in knowledge evolution. Many novel ideas have emerged in different fields, including scientometrics, econometrics, and computer science, and these ideas have been investigated using traditional machine learning approaches (e.g. Abrishami & Aliakbary, 2019; Ma et al., 2021) and neural network approaches (e.g. Ruan et al, 2020; Weis & Jacobson, 2021). However, the explanatory power of AI models is limited by their nonlinear activation functions and parameters on hidden layers, and general models without neural networks exhibit poor performance in the early detection of significant features in terms of accuracy, precision, and recall scores. The trade-off between explanatory and predictive performance makes it challenging to apply scientometrics theories in practice and prevents the use of high-performing models to generate new theories. Here, we demonstrate a SOTA-level model with both predictive and explanatory power by direct control of a black box in artificial intelligence by perturbing a simple model that has learned up to negative information. For the purpose of validating the predictive power of the model, data from the Computer Security domain is used to compare its performance to that of existing models and to provide an illustration of its predictive power. An analysis of the features that can detect impact and their relationship is presented as an example. Direct control of models through knowledge distillation and perturbation may provide a new approach to scientometrics questions regarding how the various features of a scientific-scholarly communication system interact with one another and how these interactions affect the system's overall performance, with direct applications in the development of academic-scholarly evaluation systems. Keywords Impact measurement, Knowledge distillation, Perturbation, Scientometrics indicator
AWARD