Predicting publication productivity for authors: Shallow or deep architecture?

TL;DR

This paper compares deep neural networks and model-based methods for predicting research publication productivity, proposing a hybrid model that outperforms individual approaches on a large dataset.

Contribution

It introduces a hybrid prediction model combining data-driven and model-based methods for better accuracy in research productivity forecasting.

Findings

Neural networks struggle with long-term group predictions.

Model-based approaches are limited for short-term individual predictions.

The hybrid model outperforms standalone methods on dblp dataset.

Abstract

Academic administrators and funding agencies must predict the publication productivity of research groups and individuals to assess authors' abilities. However, such prediction remains an elusive task due to the randomness of individual research and the diversity of authors' productivity patterns. We applied two kinds of approaches to this prediction task: deep neural network learning and model-based approaches. We found that a neural network cannot give a good long-term prediction for groups, while the model-based approaches cannot provide short-term predictions for individuals. We proposed a model that integrates the advantages of both data-driven and model-based approaches, and the effectiveness of this method was validated by applying it to a high-quality dblp dataset, demonstrating that the proposed model outperforms the tested data-driven and model-based approaches.