| Predicting publication productivity of researchers is a basic task for academic managers and funding agencies to evaluate the capabilities in scientific research.However,it remains an elusive task due to the influence of many random factors on individual behaviour in publishing and the diversity of different researchers’ production capacity.Based on integrating a shallow architecture with a deep one,this paper conducts a research on predicting the number of published papers by researchers.In this research,we proposed a piecewise Poisson model,and constructed a prediction model wihch integrates a shallow Poisson model and a deep LSTM architecture.Here,researchers refer to those who have published at least one academic paper in scientific research,and research groups refer to groups of researchers with the same publication productivity.The data set used here is DBLP computer science bibliographic data set,which covers the literature information of most journals and conference records in the field of computer science.DBLP dataset spans a long time and updates real time information with a high effectiveness,and is suitable for extracting the quantitative information of papers and that of their authors.In this research,by applying DBLP data set,we completed the following tasks:(1)Given a specific subset of researchers,analyze the distribution of the number of researchers.The K-S test shows that it follows a Poisson distribution.Based on this result,we proposed a piecewise Poisson model,which belongs to shallow architectures due to the limited parameters of linear regression.The model is applied to the DBLP data set to verify its effectiveness.The results show that the piecewise Poisson model can characterize the distribution of the number of low-productivity researchers,but it can only predict the average number of papers published by a group of researchers with similar productivity and the publication events of low-productivity researchers in the next year with high accuracy.(2)In view of the LSTM neural network can solve the problem of long term dependence,and a strong correlation between the number of papers published in the future and those published in the past,we build a deep LSTM neural network to predict the number of papers by researchers.A time series,comprising the number of papers published by researchers in the past,is input into the neural network,and then a predictive value can be obtained at the output of the neural network.Compared with the piecewise Poisson model,the LSTM neural network has many layers containing large number of parameters,so it belongs to deep architectures.The LSTM neural network is applied to DBLP data set to verify its performance.The results show that it has a high prediction accuracy only on publication events in the next year of researchers,but cannot characterize the distribution of the number of researchers.(3)On the one hand,although the piecewise Poisson model can characterize the distribution of the number of low productivity researchers,its prediction accuracy needs to be improved.On the other hand,the LSTM neural network can capture the long term dependence,and has a high prediction accuracy on publication events of researchers.In view of the above two aspects,we combined the advantages of the piecewise Poisson model and the LSTM neural network,to construct a prediction model which integrates a shallow Poisson model and a deep LSTM architecture.The validity of this model was verified by applying it to DBLP data set.Compared with the prediction results of the piecewise Poisson model and the LSTM neural network respectively,this model can better characterize the distribution of researchers on the data set,and its accuracy on predicting the publication events can match that of the LSTM neural network.However,this model can only predict the average number of papers published by a group of researchers with similar productivity,but cannot accurately predict the number of papers published by a single researcher. |
PDF Full Text Request