Application Of BP Network Model For Predicting Factors Affecting The First Fund Research Performance

[Objective](1) We would analyze Factors Influencing Research Performance by establishing BP neural network model.(2) Explore the suitability and feasibility of BP neural network.[Methods](1) Using correlation and variance analysis to explore the relationship between total output and performance factors, the subjects are projects for three consecutive years from2003to2007, the purpose is to initially select factors of overall performance.(2) Using stratified random sampling method in which project type as a stratification factor to divide the total data set, we extracted75%of the data as the training set, the remaining25%of the data as the test set.(3) We would establish BP neural network model and multiple linear regression model on total output and performance factors using the training set, then we would compare prediction error of two models using the test set.(4) Using BP neural network model to put factors in order, the subjects are projects for three consecutive years from2003to2007, and the purpose is to analyze the importance of factors included in the model, then we would explore the relationship between total output and performance factors with charts.(5)Statistical software using internationally accepted software SAS9.13; BP neural network analysis was performed using SPSS Clementine software.[Results](1)We chose factors including characteristics of projects and the project leaders as influencing factors initial framework, Project performance output comprised the published papers, published works, subsequent commitment issues, patented, personnel training, achievement awards and new technologies, new methods, new regulations etc. Through statistical description of research performance, we found the performance of Capital Medical Development Research Fund presented a skewed distribution, research and performance output is not high, the highest score is published papers in seven outputs, followed by the training of personnel, the score of patents is the lowest.(2) We used univariate analysis and correlation analysis to screen the influencing factors, the factors which have significant differences could be retained.(3) Using stratified random sampling method, we extracted75%of the data as the training set, the remaining25%of the data as the test set, and then we tested the distribution of factors in two data sets, the results showed that factors in datasets are balanced.(4) We established BP neural network model and multiple linear regression model using the training set, then we compared prediction error of two models using the test set. The results showed that prediction error of multiple linear regression model is73.38%, and prediction error of BP neural network model is87.91%, in the same conditions, prediction accuracy of neural network model improved as much as10percentage points more than the multiple linear regression model.(5) We chose factors based on predict changes of neural network, factors should be retained in the model when prediction accuracy was no longer increasing after removing certain factors, when process of screening factors was terminated. Factors ultimately retained in the model are project type、tutor、nature of units、unit level、 degree、titles、whether it is the first time to get research project、whether it is the highest level, the importance of the variables are further sorted by using neural network, the results showed that the greatest impact on the research project was project type, followed by tutor、whether it is the highest level、degree、nature of units、unit level、titles, minimal impact on research performance factors was whether it is the first time to get research project.(6) We further researched the relationship between factors and research performance, we found joint research showed the highest scientific performance, followed by the key support, The lowest scientific performance was Independent innovation Project. For factors of tutor、degree、 nature of units、unit level、titles, research Performance increased with the level of upgrade; For factors of whether it is the first time to get research project and whether it is the highest level, Research performance of the new was lower then the elder.[Conclusion] Using the neural network model to analyze the performance of scientific research have more advantages, the neural network model can be used into complex and nonlinear problems; According to the relationship between various factors and research performance of capital medical development research fund, we proposed the following recommendations to improve research performance:(1) Fund management should effectively identify the characteristics of the applicant’s academic, and projects can be established depending on the circumstances.(2) Strengthening the development of talent is in order to improve the level of people who are in charge of projects.(3) Strengthening management of work unit is in order to Improving conditions for research support.(4) Fund management should establish a database of talent, and increase efforts to support new.