Expert recommendation algorithm for university procurement review based on user portraits
[Objective]With the advancement of"government function reform"in the scientific research field,central universities and research institutes can now select experts for government procurement projects involving complex technologies.Currently,two expert selection methods are mainly used:self-selection and database recommendation.The self-selection method enhances the autonomy of university procurement but also brings certain integrity risks.The database recommendation method relies on the precision of the professional settings in the review expert database.It may face issues such as reduced recommendation efficiency or mismatches in expertise.However,it effectively reduces the risk of corruption.[Methods]Considering the correlation and continuity of expertise among self-selected review experts in universities,this study aims to enhance the matching accuracy of the database recommendation method and address data sparsity issues and the cold start problem.A tunable hybrid recommendation algorithm was designed by combining the area under the curve AUC-MF model with an expert and project portrait-based recommendation method.The AUC-MF algorithm converts the recommendation problem into a classification problem by treating reviewed and unreviewed projects as positive and negative samples.Furthermore,it converts the data sparsity problem into a data imbalance problem,improves the matrix factorization algorithm by maximizing the AUC,and optimizes the gradient descent algorithm using LambdaMF.It can better solve the data sparsity problem in the expert sample of the university procurement project review.The recommendation algorithm models expert and project portraits using text mining,incorporating dynamic attributes(project reviews)and static attributes(expertise)for expert portraiture and using project names and summaries for project portraiture.Combining expert and project portraits as input and applying logistic regression for recommendations substantially addresses the cold start problem.[Results]Experimental results using historical review data from government procurement projects over the past six years and project text information from a certain university demonstrate the following features.① At K=3,5,and 10(where K represents the top K recommended experts),the F1 scores of the recommendations are 0.514 4,0.571 2,and 0.454 5,respectively.② The dimensionality of portrait vectors impacts recommendation outcomes with optimal dimensions identified between 50 and 60 for K=3,5,and 10.③ The parameter α(where α represents the weight of the recommendation algorithm)also impacts recommendation effectiveness,with the best result at α=0.7.④ The effectiveness of the proposed algorithm is validated on five practical university projects.[Conclusions]To address the challenge of low professional matching in the recommendation of review experts for university procurement projects,this paper leverages historical project review data and textual information to design an algorithm based on the AUC-MF model,expert portraits,and project portraits.This integrated approach effectively reduces issues related to data sparsity and cold starts,enhances the expert recommendation process,and improves the user experience for expert recommendation.
university procurementexpert recommendationmatrix factorizationexpert portrait
NETL
NSTL
万方数据
