Construction and Empirical Study of Scientific and Technological Literature Recommendation Model Based on Dynamic Multi-task Learning
[Purpose/Significance]In order to enhance the interaction of scientific and technological litera-ture recommendation scene elements and transform the problem of capturing the interactive characteristics of each element into a multi-task joint optimization learning problem,this paper constructs a scientific and technological literature recommendation model based on dynamic multi-task learning to further improve the performance of scientific and technological literature recommendation.[Method/Process]Based on multi-task learning method,the sub-tasks are deconstructed according to the key features collected from scientific and technological literature recommendation elements,and the multi-head attention mechanism was used to dynamically learn the interactive relationships of sub-tasks.A scientific and technological literature recommendation model was designed through dynamic learning of the interaction of each task.[Result/Conclusion]According to the experimental results of Ci-teULike data,the DMRSTL model constructed in this article is significantly better than the comparison model in three evaluation indicators.The highest difference is the increase of AUC indicator by 15.51％,MRR by 11.90％,and nDCG@5 indicator by 16.45％.The task combination comparative experiments further show that the interac-tive enhancement of recommendation elements can effectively improve the recommendation performance of sci-entific and technological literature.
scientific and technological literature recommendationmulti-task learningmulti-head atten-tion mechanismtask interaction
NETL
NSTL
万方数据
