Research On Academic Resource Recommendation Based On Multi-dimensional Feature Fusion Of Scholars’ Academic Preference

With the rapid growth of interdisciplinary academic activities and the number of journals,finding suitable academic journals has confronted as one of the most challenging in academic resource recommendation research.Academic journal recommendation system has become an emerging research field as a decision-making method to effectively alleviate the problem of journal overload.However,existing research results cannot meet the needs of pluralistic,diverse,and interdisciplinary academic circumstances,as well as the drift of scholars’ research interests and the dynamic changes in the scope of journal topics.To resolve such problems reasonably,this article proposes a journal matching strategy that integrates multi-dimensional feature fusion of scholars’ academic preferences,focusing on exploring multidimensional semantic features using hybrid deep learning model of multi-perspective scholars’ academic preferences contained in academic texts to improve the accuracy of journal recommendation models,to meet the personalized journal needs of scholars.Based on the feature modeling of scholars’ academic preferences,this paper studies how to extract multi-dimensional scholars’ academic preference features from academic texts to improve the accuracy of journal recommendation.Due to the different characteristics of scholars’ preference in different application situations,this paper explores and studies scholars’ academic preference modeling in contextual semantic situation,global semantic situation,personality trait situation and sample feature missing situation.Its main work and contributions are as follows:(1)Aiming at the problem of polysemy and lack of deep semantics in existing journal recommendation models,a journal recommendation model integrating the local features and higher-order features of academic text semantics is constructed.The model uses a natural language preprocessing language model to obtain the contextual semantic features of academic texts and extracts their deep semantic features by convolutional neural networks.Then,based on the real academic dataset ISa M,the recommendation performance of the model and its difference from benchmark model are analyzed,as well as the validity of the model in the cold start environment.Furthermore,considering the differences of scholars’ academic preferences,the diversity and universality of the proposed model recommendation results are studied.(2)Aiming at the problem of lack of semantic association between academic samples in existing recommendation models,a journal recommendation model fusing the local features and global features of academic texts is proposed.The model constructs a "worddocument" heterogeneous graph based on an academic corpus,and uses a graph convolutional neural network to obtain a vector representation of document nodes to quantify the global features of scholars’ preferences.During the experimental phase,the differences between the model and the benchmark model are studied,and the key factors affecting the performance of the model are analyzed.The dependencies between the parameters and the model and the robustness of the model in the cold-start environment are also analyzed.In addition,considering the differences in academic preferences of scholars,the diversity of proposed model recommendation results is also analyzed.(3)Considering the sparseness of academic text features in the dataset,introducing scholars’ personality traits,a journal recommendation model is established that enhances the personality traits of scholars and integrates local features of academic texts.The model proposes three preference factors of text explicit traits,text implicit traits and scholars’ personality traits based on the relevant writing style information of academic texts,and uses different methods to quantify these preference factors.To address the problem of personality trait factor quantification,a binary prediction model based on transfer learning strategy is proposed to identify personality traits of scholars across domains.The introduction of personality trait information,which determines scholars’ preferences,effectively alleviates the problem of sparse academic text features.Then,using real personality trait datasets(Essays)and academic datasets to verify the validity of the personality trait classification model and the journal recommendation model,respectively.(4)In view of the lack of academic text features in the dataset,integrating the semantic information of journal names,a journal recommendation model with semantic enhancement of journal names and local features of academic texts is designed.The model reconstructs the explicit features of academic texts with the help of "small sample learning" strategy and constructs a representation learning model of local features of academic texts based on the realization principle of the two-tower model.Then,the effectiveness of the journal recommendation model is verified using real academic datasets ISa M.In summary,this paper validates the effectiveness of the proposed four journal recommendation models using real academic datasets in Soupus.The results indicate that the proposed models outperform the current mainstream benchmark reference methods and have better recommendation results.In addition,the research results also provide valuable references and guidance for other recommendation scenarios in academic big data.