Research On Key Issues Of Academic Paper Recommendation

As a presentation form of scientific research results,academic papers are the main reference for researchers to obtain the latest scientific progress.In recent years,interdisciplinary research has been highly valued by the academic community and has developed rapidly.It is not only reflected in the increasing number of published interdisciplinary papers,but also in the increasing proportion of different discipline related works in the paper references.The existing paper recommendation models mainly focus on the semantic relevance between paper contents.Such calculation methods of symmetrical relevance on general semantics are not suitable for dealing with the asymmetric knowledge propagation across disciplines.Also,the word semantic ambiguity between different disciplines makes the paper correlation difficult to be analyzed.In addition,with the rapidly increasing numbers of papers published,it has become more difficult for researchers to select high-quality works from a large number of papers in order to obtain new knowledge.The existing methods for evaluating the quality of newly published papers mainly consider the potential citations and the authority of paper authors,and give a unified quantitative evaluation.However,since the academic innovation forms are diverse and the paper contents has different aspect,it is difficult to evaluate a paper under a unified quantitative framework.It is important to model the academic knowledge propagation and the potential influence of new papers for helping researchers use literature resources more efficiently.The researches on personalized interdisciplinary paper and new paper recommendation have profound significance and practical value.This dissertation focuses on the researcher requirements for interdisciplinary papers and newly published papers.The main contributions and innovations are as follows:(1)Aiming at the cross-domain paper recommendation problem,this dissertation proposes domain semantics and semantic correlation extraction models,as well as a personalized cross-domain paper recommendation method.To reduce the semantic ambiguity in difference academic domains,a topic model based generative method is adopted to represent a paper content as the probabilistic association with an existing hierarchically classified discipline.Compared with the classic distributed word embeddings,the domain semantics modeled against the hierarchical categories reveals word usages and language features in different scenarios,and supports the introduction of new words and semantic evolution.A neural network is trained to model the asymmetrical academic knowledge propagation between different disciplines as an influence function.Different from the previous cross-domain product recommendation methods that model the entity associations between different platforms as symmetric relationships,academic influence propagation conforms to the dissemination law of academic knowledge.User interests are represented as probabilistic distribution over the target domain semantics and the correlated papers are recommended.The experimental results on the real data set verify the effectiveness of the recommended model,and the implicit features in the results increase the interpretability of the model.Experimental results on real datasets show the effectiveness of these methods.The intrinsic factors of results are also discussed in an interpretable way.Compared with traditional word embedding based methods,this approach supports the evolution of domain semantics that accordingly lead to the update of semantic correlation.Another advantage of this approach is its flexibility and uniformity in supporting user interest specifications by either a list of papers or a query of key words,which is suited for practical scenarios.(2)Aiming at analyzing paper differences,this dissertation proposes a paper subspace embedding model that integrates multiple expert rules,and a multiple aspects paper difference comparison method.Consider different aspects of paper academic innovations,the concept of subspace is introduced to distinguish the research background,research methods,and experimental verification in the paper text.A pre-training model based subspace semantics embedding method is proposed to analyze the correlation on paper difference with influence.Compared with traditional paper text representation methods,subspace embedding can better reflect the different characteristics of academic innovation.In order to analyze the paper difference at multiple aspects,the expert rules are used for classifying papers to automatically label paper differences,based on which we a twin-network is proposed with the contrast loss for learning the paper embeddings in different subspace.Compared with the supervised models on predicting paper citations,subspace embedding based method can eliminate the influence of numerical deviations such as citations caused by different innovative forms and subject characteristics.It can also eliminate the influence of different rules and scoring scales while integrating expert knowledge,and increase the model's robustness.The experimental results on the real datasets verify the positive correlation between subspace differences and paper citations.Compared with other text embedding methods and paper quality prediction methods,our method is more relevant to the actual citations.Specially,we analyze the highly-cited papers in different disciplines and discuss the distribution characteristics of their subspace embeddings.We also perform the experiments to analyze the influence on different parts of our method.(3)Aiming at the newly published paper recommendation problem,this dissertation proposes a joint multi-feature embedding model of academic network entities,as well as personalized new paper recommendation and citation recommendation methods.Considering the research contents,referring paper habits,and the asymmetric influence of new papers,a paper embedding method combining text and academic network is proposed.Aiming at the paper research content on multiple aspects,the pre-trained paper subspace embedding is used as the text representation.The academic influence between papers is the asymmetric knowledge propagation,which can be inferred by the citation relationships,the author authorities and so on.A joint multi-feature embedding method of academic network based on graph convolution is proposed,which embeds the citation habits and potential influence features of papers respectively.Compared with traditional paper evaluation methods,our model takes into account the personalized concerned internal influence features when referring papers.In order to improve the training effect of the recommendation model,a data labeling strategy is proposed for ignoring blurry samples.The researcher publications are used to model user interests,and the papers are recommended based on the correlation between research interests and candidate papers.The experimental results on real data sets verify the effectiveness of this method in personalized paper recommendation and citation recommendation.The necessity of multi-feature embedding is verified by analyzing the embedded semantics of authors and paper elements,and the citation habits and influence of different types of authors and highly cited papers are reflected in a visual form.The reusability of the method in this paper is verified on other types of academic resources with fewer features.(4)In response to the paper recommendation requirements in colleges,this dissertation designs and implements a paper recommendation prototype system.Based on the academic community's emphasis on interdisciplinary and the latest innovative researches,the academic paper recommendation requirements from colleges are analyzed.In response to the urgent needs of university researchers for interdisciplinary papers and newly published papers,a research-driven paper recommendation system for colleges is designed and implemented.Different from the general paper recommendation and advanced retrieval functions provided by the previous academic service platforms,this prototype system caters to the development trend of cross-research and realizes the cross-disciplinary paper recommendation function,based on the advanced retrieval requirements proposed by users,such as source domains,retrieval target domains,representative works,query keywords and other forms of requirement representations,recommend the most relevant academic papers for users.In order to cope with the current challenges of the large number and high speed of newly published papers,this prototype system implements a personalized recommendation function for new papers,which models user portraits based on user registration data,paper downloading and other behavioral data,and recommends the most relevant new papers.The illustrative examples and feedbacks of multidisciplinary researchers on the system verify that the prototype system can accurately receive,process,retrieve and return data,the key algorithms are implemented properly,and the programs run in accordance with the expected logic.It also verifies the feasibility of the recommended algorithm in this paper in practical applications.The continuous optimization strategies of the prototype system are also discussed,such as designing an iterative mechanism to expand the corpus of newly published papers and background model knowledge,adding a feedback mechanism based on user ratings to facilitate optimization of the recommendation model,designing user subscriptions,automatic generation of search terms and other functions.