Deep Aligned Matrix Factorization Model For Academic Paper Recommendation And Paper Representation

In the age of information explosion,it is important to help researchers to find their interesting academic papers as efficiently as possible.Academic paper recommenda-tion is the key solution to this problem which recommends users(researchers)some interesting academic papers by modeling their interests to give them more time to do scientific research work.In general,the academic paper recommendation algorithms can be divided into content-based algorithms,collective filtering algorithms,and multi-source hybrid al-gorithms.Among them,the content-based algorithms are the most widely used algo-rithms in the early days.Because they recommend users academic papers with similar content information to their interacted papers,such recommendation can satisfy users to a certain extent,but can not surprise them.Recently,academic social networks such as CiteULike and Mendeley provide a large number of "user-paper" interaction data,and make the collective filtering algorithms based on these data being the mainstream direction in this problem,including matrix factorization as the main model.However,this kind of methods face serious problems of interaction data sparsity and academic papers cold start.Multi-source hybrid algorithms are the popular researches in this task,which introduce content information into matrix factorization model to combine the above two methods to make better recommendation.However,the key problem is how to combine multi-source information and fully use the correlation and complementarity between different sources.Focusing on such question,combining "user-paper" interactive data with the textual content information of academic papers,this thesis firstly proposes a deep matrix factorization based on"alignment" and uses it to do academic paper recommendation.In this thesis,the fol-lowing works are done.1.This thesis focuses on the task of academic paper recommendation and proposes a new hybrid matrix factorization model named Deep Aligned Matrix Factorization(DAMF)model which can take both the heterogeneity and information complemen-tarity into consideration.Firstly,by considering the heterogeneity between differ-ent data sources,DAMF model learns different representation for every data source.Secondly,to fully use the correlation between different sources to achieve the goal of information complementarity,this thesis firstly proposes an idea of "alignment",which means that different representations for one paper should be as close as pos-sible.On two datasets in the field of academic paper recommendation,this thesis compares the proposed recommendation algorithm based on DAMF model with some existing recommendation algorithms and proves the validity of DAMF model in the task of academic paper recommendation.2.To further prove the validity and versatility of the proposed DAMF model,this thesis applies it to the task of academic paper representation.By using two data sources of citation information and papers'content information as inputs,this thesis learns papers' representations and utilizes them to do academic paper classification to prove the validity of DAMF.Besides,this thesis uses several methods like 0/1,tfidf,pmi to construct the "paper-word" matrix to explore their impacts on academic paper representation.The experimental results show that using 0/1 matrix as the original "paper-word" input can achieve the best results.All the related data of the proposed model and its applications can be available in http://114.212.189.51:2012/projects/damf.