Network Joint Representation Learning And Applied Research

The research goal of network representation learning is to learn the transformation approach of new feature representations based on the supervised or unsupervised learning methods for the original networks,so that new features can adapt to various machine learning tasks without heavy feature engineering from people.Network representation learning can be regarded as the network encoding task,namely,each node in the network is given a unique representation vector,consequently,nodes with similar structures have a closer distance in the network representation vector space.Based on the above research objectives and the existing network representation learning algorithms,we first study more effective network joint representation learning algorithms so as to fully mine the hidden feature factors of the networks,and then the learnt network representation vectors contain various feature hidden factors of the networks,rather than just the structure features of the networks.There exist many classical algorithms on network representation learning,which can fully mine the structural correlations between vertices.There are also some classical algorithms for the network joint representation learning.These researches in this thesis are the important supplements and further optimizations for the existing network joint representation learning algorithms,which provide some theoretical and experimental foundations for the follow-up researches.Therefore,the network joint representation learning model studied in this thesis is mainly divided into three parts,which are introduced as follows:(1)The network representation learning based on random walk strategiesHere,the random walk strategies of equal probability and unequal probability are adopted to improve the random walk procedures of the network representation learning.This section introduces the basic knowledge and theories of network representation learning in detail,and it is also the research basis of the follow-up works of this thesis.(2)The network representation learning based on relationship modelingHere,the texts.hierarchical relationships and position information of neighboring nodes are embedded into the network representation vectors.In order to embed multi-type features into learning model,we introduce the thought of multi-relational modeling to transform these features into vertex triplet forms,which can constrain and guide the network representation learning procedures.(3)The network representation learning based on multi-view feature embeddingHere,text view features,weight view features and structure view features are embedded into network representation vectors.This learning procedure is based on the multi-view joint representation learning.Therefore,two different multi-view joint representation learning frameworks are introduced in this thesis.Network representation learning algorithms based on neural networks originate from word vector representation learning algorithms.Therefore,there exists a stronger correlation between network representation learning and word representation learning.This thesis begins with network representation learning,and absorbs the frameworks and ideas of the advanced network representation learning algorithms,and then applies them to word representation learning algorithms.Due to the inconsistency of word representation learning algorithms in corpus collection,word frequency filtering,noise processing and part-of-speech tagging,it is difficult to conduct comparison tests.However,there exist already some standard datasets for network representation learning.If we introduce the advanced algorithms of network representation learning into the word representation learning algorithm,we can save enough time and energy for experimental verification.Based on such assumption,we conduct the following two kinds of application researches on word representation learning:(1)To improve the acquisition procedures of context words in word representation learning,namely,we only retain the current central word and its dependencies related to the current central word,and delete the words and the dependencies which have no dependency relations with the current central word.In addition,the polysemous words in the language model are identified and tagged,and then we train the language model based on polysemous word analysis.(2)The multi-view joint representation learning method is adopted to embed the features of context structure,attribute semantics features,synonym and antonym features into the low-dimensional word representation vectors.Moreover,this thesis studies the link prediction based on network joint representation learning algorithm,which provides another method and solution for link prediction and other tasks.This thesis does not directly use the existing network representation learning algorithm to conduct the link prediction evaluations,we propose a novel high-order network joint representation learning algorithm in this thesis,and verifies its feasibility and stability in link prediction tasks based on the proposed algorithm.In summary,the innovations of this thesis are to research how to use the joint learning model and framework to improve the performance of the network representation learning algorithms.The joint learning model includes the method of joint learning between algorithm and data,and the approach of joint learning between algorithm and algorithm.