Network Representation Learning Based Node Attribute Inference

The rapid development of online social networks and e-commerce attracts more and more people to participate in it.Obtaining rich user attributes is beneficial for operators to better target users and design personalized recommendation systems.Attribute inference is a technological means used to predict the unknown attributes and potential traits of a node.However,the features extracted by existing attribute inference methods suffer from high dimensionality and sparsity.Furthermore,existing paradigms typically employ single attribute inference model which ignores the inherent correlation between attributes.Therefore,existing attribute inference methods have low accuracy and low efficiency.This thesis mainly studies the node attribute inference algorithm based on network representation learning.An efficient network representation learning algorithm is designed to automatically learn the low-dimensional feature representation of nodes in different application scenarios,so as to solve the problem of traditional feature extraction in attribute inference.At the same time,we take into consideration the correlation among different attributes by exploring an appropriate prediction model to infer multiple attributes simultaneously.The main points of this thesis are summarized as follows:(1)A structured attribute inference algorithm based on social embedding is proposed.Firstly,it embeds the user's friend relationship of an online social network into a low-dimensional vector space.Then,a structured attribute vector is proposed to encode multiple user attributes.Finally,a multi-layer neural network model is designed to capture the complex nonlinear mapping relationship between the user's social embedding vector and the structured attribute vector,so as to improve the accuracy of attribute inference.(2)A multi-attribute inference algorithm based on user node embedding is proposed.Based on the user's purchase records in a period of time,a novel node embedding algorithm is proposed.Firstly,a directed and weighted bipartite graph is constructed.Then,a biased random walk is designed to generate the sequences which embody context corpus.Finally,the vector representation of each node is learned by employing Word2 Vec algorithm on the aforementioned sequences.The node representation learned by this embedding algorithm contains more feature information than the traditional dimensionality reduction methods.In addition,a multi-attribute inference model is proposed,which utilizes the correlation between multiple attribute inference tasks to achieve mutual promotion.The model takes the vector representation of the user node as input,and jointly optimizes the loss of multiple tasks to improve the inferred accuracy.In general,this thesis mainly studies the network representation learning algorithm to automatically learn the low-dimensional feature representation of nodes.Moreover,it considers the intrinsic correlation of multiple attributes to infer multiple attributes simultaneously by exploring new models.Extensive experiments on real social network and purchase scenario datasets are performed.The results demonstrate that the feature representations learned by user node embedding algorithm contain richer attribute information.Besides,both of the proposed structured attribute inference method and multi-attribute inference model improve the efficiency and accuracy of attribute inference.