Research On Representation Learning And Zero-Shot Learning Of Network

With the continuous development of Internet,the number of users on the internet is increasing.Network data is becoming more and more common in our daily life.The goal of our research is to mining and utilize the latent information of network data.The main work of this paper is organized around three issues on network data.The first problem is that sampling progress of traditional network repre-sentation learning lacks flexibility.The second problem is zero-shot learning of network data.The last problem is that there are few systems to analyze net-work data and visualize the network representation learning methods,which needs to be improved immediately.Network representation learning technology is the way to make full use of such kind of data.With the help of network representation learning tech-nology,we can mining latent information from network data.For example,we can obtain portrait of users through their social network.Besides,we can extract the information in the protein-protein interaction network to help us predict the interaction between other protein.Network representation learning has evolved from dimensionality reduction based models to matrix decomposition based models and then neural network based models.The use of network data has developed from single data to multi-source heterogeneous data such as text,images,attribute tags and so on.On account of the inflexibility of traditional random walk-based network representation methods,this paper utilizes the Markov Chain to control the process of nodes sampling.The model provides a more flexible sampling strategy to improve the quality of representations for the nodes.However,the strategy is based on the distance metric which brings about the problem of excessive computational complexity.However,we successfully reduce complexity of the algorithm through bit operation.Finally,the model has achieved best results in the tasks of node classification and link prediction.In addition,the paper also analyzes the influencing factors of the order in the model and the scalability of the model(including nodes scalability and edges scalability)by theory and experiments.For the problem of zero-shot learning in network data,the existing net-work representation learning methods cannot classify the classes with no training data.To solve the problem of zero-shot in network data,this pa-per proposes a solution to it:utilizing autoencoder and optimization based on semantic constraints.The encoder of the model is graph convolutional network and the decoder of the model is inner product decoder.We use semantic constraints to optimize the model.At last,the classification is per-formed by calculating the distance between encoded node vectors and feature vectors of classes.Finally,through the experiments in 3 datasets,the model is proved to be valid which means it can combine the network information,node semantic information and category semantic information very well.Nowadays,there are few network analysis and visualization systems.This paper designs a network analysis and visualization system based on the situa-tion.The system allows users to perform operations such as data uploading,network representation learning,data analysis,data visualization,etc.In ad-dition,the paper introduces the network analysis and visualization system from multiple perspectives.The system integrates the relevant network rep-resentation learning methods.Besides,it allows users to perform interactive operations and conduct error analysis.It helps researchers reduce the burden of programming which has important practical value.