Research Of Models And Methods For Multi-Granular Network Representation Learning

With the development of the Internet,a large amount of data such as user contents,user behaviors,and user relationships have been accumulated on social networks.People can discover useful knowledge from these data to predict the evolution of relationships among users,explain users’ behavior patterns,depict users’ preferences,analyze the evolution pattern of network events,and so on.The network representation is related to the effectiveness and efficiency of analysis on the networks.The large scale and complexity of network data impose high requirements on the representation methods of network data.The traditional adjacency matrix-based network representation method has many shortcomings,while the representation learning technique has great advantages over the adjacency matrix.Network representation learning uses a set of low-dimensional vectors to represent the original network components,which reduces the complexity of the representation results while preserving the original network features.By analyzing the low-dimensional representation vector,the knowledge contained in the original network can be revealed,thus helping to solve various network inference problems.Multi-granularity cognitive computing is an important research direction developed based on the theory of granular computing,which draws on the multi-granular mechanism of the human brain in dealing with complex problems and attempts to solve problems collaboratively at multiple levels of granularity.Network structure is the basic research object in networks,which presents different functional properties at different granularities.Network structure is characterized by multi-level and multi-granularity.Therefore,it is necessary to draw on the ideas of multi-granularity cognitive computing to analyze and mine the multi-granularity knowledge in the network,and solve the problems in the field of network representation learning.This thesis provides an in-depth analysis and study of the representation learning problem of multi-granular structural features,and proposes representation learning methods for collaborative extraction of multi-granular structural features for crossnetwork user alignment,and interpretable multi-granular structural feature representation learning model for hierarchical structure semantics of network data.The main research contents of this thesis include: multi-granular network representation learning model for high-order structural feature extraction,multi-granular network representation learning model for nonlinear structural feature extraction,graph leading tree model for multigranular hierarchical semantic extraction,and multi-granular interpretable network representation learning method based on graph leading tree model.The main innovative research work of this thesis includes:(1)Aiming at the problem that network representation learning models have difficulty discovering LAP(Latent Anchor Pair)with high-order structural proximities,a multi-granular network representation learning model for high-order structural features extraction is proposed.With the idea of multi-granularity cognitive computing,the heuristic weighted random walk mechanism is designed to extract the high-order structural features and enhance the ability of network representation learning algorithm to capture high-order structural features.The representation vectors help to align users across networks and the alignment effect is improved.(2)Aiming at the problem that network representation learning models have difficulty discovering LAP with nonlinear structural proximities,a multi-granular collaborative representation learning model for nonlinear structural feature extraction is proposed.A deep learning-based edge wighter is designed to improve the ability of representation learning model to capture nonlinear features in supervisor anchor pair(SAP).By co-extracting high-order structural features and nonlinear structural features,the similarity of users with the same identity is enhanced,and the effect of representation learning model on cross-network user alignment task is further improved.(3)Aiming at the problem that the density-based leading tree model cannot express the multi-level structural features in network data,a graph leading tree(Graph LT)model is proposed based on the multi-granular hierarchical data representation model of leading tree.By mapping the network nodes onto a tree structure,the multi-granular hierarchy in the network is expressed intuitively.This work solves the limitation problem of former network representation learning models that focus only on local structural similarity extraction and the distribution of representation vectors cannot reflect the global hierarchical relationships.(4)Aiming at the problem of insufficient interpretability of network representation learning models,a multi-granular interpretable network representation learning method based on the graph leading tree model is proposed.The interpretable tree structure guides the generation process of representation vectors,and the distribution of the generated representation vectors can reveal the semantics of hierarchical structures in the network,which improves the effectiveness on network analysis tasks such as network node classification and link prediction.(5)Based on the research archievements of this thesis,a multi-granularity network structural characteristics analysis system(MGSAS)is developed and functionally verified.In the cross-network user information fusion analysis task,as a supplement to the explicit information analysis method,the system uses structural information for analysis and achieves improved results.In summary,based on the academic thought of multi-granularity cognitive computing,in the aspect of structural feature mining in social network data,this thesis takes network representation learning model as the main research object and researches the multi-granular collaborative network representation learning method and multigranular hierarchical semantic representation learning model.The problem of insufficient representation of multi-granular structural features in network representation learning models is solved.The proposed network representation learning approach improves the effects of network inferences such as cross-network user alignment,multi-level community detection,node classification,link prediction,etc.