Study On Vector Representation Of Tree Structured Data

With the continuous advancement and development of the Internet technology,the scale of network data is expanding.The organizational form of the data is varied according to the different application scenarios.As an effective way of data organization,tree structure can reflect the hierarchical relationship of data,which is conducive to complementary information between different levels and heterogeneous information integration.For example,in the traditional text processing,a book can be solved by the "bag of words" model,but this approach ignores the hierarchical structure of information which is determined by the characteristics for the structure of books.And the tree structure can effectively express the hierarchical structure of the data.Therefore,it is of great significance to study the quantitative representation of tree structure data.The quantitative representation of tree structure data has not been studied.In the existing related works,it mainly achieved the low dimensional mapping of the bottom information of tree structure,so as to achieve information fusion,but it does not consider the hierarchical characteristics of tree structured data.Therefore,this study considers the characteristics of the tree structured data and achieve the information fusion from bottom-up layer by layer.Then,we can get the unified vector representation which can be used for further application such as classification or clustering.This research proposes an effective framework for vector representation of tree structure,which can be divided into two stages,namely,the representation of tree structure and the integration of hierarchical information.The information fusion for each level of tree structure can be divided into two categories,namely,the tree structure formed by the homogeneous data,and the tree structure formed by heterogeneous data to quantify.The tree structure formed by homogeneous data organization,namely the tree node information is inherited from its parent node,this research designed two local reconstruction models based on the theory of sparse encoding which reconstructs the parent node information using the children node information,in order to obtain the hierarchical information hiding in the tree structure to enhance the vectorial representation.The tree structure formed by the heterogeneous data of which each level comes from the different domains.We obtain the clusters of nodes in each level by clustering algorithm,which is used to cluster for the data coming from the same domain,i.e.,the global information is acquired from each level,and realizes the location mapping for the nodes from each hierarchical.The information fusion process mentioned above is proceed from bottom to up,so as to obtain a unified vector representation.Further,the vector representation of tree structure is applied for the e-book recommendation,the author recommendation and image retrieval,and the comparison experiments with various algorithms are carried out to verify the effectiveness of the framework.