Reading Notes: Ranking Graph Embedding for Learning to Rerank

Mainidea  In thispaper, the author proposes a method that introducing ranking information intodimensionality reduction significantly increases the performance of imagesearch reranking.

1)      The proposed method transformsgraph embedding, a general framework of dimensionality reduction, into rankinggraph embedding (RANGE) by modeling the global structure and the localrelationships in and between different relevance degree sets, respectively.

2)       The proposed method also defines three types of edge weightassignment between two nodes: binary, reconstruction, and global.

3)       In addition, a novel principal components analysis based similaritycalculation method is presented in the stage of global graph construction.

 

Motivation  Dimensionality reduction is a key step to improving thegeneralization ability of reranking in image search. However, existingdimensionality reduction methods are typically designed for classification,clustering, and visualization, rather than for the task of learning to rank. Withoutusing of ranking information such as relevance degree labels, directutilization of conventional dimensionality reduction methods in ranking tasksgenerally cannot achieve the best performance. So the authors propose a dimensionalityreduction approach based on ranking information like different relevance degreeset and global structure.

 

Contribution Contributionsin this paper are summarized as follows:

1) Considering ranking information in dimensionalityreduction, the author proposes a novel algorithm by extending graph embeddingwith relevance degree labels information.

2) A new semi-supervised featuredimensionality reduction based image search reranking framework is proposed.

3) Develop a PCA-based adaptive edge weightassignment method in graph construction.

4) divide the edge weight assignment intothree types: binary, reconstructive, and global. Binary edge weight assignmentis used for modeling the relationships between two very relevant nodes, tworelevant nodes, two irrelevant nodes, and a very relevant node and anirrelevant node. The reconstruction one is used for modeling the relationshipsbetween a relevant node and a very relevant node, a relevant node and anirrelevant node. The global one is used for modeling the “intrinsic structure”over the whole examples.

 

PersonalOpinion  Consideringour work on hashing and image retrieval, the model of this paper can be seen asa semi-supervised hashing model. It seems like that it just learn a projectionmatrix W for dimensionality reduction. As discuss above, the contributions ofthis paper lie in that for the first time, a dimensionality reduction method consideringthe label information and global structure is proposed. The idea and optimizationmethod are simple, but it works.