Semi-supervised LPP algorithms for learning-to-rank-based visual search reranking

• Two feature dimensionality reduction methods for search reranking are proposed.
• The SSLPP algorithm improves LPP by incorporating the relevance degree information.
• The proposed PCC-SSLPP algorithm models the correlation information into SSLPP.
• The proposed methods achieve better NDCG performance.

Recent years have witnessed the success of learning-to-rank technique in visual search reranking. Generally, dimensionality reduction is one of the important stages in this kind of method to deal with the high-dimensional visual features. However, these methods are typically designed for classification and retrieval applications, rather than ranking. Direct utilization of them cannot guarantee the best performance in ranking applications as ranking information (e.g., relevance degree label) is absent. To solve this problem, a Semi-Supervised LPP algorithm called SSLPP is proposed. The method incorporates the relevance degree information into a manifold dimensionality reduction method, i.e., Locality Preserving Projections (LPP). To accurately preserve the correlation information between two images in the low intrinsic dimensional space, a Pearson Correlation Coefficient (PCC) based SSLPP algorithm called PCC-SSLPP is further developed. The PCC-SSLPP method models the correlation information between two images in the graph construction step of SSLPP. Besides, a novel visual search reranking scheme is also presented based on the proposed SSLPP/PCC-SSLPP and learning-to-rank. Experiments on two popular real-world image reranking datasets demonstrate the effectiveness of the proposed scheme and algorithms in visual search reranking.