Click data guided query modeling with click propagation and sparse coding

We address the problem of fine-grained image recognition using user click data, wherein each image is represented as a semantical query-click feature vector. Usually, the query set obtained from search engines is large-scale and redundant, making the click feature be high-dimensional and sparse. We propose a novel query modeling approach to merge semantically similar queries, and construct a compact click feature with the merged queries. To deal with the sparsity and in-consistency in click feature, we design a graph based propagation approach to predict the zero-clicks, ensuring similar images have similar clicks for each query. Afterwards, using the propagated click feature, we formulate the query merging problem as a sparse coding based recognition task. In addition, the hot queries are utilized to construct the dictionary. We evaluate our method for fine-grained image recognition on the public Clickture-Dog dataset. It is shown that, the propagated click feature performs much better than the original one. In the query merging procedure, sparse coding performs better than traditional K-mean algorithm. Also, the “hot queries” outperform K-SVD in dictionary learning.