Image Tag Recommendation Algorithm Using Tensor Factorization

This paper aims to provide high quality tags for digital images according to users’ interest. As there are three main elements in image tag recommendation problem, tensor factorization technology is utilized in this work. In this paper, the parameters of the tensor factorization model are represented as latent variables, and the key functions of the tensor factorization model can be implemented by integrating three matrices(person matrix, image matrix, and tag matrix) into one tensor. The key problem of image tag recommendation is to obtain the top ranked tags which are suitable not only to image visual contents but also to users’ interest. Afterwards, the top ranked tags are obtained by a predictor utilizing the proposed tensor factorization model. Therefore, the image tag recommendation problem can be converted to calculate the ranking scores by maximizing the ranking statistic AUC. Finally, performance evaluation is conducted on the NUS-WIDE dataset using MRR, S@k, P@k, and NDCG metric. Experimental results show that the proposed image tag recommendation algorithm performs better than other methods.