Abstract: This paper proposes an LDL algorithm, which integrates the local correlation of labels. The algorithm is divided into three stages. In the initial prediction stage, this paper constructs a multi-layer neural network, which takes the original features as the input and the initial prediction label distribution as the outputs. In the local correction stage, first we use k-means to obtain the local information described by different clusters. Then, for each class, we calculate the corresponding covariance matrix, and finally use this matrix to correct the initial predicted label distribution to obtain the corrected label distribution. In the label fusion stage, the corrected label distribution is weighted, and then fused with the initial predicted label distribution to obtain the final predicted distribution. Compared with night popular LDL algorithms, experiments were conducted on eight different public datasets. The results show that our algorithm can better describe the local correlation of labels, and ranks higher on mainstream evaluation measures.

Key words: label distribution learning, label correlation, label fusion, k-means