Instance-ranking: A new perspective to consider the instance dependency for classification

Single-label classification refers to the task to predict an instance to be one unique label in a set of labels. Different from single-label classification, for multi-label classification, one instance is associated with one or more labels in a set of labels simultaneously. Various works have focused on the algorithms for those two types of classification. Since the ranking problem is always coexisting with the classification problem, and traditional researches mainly assume the uniform distribution for the instances, in this paper, we propose a new perspective for the ranking problem. With the assumption that the distribution for the instance is not uniform, different instances have different influences for the distribution, the Instance-Ranking algorithm is presented. With the Instance- Ranking algorithm, the famous K-nearest-neighbors (KNN) algorithm is modified to confirm the validity of our algorithm. Lastly, the Instance-Ranking algorithm is combined with the ML.KNN algorithm for multi-label classification. Experiment with different datasets show that our Instance-Ranking algorithm achieves better performance than the original state-of-art algorithm such as KNN and ML.KNN.