Hierarchical Affinity Learning for Training Evaluation

Appropriate affinity/similarity measures always play a critical role in data mining. The complex interactions among multiple features and personality of each individual object makes it still a challenging problem. Existing methods simply consider the relevance in a feature-pair manner, and they treat the features for each object equally without considering the personality. In this paper, we propose a hierarchical affinity learning method on categorical data with unsupervised personalized feature weighting, called HAL. HAL captures the interactions by exploring the affinities among objects, features and values, which carry intrinsic data characteristics, via hierarchical affinity learning to handle this complex data. The inferred affinities between objects and features can be treated as the personalized feature weights which is used to refine the initial affinity matrix. The learned affinities between objects obtained by reinforcement affinity learning can be exploited for clustering. Experimental results on 16 real world datasets with diverse characteristics from 6 different domains confirm the superiority of our method. Compared to the state-of-the-art measures, it averagely achieves 8.8% improvement in terms of F-score.