Semi-supervised Classification with Multiple Ants Maximal Spanning Tree

Xiao-hua Xu, Lin Lu, Ping He, Yue Ma, Qi Chen, Ling Chen

Abstract: This paper presents a multiple ant colonies random walk model for semi-supervised classification task. Taking the multi-species competition mechanism into consideration, this model treats the limited labeled data as heterogeneous ant colonies according to the predefined classes, these colonies live in their nests and communicate indirectly by their unique types of pheromone. The unlabeled data set is taken as the food resources, which transforms the conventional classification problem into the scramble for resources among multiple ant colonies. During the swarm random walk in such an environment, every nest fights against others by the accumulation of different class pheromone on these resources for the possessions. The proposed algorithm is evaluated with real life benchmark datasets in terms of classification accuracy. Meanwhile the method is compared with other techniques. Experimental results show the potentiality of the proposed algorithm.