The curse of dimensionality is a well-known problem in pattern recognition in which the number of patterns is smaller than the number of features in the datasets. Often, many of the features are irrelevant and redundant for the classification tasks. Therefore, the feature selection becomes an essential technique to reduce the dimensionality of the datasets. In this paper, unsupervised and multivariate filter-based feature selection methods are proposed by analyzing the relevance and redundancy of features. In the methods, the search space is represented as a graph and then the ant colony optimization is used to rank the features. Furthermore, a novel heuristic information measure is proposed to improve the accuracy of the methods by considering the similarity between subsets of features. The performance of the proposed methods was compared to the well-known univariate and multivariate methods using different classifiers. The results indicated that the proposed methods outperform the existing methods.