Learning similarity with cosine similarity ensemble

• This paper proposes a cosine similarity ensemble (CSE) method to learn similarity.
• CSE is a selective ensemble and combines multiple cosine similarity learners.
• A learner redefines the pattern vectors and determines its threshold adaptively.
• Experimental results show the superiority of CSE.

There is no doubt that similarity is a fundamental notion in the field of machine learning and pattern recognition. How to represent and measure similarity appropriately is a pursuit of many researchers. Many tasks, such as classification and clustering, can be accomplished perfectly when a similarity metric is well-defined. Cosine similarity is a widely used metric that is both simple and effective. This paper proposes a cosine similarity ensemble (CSE) method for learning similarity. In CSE, diversity is guaranteed by using multiple cosine similarity learners, each of which makes use of a different initial point to define the pattern vectors used in its similarity measures. The CSE method is not limited to measuring similarity using only pattern vectors that start at the origin. In addition, the thresholds of these separate cosine similarity learners are adaptively determined. The idea of using a selective ensemble is also implemented in CSE, and the proposed CSE method outperforms other compared methods on various data sets.