The proposed scheme of computing the similarity by equation (3) has the higher complexity as payment for obtaining the more discrimination among sparse vectors. Let us assume that two d dimensional numerical vectors are given as the input for computing the similarity between them. It takes only linear complexity, O(d), to compute the cosine similarity as the traditional one. However, in the proposed scheme takes the
quadratic complexity, O(d2). We may reduce the complexity by computing similarities of some pairs of features, instead of all.