The reasons why the classifier can work without the user specifying the reference similarity r and the number of classes K can be summarized as follows.
Then, the average of the centroids of the two clusters is taken as a reference similarity r to separate the two distributions.
The fact that the similarity values between each SPN and the rest of the dataset can be grouped into intra-class and inter-class as described in Section 2.3.2 facilitates adaptive determination of the reference similarity r automatically.
Based on the pairwise correlation matrix, a reference similarity is determined using the k-means (k=2) clustering algorithm and a membership committee, which consists of the most similar SPNs of each SPN, is established.
Based on the similarity matrix, each SPN is treated as a random variable to be assigned a class label, and a reference similarity r is estimated to serve as a rough boundary between the intra- and inter-class similarities in order to encode a cost function using a Markov random field (MRF).
The details of the algorithm are given in Procedure 2. We assume that the reference similarity r to be determined lies in between [low,high], so the sum and size of the similarities between (0,low) and (high,N) can be pre-calculated before the iterative update.
where W ij is the similarity (see Eq. (2)) between n i and n j (j∈C i ), r i is the reference similarity described in Section 2.3.2, and s l,f j ) is a sign function defined as s l,f_{j})= left{ begin{array}{ll} +1, quad l neq f_{j} -1, quad l = f_{j}.
The term W ij −r i in Eq. (6) also provides adaptability and helps the clustering to converge because it gives more say to the SPNs with the similarity value farther away from the reference similarity r in determining the class label for the SPN in question.
article similarity according to their cited references and reference similarity according to their source articles was significantly higher within than among clusters.
