||We propose an algorithm for single label classification, multi-label classification, and regression estimation which incorporates a rotating similarity, weighted relevance, hybrid learning, and threshold checking.|
Firstly, the rotating cluster similarity is more suitable of the distribution of the data set with hyper-spherical, hyper-ellipsoidal, or oblique hyper-ellipsoidal shapes and it is used to transform each input instance into a rotating cluster similarity. Then, the similarity of the input instance will be combined to obtain the weighted relevance of the input instance to each particular category or output value. Next, we use the hybrid learning method to refine the parameters which is in this algorithm to get better performance. Finally, the threshold checking is used to obtain the output. We will set different kind of threshold functions to determine the output due to the kind of problems.
The number of rotating clusters do not need to be specified in advance. Each cluster will self-construct during the training phase. A number of experimental results are shown the effectiveness of our proposed method.