Improving Self-Organizing Feature Map (SOFM) Training Algorithm Using K-Means Initialization

Self-Organizing Feature map (SOFM) is a competitive neural network in which neurons are organized in an l-dimensional lattice (grid) representing the feature space. The principal goal of the SOFM is to transform an incoming pattern of arbitrary dimension into a oneor twodimensional discrete map, and to perform this transformation adaptively in a topologically ordered fashion. Usually, SOFM can be initialized using random values for the weight vectors. This paper presents a different approach for initializing SOFM. This approach depends on the K-means algorithm as an initialization step for SOFM. The K-means algorithms is used to select N (the size of the feature map to be formed) cluster centers from the data set. Then, depending on the interpattern distances, the N selected cluster centers are organized into an N x N array so as to form the initial feature map. Later, the initial map will be fine-tuned by the traditional SOFM algorithm. Two data sets are used to compare between the proposed method and the traditional SOFM algorithm. The comparison results indicated that: using the first data set, the proposed method required 5,000 epochs to fine tune the map while the traditional SOFM required 20,000 epochs (4 times faster). Using the second data set, the traditional SOFM required 10,000 epochs while the proposed method required only 1,000 epochs (10 times faster)