Overlapping spikes sorting using feature fusion

In this paper, a new feature fusion strategy is proposed directing at the classification issue of overlapping spikes. Firstly, wavelet coefficient features and PCA (principal component analysis) features of spikes are extracted. Secondly, based on the idea of local preserving projection, for different features, the graph incorporating neighborhood information of the data sets was built. A key issue in feature fusion is to derive a fusion graph to reflect relative importance of different features, which can be achieved by a weighted method. Using the notion of the Laplacian of the graph, then a transformation matrix, which maps the data points to a subspace was computed. The two features extracted are fused into a new eigenvector by the transformation matrix. According to the experiment under different noise, this method can obtain an ideal classification effect by adopting lower-dimension features. When the noise level is 0.4, the accuracy rate of this method is increased by 9% compared with single-feature classification on average.