To reduce the dimensionality constraints of neural network decoders for polar codes during the training phase, a partitioned successive cancellation (SC) decoder based on fully connected neural networks (FCNN) is designed. By dividing the polar code decoding tree into two regions and processing each with differently parameterized FCNNs, the need for large-scale training data is reduced. The simulation results show that in an additive white Gaussian noise (AWGN) channel, when the signal-to-noise ratio (SNR) is between 1 to 5 dB, the performance of the FCNN-SC decoder approaches that of the SC decoding algorithm. When the SNR is between 1.5 to 3 dB, the FCNN-SC decoder achieves approximately 0.5 dB coding gain compared to the FCNN decoder, and requires a smaller dataset during the training phase, being roughly half the size needed for the FCNN decoder.