Instead of arrayed -waveguide grating (AWG) coder/decoders approach, we aim at intelligent coding computations to mitigate interference noises from noises-accumulated multi-carriers transmissions. Recursive interference cancellation result will be worse off if there is strong noise in the transmission channel. An interference cancellation method based on convolutional neural network (CNN) was proposed to the increased or cancel noise to improve accuracy of retrieving optical target data signals in multiuser systems. In this paper, we focus on the training gathering, and analysis. The training data for CNN model building was discussed with different decision rules. The performance of CNN-based interference cancellation method was defined according to the analysis result of training data collection. The analysis result shows that if we can retrieve the estimation noise which approach the true noise, the better BER will be acquired.