In this paper, we demonstrate a newly designed multi-source domain transfer learning (MST) scheme to reduce the training cost of deep neural network (DNN) based equalizer in intensity-modulation and direct-detection (IMDD) systems. Different from a common transfer learning algorithm, in this scheme, data with different channel parameters is selected and proportionally used to construct a multi-source domain dataset. This allows training the source domain in a single task while ensuring the model's generalization ability and stability. In an 80Gb/s PAM-4 IMDD short reach system, our proposed MST equalizer was proven effective. The corresponding results demonstrate that, compared to a conventional DNN equalizer, the proposed MST equalizer can achieve a bit error rate that meets the hard decision-forward error correction threshold while saving 87% of the iteration epochs and 65% of the training data.