• Applying low-cost 10 G-class optical transceivers to 50 Gb/s PAM4 IM-DD transmissions. • The proposed algorithm combines traditional equalization and machine learning. • The impact of data pre-processing and algorithm structure on algorithms is studied. • Complexity and performance analysis of the algorithm is performed. A light gradient boosting machine LGBM-improved-MLSE equalizer for 50 Gb/s PAM4 IM/DD PON with 10 G class optical transceivers is proposed and experimentally demonstrated. The impact of delay-taps and delay-spans on equalization performance are studied. We also analyze the performance of LGBM-modified-MLSE as the different numbers of tree. Experimental comparisons of LGBM-modified-MLSE algorithm with 16-tap and 32-tap, MLSE (L = 4), and 100-tap FFE with MLSE (L = 4) are made. The results show that the proposed algorithm can get a power budget of 21 dB after transmitting 15 km of SSMF, which is 3 dB superior to MLSE (L = 4) and 1 dB superior to 100-tap FFE with MLSE (L = 4), respectively.