In the high-speed next generation passive optical network (NG-PON) transmissio n system based on low bandwidth optical devices,severe inter-symbol interference (ISI) p roblem would be caused due to the limited bandwidth.In this paper,we proposes a support vector machine (SVM) equalization algorithm modified by feature vector c onstruction to improve the performance of high-speed bandwidth limited NG-PON system.Considering the characteristics of these systems,a novel feature vector i s constructed to achieve a new SVM-based equalizer suitable for band-limited NG -PON systems.In addition,a many-to-many (MVM) approach to train three independent SVM classifi ers is employed to achieve multi-classification tasks for PAM-4signals.A 50Gb/s PAM-4signal t ransmission experiment based on 10-G bandwidth transceiver is demonstrated to prove the feasibility of our new SVM-based equalizer.Experimental results show that,compared with the traditional feed-for ward equalizer (FFE) and decision feedback equalizer (DFE),the receiver sensitivity achieved by our s cheme can be improved by about 1dB under the HD-FEC threshold (BER=3.8E-3), which allows the system to achieve a power budget of 20dB without the use of any amplifier.Inaddition,our scheme is more s table for CDR timing offset errors,which maintains reliable equalization performance over a relativel y large margin of timing offset.Therefore,we believe the proposed modified SVM-based equalizer could be considered prospective candidate for NG-PON systems.