The　performance　of　the　machine　learning-based　arc　fault　diagnosis　method　is　significantly　influenced　by　the　effectiveness　of　features.　This　article　proposes　the　use　of　the　feature　selection　(FS)　method　to　optimize　the　feature　quality　and　classification　result　for　arc　fault　diagnosis.　An　experimental　system　is　constructed　to　generate　arc　fault　current　signals　under　various　conditions.　The　limitations　of　the　predetermined　threshold-based　methods　in　time,　frequency,　and　time-frequency　domains　are　outlined,　and　the　dependence/redundancy　issues　present　in　the　feature　dataset　are　analyzed.　To　mitigate　the　impact　of　these　issues,　a　light　gradient　boosting　machine　(LightGBM)-based　FS　method,　which　establishes　the　feature　importance　evaluation　criterion　for　the　selected　optimal　feature　subset　based　on　permutation　importance　(PI),　is　proposed　in　this　article.　The　fitness　proportionate　sharing-based　feature　clustering　(FPS-FC)　method　searches　for　potential　feature　clusters　and　selects　a　feature　subset　with　low　redundancy　to　accommodate　unlabeled　data.　Finally,　both　the　proposed　FS　methods　are　compared　with　six　popular　supervised/unsupervised　methods　using　two　classifiers　across　seven　datasets.　The　results　validate　the　effectiveness　of　the　LightGBM-based　and　FPS-FC-based　FS　methods　in　enhancing　the　performance　of　arc　fault　diagnosis.　In　addition,　a　designed　case　on　Raspberry　Pi　verifies　the　feasibility　of　the　proposed　methods　in　real-world　applications.　©　1963-2012　IEEE.