Previous　studies　have　demonstrated　diverse　effects　of　different　factors　on　occupant　window　behaviours.　It　is　necessary　to　choose　appropriate　subsets　of　different　behavioural　window　opening　features,　and　to　eliminate　irrelevant　and　redundant　features　so　as　to　avoid　overfitting,　noise　and　random　fluctuations　being　learned　by　the　model,　and　improve　the　accuracy　of　predictive　models　of　window　opening.　The　choice　of　protocols　for　the　selection　of　features　has　been　widely　accepted　as　one　of　the　most　important　steps　in　developing　machine　learning　prediction　algorithms.　This　study　employed　the　use　of　both　a　recursive　and　a　non-recursive　feature　selection　method　designed　to　consider　all　influencing　factors　simultaneously　to　explore　the　confounding　effects　inherent　in　various　factors　pertaining　to　the　prediction　of　window　opening　behaviour.　Two　machine　learning　algorithms　were　applied　as　estimators　in　a　recursive　selection　process,　namely　support　vector　classification　(SVC),　logistic　regression　(LR),　and　one　in　a　non-recursive　process,　namely　random　forest　(RF).　Additionally,　two　processing　schemes　in　the　recursive　method　analysis　were　tried　to　determine　the　optimal　feature　subset　based　on　corresponding　algorithms,　namely　recursive　feature　elimination　(RFE)　and　recursive　feature　elimination　with　cross　validation　(RFECV).　Seven　factors　were　considered　in　the　feature　selection　process　based　on　collected　data,　including:　indoor　temperature,　outdoor　temperature,　relative　humidity,　concentrations　of　PM2.5,　air　quality　index　(AQI),　wind　speed　and　wind　direction　respectively.　The　results　showed　that　different　feature　subsets　can　generate　different　prediction　accuracy　within　the　recursive　method.　RFECV　can　determine　the　most　appropriate　feature　subset　effectively　with　the　consideration　of　the　correlation　among　various　factors.　Both　LR　and　SVC　were　proved　to　be　effective　as　estimators　embedded　in　RFECV,　however　SVC　is　more　computationally　expensive　and　LR　shows　a　larger　variance　within　the　feature　subset　space.　RF,　as　a　non-recursive　method,　demonstrated　real　advantages　in　eliminating　redundant　features　compared　to　the　recursive　feature　selection　process.　©　2018　Proceedings　of　10th　Windsor　Conference:　Rethinking　Comfort.