Exploring　new　thermal　preference　prediction　models　or　methods　to　precisely　analyze　occupants＇　unconscious　feedback　on　the　thermal　environment　without　disturbing　them　is　essential　for　increased　building　efficiency,　comfort　and　productivity.　In　this　study,　we　propose　a　novel　method　for　developing　a　comfort　model　that　uses　occupant　self-adjusting　behavior　to　predict　thermal　preference.　The　model　development　draws　from　field　data　including　thermal　self-adaptive　behaviors,　environmental　parameters,　and　thermal　preferences　collected　from　34　occupants　in　a　single　multi-occupancy　room　in　a　research　office　in　a　university　in　south-east　China,　and　it　employs　four　machine　learning　algorithms　including　Support　Vector　Machine,　Random　Forest,　K-Nearest　Neighbor,　and　Decision　Tree.　The　results　indicate　that　there　is　a　close　relationship　between　thermal　preference　and　thermal　self-adaptive　behavior.　The　highest　prediction　accuracy　of　thermal　preference　using　a　single　set　of　input　parameters　is　0.81,　obtained　using　self-adaptive　behaviors.　When　the　input　parameters　increase　to　four　sets　or　more,　the　highest　median　accuracy　of　thermal　preference　prediction　(0.85)　does　not　change　significantly.　When　it　is　difficult　to　obtain　personal　information　(gender,　height,　and　weight)　and　data　on　clothing　thermal　resistance,　it　is　recommended　to　use　a　combination　of　three　sets　of　parameters　(self-adaptive　behavior,　indoor　temperature　and　relative　humidity　(T/RH),　and　outdoor　T/RH)　or　a　combination　of　four　sets　of　parameters　(self-adaptive　behavior,　indoor　T/RH,　outdoor　T/RH,　and　globe　temperature)　as　the　input　parameters　of　the　thermal　preference　prediction　model　based　on　Random　Forest,　which　results　in　an　expected　prediction　accuracy　of　0.84　and　0.85,　respectively.