Background:　To　reduce　calibration　time　of　brain–computer　interface　(BCI)　or　even　implement　zero-training　BCI,　researchers　have　been　studying　how　to　effectively　apply　transfer　learning　in　the　field.　In　order　to　thoroughly　investigate　the　performance　of　transfer　learning　in　BCI　and　the　key　factors　affecting　transfer　performance　in　the　field,　we　carried　out　a　comprehensive　study.　New　method:　In　general,　transferring　knowledge　in　BCI　is　implemented　in　two　ways:　ensemble　or　pool.　In　this　work,　we　propose　two　different　transfer　approaches.　One　is　to　transfer　the　information　of　all　channels　as　a　whole　from　the　source　subjects　to　a　target　subject.　The　second　approach　is　to　transfer　the　information　of　corresponding　channels　between　the　subjects.　A　subject　transfer　framework　is　built　by　combining　the　two　approaches　with　ensemble　or　pool.　Results:　We　investigated　the　performances　of　eight　implementations　of　this　framework　on　a　data　set　acquired　by　an　interpersonal　interaction　(Chicken　Game)　experiment　based　on　code-modulated　visual　evoked　potential　(c-VEP)　BCI.　The　results　show　that　transfer　learning　generally　provides　acceptable　classification　performance.　Additionally,　an　in-depth　analysis　reveals　that　a　target　subject　usually　shares　different　brain　signal　distribution　with　different　source　subjects.　In　fact,　this　is　a　hypothesis　usually　implied　by　this　kind　of　research.　Conclusions:　Transfer　learning　for　c-VEP　BCI　can　be　qualified　for　reducing　calibration　time　or　starting　the　recognition　of　BCI　without　sufficient　subjects’　own　data.　In　addition,　our　finding　suggests　a　solid　validity　of　the　hypothesis　underlying　transferring　knowledge　in　BCI.　©　2020　Elsevier　B.V.