When　recognizing　facial　expression　sequences　by　discrete　Hidden　Markov　Model,　it　is　necessary　to　cluster　the　image　frames　into　several　observation　states.　Considering　the　complexity　of　face　images　in　emotion　expressing,　we　partition　the　whole　face　into　several　sub-regions　to　do　clustering　separately.　Then　the　categorical　values　from　clustering　results　of　each　subregion　are　combined　to　do　the　further　clustering.　In　the　paper,　we　propose　an　improve　K-mode　clustering　algorithm　called　K-frequency　to　do　the　category　data　clustering.　Instead　of　using　a　simple　0-1　method　to　determine　the　similarity　between　different　samples,　the　K-frequency　statistics　the　frequency　of　each　categorical　values　of　an　image　in　a　given　cluster　and　sums　them　as　the　similarity　between　a　sample　with　this　cluster.　Experiment　results　on　CK+database　show　that　the　bi-level　clustering　method　with　K-frequency　algorithm　outperforms　the　Single-layer　clustering　using　K-means　and　bi-level　clustering　with　K-mode.　The　improved　k-modes　algorithm　K-frequency　is　more　robust　than　the　original　k-mode　algorithm　in　dealing　with　isolated　points.