Aiming　at　the　hybrid　algorithm　I-B　and　B-MDL,　an　improved　method　is　proposed.　Firstly,　it　uses　order-0　and　partial　order-1　independence　tests　to　obtain　an　original　Bayesian　network　structure.　This　reduces　the　number　of　independence　tests　and　database　passes　while　effectively　restricting　the　search　space.　Secondly,　it　takes　mutual　information　between　nodes　as　heuristic　knowledge　perform　sort　order　for　candidate　parent　nodes,　which　increases　the　cut-offs　of　B　and　B　search　trees　and　accelerates　the　search　process.　The　experimental　results　on　the　currency　database　show　that　the　modified　algorithm　is　quicker　than　some　hybrid　algorithms　while　keeping　a　high　accuracy,　and　it　can　handle　large　data　sets.