（五）规则化和模型选择.pdfVIP

瑙勫垯鍖栧拰妯″瀷閫夋嫨 锛圧egularization and model selection 锛 JerryLead csxulijie@ 2011 骞? 鏈?4 鏃ユ槦鏈熷洓 1 闂 妯″瀷閫夋嫨闂锛氬浜庝竴涓涔犻棶棰橈紝鍙互鏈夊绉嶆ā鍨嬮€夋嫨銆傛瘮濡傝鎷熷悎涓€缁勬牱鏈偣锛屽彲浠ヤ娇 鐢ㄧ嚎鎬у洖褰?y = 饊寖肀 肀? 锛屼篃鍙互鐢ㄥ椤瑰紡鍥炲綊(y = 饊寖肀 肀 1~肀 ) 銆傞偅涔堜娇鐢ㄥ摢绉嶆ā鍨嬪ソ鍛 锛堣兘澶 鍦ㄥ亸宸拰鏂瑰樊涔嬮棿杈惧埌骞宠　鏈€浼橈級锛 杩樻湁涓€绫诲弬鏁伴€夋嫨闂锛氬鏋滄垜浠兂浣跨敤甯︽潈鍊肩殑鍥炲綊妯″瀷锛岄偅涔堟€庝箞閫夋嫨鏉冮噸w 鍏紡閲 鐨勫弬鏁跋 锛 褰㈠紡鍖栧畾涔夛細鍋囪鍙€夌殑妯″瀷闆嗗悎鏄 = *饊亐 , 饊亐 , 鈥? 饊亐 + 锛屾瘮濡傛垜浠兂鍒嗙被锛岄偅涔圫VM銆乴ogistic 1 2 饊亼 鍥炲綊銆佺缁忕綉缁滅瓑妯″瀷閮藉寘鍚湪M 涓€ 1 浜ゅ弶楠岃瘉锛圕ross validation 锛 鎴戜滑鐨勭涓€涓换鍔″氨鏄浠嶮 涓€夋嫨鏈€濂界殑妯″瀷銆 鍋囪璁粌闆嗕娇鐢⊿ 鏉ヨ〃绀 濡傛灉鎴戜滑鎯充娇鐢ㄧ粡楠岄闄╂渶灏忓寲鏉ュ害閲忔ā鍨嬬殑濂藉潖锛岄偅涔堟垜浠彲浠ヨ繖鏍锋潵閫夋嫨妯″瀷锛 1銆佷娇鐢 S 鏉ヨ缁冩瘡涓€涓€亐 锛岃缁冨嚭鍙傛暟鍚庯紝涔熷氨鍙互寰楀埌鍋囪鍑芥暟鈩 銆傦紙姣斿锛岀嚎鎬фā鍨 肀 肀 ( ) 肀 涓緱鍒梆€寖 鍚庯紝涔熷氨寰楀埌浜嗗亣璁惧嚱鏁扳剮 x = 饊寖 肀 锛 肀 饊寖 2銆侀€夋嫨閿欒鐜囨渶灏忕殑鍋囪鍑芥暟銆 閬楁喚鐨勬槸杩欎釜绠楁硶涓嶅彲琛岋紝姣斿鎴戜滑闇€瑕佹嫙鍚堜竴浜涙牱鏈偣锛屼娇鐢ㄩ珮闃剁殑澶氶」寮忓洖褰掕偗瀹 姣旂嚎鎬у洖褰掗敊璇巼瑕佸皬锛屽亸宸皬锛屼絾鏄柟宸嵈寰堝ぇ锛屼細杩囧害鎷熷悎銆傚洜姝わ紝鎴戜滑鏀硅繘绠楁硶濡備笅锛 1銆佷粠鍏ㄩ儴鐨勮缁冩暟鎹 S 涓殢鏈洪€夋嫨 70%鐨勬牱渚嬩綔涓鸿缁冮泦饊亞train 锛屽墿浣欑殑 30%浣滀负娴嬭瘯闆 饊亞CV 銆 2銆佸湪饊亞 涓婅缁冩瘡涓€涓€亐 锛屽緱鍒板亣璁惧嚱鏁扳剮 銆 train 肀 肀 3銆佸湪饊亞 涓婃祴璇曟瘡涓€涓剮 锛屽緱鍒扮浉搴旂殑缁忛獙閿欒饊寑虃 (鈩 ) 銆 CV 肀 肀咰V 肀 4 銆侀€夋嫨鍏锋湁鏈€灏忕粡楠岄敊璇€寑虃 (鈩 ) 鐨勨剮 浣滀负鏈€浣虫ā鍨嬨€ 肀咰V 肀 肀 杩欑鏂规硶绉颁负hold-out cross validation 鎴栬€呯О涓虹畝鍗曚氦鍙夐獙璇併€ 鐢变簬娴嬭瘯闆嗘槸鍜岃缁冮泦涓槸涓や釜涓栫晫鐨勶紝鍥犳鎴戜滑鍙互璁や负杩欓噷鐨勭粡楠岄敊璇€寑虃 (鈩 )鎺 肀咰V 肀 杩戜簬娉涘寲閿欒锛坓eneralization error 锛夈€傝繖閲屾祴璇曢泦鐨勬瘮渚嬩竴鑸崰鍏ㄩ儴鏁版嵁鐨?/4-1/3銆?0%鏄 鍏稿瀷鍊笺€ 杩樺彲浠ュ妯″瀷浣滄敼杩涳紝褰撻€夊嚭鏈€浣崇殑妯″瀷