Share
Embed
Tree-based Algorithm by Python
Tree-basedアルゴリズムのPythonでの実装について取り扱った、下記の4章の内容を公開します。
https://lib-arts.booth.pm/items/2858666
基本的な理論の理解に基づいた平易な実装を行うにあたって重要と思われる事項についてまとめました。
LiberalArts
April 03, 2021
Transcript
38 slides
-
5 7 痥 4 畍 㹋鄲 睗 4 皹ךעյTree-based ٜؓإٛثّס㵅鍮מחַי⺅䪒ַױ
ն霄箖ס㵅鍮ע Jupyter نٜؒؕםל⩧מ牞霼ג偙ֿ־ ַסך׆׆ךע⺅䪒յ睗 2 皹睗 3 皹ס⫐㵼־ꓨ锡ם掾 Python ך牞霼כַֹ䓺䑑⺅גַכ䘼ַױն♧┫յ4-1 硼ך睗 4 皹ס阾鼥偙ꓹמחַיױכ״גסהמյ4-2 硼ך尴㴻勎ס圸 碎מֵגזיס㕈勓⭦槏מחַיյ4-3 硼ךؓ٤ئ٤هٜ㳔肪מחַ יאב⺅䪒ַױն 4.1 痥 4 畍ך鎸鯹ח֮גך倯ꆙ 4-1 硼ךע睗 4 皹ס阾鼥מֵגזיס偙ꓹמחַי祔ⷃמׇ鞃僻ױն כַֹסյ րحٞ־✑ DeepLearningցסֹמ槏韢㵅鍮־牞霼 剹禶עַׂח־⮂曫יַױֿյ㵅鍮ס鼥偙ע־ם詇脢嬐מꇙ ַֿ⮂ַכ䘼ג״ךն ┞沁㝕׀ם脝ֻ偙סꇙַכיעր؛هةؘؠع䭰⻔סֹםوٞءِٚ ٤ء䪫岺לסַׂ氠ַ־ցֿ䮕ׅסךעםַ־כ瞉脢ע脝ֻי ֽյ ր勓倀╚מע؛هةؘؠع䭰⻔ך剹ׂ׀ךעםַցכַֹסֿ瞉脢 57 -
5 8 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ס锶闋ךնאסֹמ瞉脢ֿ脝ֻ槏氮כיעյ ր勓倀╚מ؛هةؘؠ
ع䭰⻔⯼䳀כ㵅鍮鼥כ牞霼ס䩘ꪨֿ⺅ג״ցךն㵅鍮 ךֵףꫀꅙؤ٭غ䱱דׄךײױֿյ剹禶ס㖪⻉ע鞅ײꂉ⯼䳀 כꇃׁכ鞅׳⣨ס鬘䬎ֿ㝕׀ַכ䘼ױն 杅מ Tree-based סٜؓإٛثّמꫀיעյゼ꾴㴻聋䩘岺מ㖪⻉ ⮔ֿׄ㝂ׂյ؛هةؘؠع䭰⻔ךױכ״⯼䳀ך闋鞃ꅼ״כبتطّ陭 阛מꫀ锡筶ֿ䒣ׂםյٜؓإٛثّס槏闋מ겏╚ך׀םׂםסךע כַֹ䧂䗻ֵֿױն ꫀꅙ㵅鍮䱱갾ע㵅갾סؤ٭غ聁־䱱偙ֿ僃ַכ䘼ג״յط ؞تعמֽׄ闋鞃כיעꓨ锡◜꽃ⷃ✄ך雧ֹם䓺䑑כיյ㵅 鍮牞霼갾ע⯁㛿ס Jupyter ׇ牞霼ַגדׂכַֹ䓺䑑ך勓剹ע鞃 僻鉿ֹֽכ䘼ַױն㵅갾עꓨ锡⭦槏עٓةٖ٭ٜⵊסֿ┞薭溷ך ֿյوٞءِٚ٤ء篑닫ס劔摾מ䈼ֿ⮂סכյⷃ✄ך雧偙ֿ槏韢 ס槏闋מחםַׅסךյ׆׆ךע⭦槏ⷃ✄牞霼׆ככױגն כַֹ׆כך䓜ط؞تع⫐ךע㕈勓⭦槏אסױױ⮗⮂ױגֿյٓ ةٖ٭ٜⵊמֵגזיעאלٜؓإٛثّٝيٜךס㜟剳◜꽃עם ַסךյ㵅갾מٚؕهٚٛ❈氠갾מ焒锶כי䔢מ皑חסךעםַ ־כ䘼ַױնױגյ峜䟨◜꽃כיյ׆׆ךס㵅鍮ע槏韢ס槏闋潨溷כ סךֵյ㵅鍮סⲯ椙םלעאל脝䢩יַױ؆סך׆ס掾ע ׇ峜䟨ׂדַն䌐儅ם溪䞯מ㕈טַי䌐儅ם㵅鍮鉿ֹכַֹ䟨㎫ך㵅鍮 鉿ַױגסךյⲯ椙溷ם㵅鍮鉿ַגַכַֹ偙ע scikit-learn 㢼 ״յ⻄甦ٚؕهٚٛס㵅鍮ׇ牞霼ַגדׄגכ䘼ַױն 4.2 寸㹀加ך㛇劤Ⳣ椚 4-2 硼ךע尴㴻勎ס㵅鍮מֵגזיס㕈勓⭦槏מחַי牞霼鉿ַױն ♧┫յ4-2-1 硼ך⮔꿔כ㎇䊟סゼ꾴陭㴻כי氠ַظ٭ذجشعס牞霼 鉿ַյ籽ׂ 4-2-2 硼ךؙ٤عٞم٭յةؼ➳俙յ◝▗⾔靯䈼םלס雄❿㕈徙 (criteria) מחַיյ4-2-3 硼ךע杅䖇ꓪסخ٭عמחַי牞霼鉿ַױն 58 -
5 9 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ױגյ4-2-4 硼ךעؿ٭غס⮔ⰺמחַיյ4-2-5
硼ךע勎圸碎יַׂמ ֵגזיס⫐ꌃס枱䡢ס⟛䭥מחַיյ4-2-6 硼ךע㳔肪ג勎מ㕈טַג ◙廠מחַיאב⺅䪒ַױն 4.2.1 ⴓ겲ה㔐䌓ך㉏겗鏣㹀 4-2-1 硼ךע⮔꿔כ㎇䊟סゼ꾴陭㴻ס䪻䳢מֵגזיյئ٤وٜכי氠 ַظ٭ذجشعס牞霼鉿ַױն⮔꿔㎇䊟䩘꼿םظ٭ذجشع氠 ַףכַֹ׆כךלה scikit-learn מ埉徙㵅鍮סظ٭ذجشع 氠ַ׆ככױגն scikit-learn מ♕㺲ס╚ך⺅䪒ַׂ劔⻏םסכַֹ׆כךյ⮔ 꿔מחַיע irisյ㎇䊟מחַיע Boston Housing אב氠ַױն ♧┫յאבסظ٭ذجشعסٞ٭غםלמחַי祔ⷃמ牞霼ױ ױע iris dataset מחַי牞霼ױն 1: import numpy as np 2: from sklearn import datasets 3: 4: iris = datasets.load_iris() 5: X = iris.data 6: y = iris.target 7: 8: print(X.shape) 9: print(y.shape) 59 -
6 0 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.1:
iris dataset Ώ ┪阾ךעظ٭ذجشعסٞ٭غכ鉿⮬סئؕثס牞霼鉿ַױגն scikit-learn כ⻎免מ NumPy 鞅ײꁎ؆ךַױֿյ┞ꅙס鉿⮬亣✑מ䖩 锡םסך⩰מ鞅ײꁎ؆דכׇ槏闋ַגדׄגכ䘼ַױն 1: import numpy as np 2: from sklearn import datasets 3: 4: iris = datasets.load_iris() 5: X = iris.data 6: y = iris.target 7: 8: print(X.shape) 9: print(y.shape) 60 -
6 1 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.2:
iris dataset ΐ ┪阾ךע (X, y) מ㸐יىشر٭ 5 鉿ס⮂ⲇכյאבסꏕ⮬מױ 㜟俙ס㒘מחַי牞霼鉿ַױגն 籽ַי Boston Housing מחַי牞霼ױն 1: iris = datasets.load_boston() 2: X = iris.data 3: y = iris.target 4: 5: print(X.shape) 6: print(y.shape) 7: print(X[:2,:]) 8: print(y[:5]) 9: print(X.dtype) 10: print(y.dtype) 61 -
6 2 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.3:
Boston Housing dataset ㎫ 4.3 ךע鉿⮬סئؕثכىشر٭鉿ס牞霼鉿ַױגնئ٤وٜظ٭ ذמחַי⺅䪒ֹ갾ע׆סֹמ鉿⮬סئؕثכىشر٭ס䝠㖥牞霼 כ⪢✄ס䪻䳢ֿ僃ׂםסךֽ״ךն 4.2.2 鐰⣣㛇彊 (criteria) 4-2-2 硼ךע雄❿㕈徙 (criteria) סؙ٤عٞم٭յةؼ➳俙յ◝▗⾔靯䈼 סאבס㵅鍮מחַי⺅䪒ַױն׆׆ךؙ٤عٞم٭כةؼ➳俙ע ⮔꿔ゼ꾴⯼䳀כג״ iris յ◝▗⾔靯䈼ע㎇䊟ゼ꾴⯼䳀כג״ Boston Housing אב氠ַױն 62 -
6 3 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ؒٝزٗؾ٦ ؙ٤عٞم٭מחַי牞霼מֵגזיעյf(x)
= −xlog(x) − (1 − x)log(1 − x) סꫀ俙ס 0 < x < 1 מֽׄءٚن䫅ֻכ虘ַ־כ䘼ַ ױն 1: import matplotlib.pyplot as plt 2: %matplotlib inline 3: 4: x = np.arange(0.01, 1., 0.01) 5: y1 = -x*np.log(x) - (1-x)*np.log(1-x) 6: y2 = -x*np.log2(x) - (1-x)*np.log2(1-x) 7: y3 = (-x*np.log(x) - (1-x)*np.log(1-x))*1.2 8: 9: plt.plot(x, y1, color="green") 10: plt.plot(x, y2, color="blue") 11: plt.plot(x, y3, color="red") 12: plt.show() 63 -
6 4 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.4:
ؙ٤عٞم٭ (Entropy) ┪阾牞霼ג갾מյ┪מ⭾םꫀ俙ךֵ׆כֿ牞霼ך׀ױն׆ֿ䟨 ⽱׆ככיյ2 ؠٚت⮔꿔ס갾סؙ٤عٞم٭ע 1 2 ך⻄ؠٚتמⰺ 䓜י갾מ劄㝕כם׆כֿ־ױ3 ؠٚت♧┪ך㐬瞏מؠٚ ت⮔꿔㖪⻉סؙ٤عٞم٭ֿ┞沁㝕׀ַך նױגյ׆ס갾מؙ٤ عٞم٭ס阛砯מ氠ַ㸐俙ꫀ俙ס䍏מחַי宜מם偙ֽכ䘼ַ ױֿյ䍏ס㜟䳕⪪䑑氠ַ׆כך (㴻俙) (㜟䳕䔿ס㸐俙ꫀ俙) כ ׆כֿך׀ג״յ2 氠ַֹֿ e 氠ַֹֿأؕ٤阛砯尴㴻勎 64 -
6 5 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ס㳔肪מַֽיע篙卸ע㜟ױ؆ն㎫מַֽיյ귱ֿ䍏ֿ 2
ס㖪⻉յ糽 ֿ䍏ֿ e ס㖪⻉ךֿյ鰱ך銨ג糽ס 1.2 ⠨ֿ⡑鞪俠ף귱מꓨם כַֹסֿ׆׆ךסَؕ٤عךն ֹ㸴⪽✄溷מ牞霼ך׀ףכַֹ׆כך iris מחַי牞霼ױն − 3 k=1 pklog(pk ) ס俙䑑מ㕈טַיյiris סؙ٤عٞم٭ס阛砯鉿ַױն 1: _, y_counts = np.unique(y, return_counts=True) 2: res_criterion = 0 3: for i in range(0,y_counts.shape[0]): 4: ratio = y_counts[i]/y.shape[0] 5: res_criterion -= ratio*np.log(ratio) 6: 7: print(y_counts) 8: print(res_criterion) ㎫ 4.5: iris סؙ٤عٞم٭ (Entropy) ㎫ 4.5 ע iris סؙ٤عٞم٭阛砯ג篙卸מםױն׆׆ך峜䟨ֿ䖩 65 -
6 6 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 锡םסעؙ٤عٞم٭ך氠ַ㸐俙ꫀ俙ךյ䑛俙ס ratio
ֿ 0 ס㖪⻉㸐俙ꫀ 俙ס㴻聋㔔־㜽ױն┞䗎 xlog(x) ס +0 ס嚋ꮹע 0 םסךյ0 כי 虘ַסךֿյوٞءِٚ٤ءמֵגזיעؙٚ٭⭦槏ֿ䖩锡מםױ նאסג״յ׆סꁊמחַי宜מגׂםַ㖪⻉ע姌꽃ך⺅䪒ֹةؼ➳ 俙ס偙ֿ꽒⣌ךה؆ٚؕهٚٛךؙٚ٭⭦槏ע氠䟨יַס ךյⷃמ scikit-learn םל❈ֹ갾מע׆סꁊעאל宜מםׂי虘ַ כ䘼ַױ ն آص⤘侧 ةؼ➳俙ס槏闋מֵגזיע f(x) = 1 − x2 − (1 − x)2 = 2x(1 − x) = −2(x − 1 2 )2 + 1 2 סꫀ俙ס 0 < x < 1 מֽׄءٚن䫅ֻסֿ虘ַ־כ 䘼ַױն 1: import matplotlib.pyplot as plt 2: %matplotlib inline 3: 4: x = np.arange(0.01, 1., 0.01) 5: y = 1-x**2-(1-x)**2 6: 7: plt.plot(x, y, color="green") 8: plt.show() 66 -
6 7 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.6:
ةؼ➳俙 (Jini Index) ㎫ 4.6 ך牞霼ך׀ֹמةؼ➳俙עؙ٤عٞم٭כ⻎坎յ┪מ⭾םꫀ俙 כםױն㎫ 4.4 כ锶嬟כ㕈勓溷מעֵױ㜟םַךֿյ㎫ 4.4 ס偙ֿ x = 0.3 x = 0.7 סֹם劄㝕⡑כ劄㸯⡑ס╚ꪨמֵ⡑ֿ嬟 鼛溷㝕׀ׂם׆כֿ牞霼ך׀ױն׆סꁊסꫀ俙מחַיעյֵ瓦䍲槏 韢溷ם㸬⮂םלꓨ锡ךֿյֵױ鞪ꇃׁכ㝕㜟םסךյⷃמءٚن 䲾ַיْؕ٭ة䲖׳כַֹסꓨ锡ךն ױגյؙ٤عٞم٭כ⻎坎מ iris מꫀיةؼ➳俙阛砯יײױն 67 -
6 8 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 1: import
matplotlib.pyplot as plt 2: %matplotlib inline 3: 4: x = np.arange(0.01, 1., 0.01) 5: y = 1-x**2-(1-x)**2 6: 7: plt.plot(x, y, color="green") 8: plt.show() ㎫ 4.7: iris סةؼ➳俙 (Jini Index) ㎫ 4.7 ע 1 ־ 1 3 ס◝▗ךֵ 1 9 3 ㎇䑛ַג⡑כי槏闋כ虘 ַכ䘼ַױն׆׆ך阛砯篙卸ֿ 0.66...65 כ⮂יַסע阛砯┪⮂靯䈼 ךն⸮槏ס槏闋ס嫘갧ךע㕈勓溷מעאל宜מםׂי虘ַךֿյ ה؆כג㵅鍮鉿ֹ갾ע׆סꁊעֵ瓦䍲宜מ䖩锡ֵֿױն ✳⛦ㄤ铎䊴 ◝▗⾔靯䈼ס槏闋מֵגזיעյⷃ㎇䊟⮔卥םלס劄㸯◝▗岺כ⻎坎מ脝 ֻיֽׂכ虘ַ־כ䘼ַױն♧┫յBoston Housing מꫀי◝▗⾔靯䈼 68 -
6 9 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 阛砯יײױն 1:
res_criterion = np.mean((y-np.mean(y))**2) 2: 3: print(res_criterion) ㎫ 4.8: ◝▗⾔靯䈼 ◝▗⾔靯䈼מꫀיע㎫ 4.8 סֹמ NumPy ס mean ْخشغ氠ַ ׆כךب٤وٜמ㵅鍮鉿ֹ׆כֿך׀ױն 4.2.3 暴䗙ꆀךا٦ز 4-2-3 硼ךע杅䖇ꓪסخ٭عמחַי⺅䪒ַױն׆ס需ע┞锶册ꄼמ خ٭عף虘ַדׄסֹמ锶ֻױֿյאבס㜟俙מחַיخ٭ع יאמ㸐䗎潨溷㜟俙⺅⮂䖩锡䙎յ勎ס⮔ⰺ鉿ֹמֵגז יסئ٤وٜס硄槏ס䖩锡䙎ꤥײכ需עם־ם־鏿겧מםױն ׆ס㵅槁מֵגזיյNumPy ס argsort סֹם塌茣氠ַיյ杅䖇ꓪ ס꽄沁ס Index 䕑יאס Index 氠ַיئ٤وٜ־ꌃ⮔겏⻉✑䧯 כַֹ偙岺ֿ┞ח脝ֻױסך׆׆ךע׆ס偙岺מחַי牞霼ױ ն鞃僻דׄ锶י־טַכ䘼ג״յ♧┫ iris 氠ַי祔ⷃמ 牞霼יײױն 69 -
7 0 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 1: print(X[:10,0])
2: print("-------") 3: print(np.argsort(X[:10,0])) ㎫ 4.9: argsort(NumPy) ┪阾ע־ַֹמ iris ס 1 ⮬潨מ㸐յ┪־ 10 ⠥סئ٤وٜ 䑛ז䒟זיאבס⡑ס僰꽄מ Index חׄג䓺מםױն׆ה עր9(8+1) 沁潨ֿ 4.4 ך┞沁㸯ׂյ4(3+1) 沁潨ֿ 4.6 ךאס姌מ㸯 ׂյ....յ6(5+1) 沁潨ֿ 5.4 ך┞沁㝕׀ַցסֹמ槏闋ף虘ַךն ׆ֿך׀כյ1 ⮬潨ס杅䖇ꓪס㸯ַ꽄מ 5 חئ٤وٜ⺅⮂םל ס⭦槏ֿ⺪茣מםױն 1: x1 = X[:10,0] 2: asc_idx = np.argsort(x1) 3: 4: print(asc_idx[:5]) 5: print(X[asc_idx[:5], :]) 6: print(y[asc_idx[:5]]) 70 -
7 1 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.10:
argsort(NumPy) ┪阾סֹמ鉿ֹ׆כךյ1 ⮬潨ס杅䖇ꓪס⡑מ㕈טַי僰꽄מ X y ס ⡑䕑׆כֿ⺪茣מםױն 4.2.4 ؎ٝך鎘皾הظ٦سךⴓⶴ 4-2-4 硼ךעؿ٭غס⮔ⰺמחַי牞霼ױնخ٭عג杅䖇ꓪמחַ יأؕ٤ס阛砯鉿ַ訪ؿ٭غס⮔ⰺ鉿ַױն杅䖇ꓪמطإٛ㜟俙ס ײ氠ַ ID3 כע沌םյꅙ籽㜟俙脝䢩 C4.5 CART מחַי ע㕈勓溷מ 4-2-3 硼סخ٭ع䔿ס篙卸氠ַי糹䓜גך阛砯鉿ֹ䖩锡ֿ ֵױն ♧┫յةؼ➳俙氠ַג iris ס⮔꿔ゼ꾴❛מ㵅鍮מחַי牞霼鉿זי 鉿׀ױն 71 -
7 2 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 1: def
calc_criteria(y): 2: _, y_counts = np.unique(y, return_counts=True) 3: res_criterion = 0 4: for i in range(0,y_counts.shape[0]): 5: ratio = y_counts[i]/y.shape[0] 6: res_criterion += ratio*(1-ratio) 7: return res_criterion 8: 9: print(calc_criteria(y)) ㎫ 4.11: ةؼ➳俙ס阛砯 ױ iris ס⩧ظ٭ذמֽׄةؼ➳俙ס阛砯鉿ַױגն㎫ 4.11 ך篙 卸ֿ 0.6666... כםױגֿյ׆ע┫阾סֹמ㸬⮂ך׀ױն 1 − 50 150 2 − 50 150 2 − 50 150 2 = 1 − 1 3 = 2 3 4-1 硼ך闑גֹמ闋鞃ס갾ע⺪茣םꮹ؛هةؘؠع䭰⻔ꫀ 俙氠ַםַך闋鞃鉿ֹֹמיַסךֿյ׆׆ךעֵֻי calc_criteria ꫀ俙ס㵅鍮鉿ַױגն׆ס槏氮כיעյأؕ٤ס阛 72 -
7 3 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 砯מֵגזי⮔ⰺ⯼כ⮔ⰺ䔿ך⻎⭦槏鉿ַյ䈼阛砯ג״ךն מյ׆杅䖇ꓪס⡑ס俙דׄ繪ꂉסךյםֽꫀ俙ⵊי⫙⯈氠
ַ䓺מיֽׂ䖩锡ֵֿױն ׆׆ױךך criteria ס阛砯מחַיע牞霼ך׀גסךյ姌מأؕ٤ס阛砯 מחַי牞霼ױն 1: desc_index = np.argsort(X[:,0]) 2: gini_before = calc_criteria(y) 3: gini_after = ((1/2)*calc_criteria(y[desc_index[:X.shape[0]//2]]) \ 4: +(1/2)*calc_criteria(y[desc_index[X.shape[0]//2:]])) 5: gain = gini_before - gini_after 6: print(gain) ㎫ 4.12: أؕ٤ס阛砯 4-2-3 硼ך闑ג NumPy ס argsort 氠ַי杅䖇ꓪ X ס 1 ⮬潨ס⡑ خ٭عյאס⡑מ㕈טַיئ٤وٜ◝⮔ⰺג갾ס gini ➳俙阛砯յ ׆מ㕈טַיأؕ٤阛砯יַױն 1: desc_index = np.argsort(X[:,0]) 2: gini_before = calc_criteria(y) 73 -
7 4 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 3: res_gain
= np.zeros([X.shape[0]-1]) 4: for i in range(0, X.shape[0]-1): 5: gini_after = (((i+1)/150)*calc_criteria(y[desc_index[:i]]) \ 6: +((150-i)/150)*calc_criteria(y[desc_index[i:]])) 7: res_gain[i] = gini_before - gini_after 8: 9: print(np.max(res_gain)) 10: print(np.argmax(res_gain)) ㎫ 4.13: أؕ٤ס劄㝕⡑ס阛砯 ׆כ⻎坎ם⭦槏♑ס杅䖇ꓪמꈌ氠㵅鉿׆כךյ訪ؿ٭غס⮔ ⰺ鉿ַױնױג׆ס갾מյ4-2-5 硼ך祔ⷃמ闑ױֿյ⪢יס訪ؿ٭ غמחַי⻎坎ס阛砯鉿ַ劄㝕⡑阛砯䖩锡ֵֿסךյ׆סꁊס㵅 鍮עם־ם־鏿겧מם׆כמ峜䟨ֿ䖩锡ךն 74 -
7 5 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 4.2.5 粸鵤׃怴皾חֶֽ朐䡾ך⥂䭯*
4-2-5 硼ךע繪ꂉ悍砯מֽׄ枱䡢ס⟛䭥מחַי⺅䪒ַױն 虝չם偙岺ֵֿכ䘼ַױֿյ瞉脢עםׂب٤وٜמ㵅鍮ג־זג סךյظ٭ذجشع✑䧯ג訪ؿ٭غמⰺ䓜ייַֹׂם㵅鍮鉿ַ ױגն♧┫אס⯼䳀ךꀐ剹氠ַג枱䡢ס⟛䭥מחַי牞霼ױն 1: leaves = {"T1":(X, y)} 2: print(type(leaves)) 3: print(leaves["T1"][0][:3, :]) 4: print(leaves["T1"][1][:3]) ㎫ 4.14: 訪ؿ٭غכ iris dataset ┪阾סֹמ訪ס ID keyյאמ㸐䗎 (X, y) value כ׆כ ך硄槏ֹכַֹ偙ꓹךն׆סֹמ׆כך leaves מ㸐յ "T1"ס ֹם key 氠ַי X כ y ס⡑אב⺅⮂׆כֿ⺪茣מםױն 姌מ訪ؿ٭غסꃯⲎס갾ס⭦槏מחַי牞霼ױն 75 -
7 6 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 1: leaves
= {"T1":(X, y)} 2: print(type(leaves)) 3: print(leaves["T1"][0][:3, :]) 4: print(leaves["T1"][1][:3]) ㎫ 4.15: 訪ؿ٭غסꃯⲎ ꃯⲎؿ٭غעאױךסؿ٭غס⮔ⰺכי銨槁סךյ⮔ⰺ ؿ٭غמ㸐䗎ꀐ剹ס锡筶 del 氠ַי巆⹛յ偆גמ◝חס訪ؿ٭ غסꃯⲎ鉿ַױն㎫ 4.15 ךע 1 沁潨ס杅䖇ꓪס⡑ךخ٭عյ75 ئ٤ وٜחiris ע 150 ئ٤وٜT1 כ T2 מ䮴⮔ׄ鉿ַױגն׆ס갾 מ print 氠ַג⮂ⲇ篙卸ס┞沁䈱ס⮬מ濪潨כյT1 ע 4.3 ־僰꽄 ך㢼ױסמ㸐յT2 ע 5.8 ־僰꽄ך㢼ױ׆כֿ牞霼ך׀סךյ⭦ 槏סْؕ٭ةֿח־ײַ־כ䘼ַױն 訪ؿ٭غסꃯⲎמֵגזיס⭦槏מחַיע牞霼յؿ٭غ⮔ⰺסْؕ٭ ةֿ䲖״ױגֿյמ׆׆ך宜מ䖩锡ֵֿסֿ勎ס圸ꅎמחַי 76 -
7 7 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ךնכַֹס訪ؿ٭غס⮔ⰺס䍲מאס訪ؿ٭غ巆յ2 ח訪ؿ٭غ
ꃯⲎֹם⭦槏דכյלסذِؕ٤ءךלֿꃯⲎג־ֿ־ םׂםյ篙卸勎ס圸ꅎֿ־םׂם־ךն׆ךע◙廠 (䱿韢) 鉿ֹ׆כֿך׀םׂםױն ׆סゼ꾴מ㸐יյ瞉脢עؿ٭غסꃯⲎ免מ㸐骭כםזגؿ٭غס䝠㖥 媘׆כך㸐⭦׆ככױגն ㎫ 4.16: 訪ؿ٭غסꃯⲎ 77 -
7 8 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.17:
訪ؿ٭غסꃯⲎ꽄כ勎ס圸ꅎ גכֻף㎫ 4.16 סֹמ⟛㲽鉿ֹדׄךյ㎫ 4.17 סֹם勎圸ꅎ⫙ 槁׆כֿך׀ױն㎫ 4.17 ס؛ٝ٤ةס訪ؿ٭غֿאב◙廠⡑מ 㸐䗎כַֹْؕ٭ةךׇ槏闋ַגדׄגכ䘼ַױն׆ס⭦槏ꇃ瓦מ חַיյ┫阾ך祔ⷃמ⪽✄ⵊיײױגն 1. T1 ⮔ⰺי T1 כ T2 ✑䧯 2. T2 ⮔ⰺי T2 כ T3 ✑䧯 3. T2 ⮔ⰺי T2 כ T4 ✑䧯 4. T3 ⮔ⰺי T3 כ T5 ✑䧯 5. T1 ⮔ⰺי T1 כ T6 ✑䧯 ⮔ⰺס꽄沁阾ꜗיֽׂדׄךյ┪阾סֹמ勎ס圸ꅎ⫙槁ך׀כ ַֹסْؕ٭ةַגדׄגסךעםַ־כ䘼ַױն 78 -
7 9 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.18:
⮔ⰺסꫲ⡑ס⟛㲽 ױגյ勎ס圸ꅎֿגל׆כךյ⮔ⰺס匛⚂מחַיע㎫ 4.18 סֹ מյ氠ַ杅䖇ꓪכאסꫲ⡑דׄ⟛䭥ף虘ַ׆כ־ױն 4.2.6 ✮庠* 4-2-6 硼ךע◙廠מחַי牞霼鉿ַױն4-2-5 硼ך⟛䭥ג枱䡢מ㕈 ט׀אבס訪ؿ٭غס⡑㳔肪免מ⟛䭥յא⩧מ◙廠鉿ַױ ն׆׆ךꓨ锡םסעյ4-2-5 硼ס leaves 氠ַ׆כ⯼䳀ככظ٭ ذجشعסئؕث⮔䊬מ⟛䭥䖩锡ֿ气յؓ٤ئ٤هٜ㳔肪鉿ֹ׆ כ脝ֻכ劳ױׂםַכַֹ׆כךն כַֹ׆כךյ4-2-5 硼ך牞霼ג㜟俙סֹה leaves ע⟛䭥םַך split_leaf כ thresholds מⲎֻיאבס訪ؿ٭غס⡑ס┩חדׄ氠ַ י◙廠׆כ脝ֻױնsplit_leaf כ thresholds מחַיע 4-2-5 硼ך 牞霼גסךյ4-2-6 硼ךעױ訪ؿ٭غס⡑מחַי牞霼鉿ַױն 1: desc_index = np.argsort(X[:,2]) 2: leaves = {"T1":(X, y)} 3: del leaves["T1"] 4: leaves["T1"] = (X[desc_index[:52], :], y[desc_index[:52]]) 79 -
8 0 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 5: leaves["T2"]
= (X[desc_index[52:], :], y[desc_index[52:]]) 6: 7: print(leaves["T1"][1]) 8: print(leaves["T2"][1]) ㎫ 4.19: אבס訪ؿ٭غס y ס⡑ ַ׀ם♣銨⡑דׄ牞霼כざ痒מםסךյ 3 沁潨ס杅䖇ꓪס 52 沁 潨ס⡑ך⮔ⰺ鉿זיײױגն㎫ 4.19 ־牞霼ך׀ֹמյ iris dataset מַֽי 3 沁潨ס杅䖇ꓪ氠ַי 150 沁潨כ 51150 沁潨מ⮔ׄ׆כ ךؠٚت 0 ֿ㴞櫼מ⮔꿔ך׀ֹךնכעַֻյֵױ㴞櫼מ⮔꿔鉿 ֻ杅媗ם塌⚶עאלםַג״յֵֻי 52 沁潨ס⡑ꫲ⡑מ陭ׄ׆ ככױגն 1: print(np.mean(leaves["T1"][1])) 2: print(np.mean(leaves["T2"][1])) 80 -
8 1 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.20:
訪ؿ٭غס♣銨⡑ (㎇䊟) 訪ؿ٭غס♣銨⡑寛״갾מע⮔꿔㎇䊟ס偙ֿ䌐㐬寛״דׄ םסך阛砯ֿب٤وٜךն ┞偙ך⮔꿔מֵגזיעյאבס訪ؿ٭غך┞沁㝂ַ⡑♣銨⡑כ יꈷש䖩锡ֵֿյ׆ס㵅鍮מֵגזיע┫阾סֹמ┞䩘ꪨ䖩锡מם ױն 1: cate_T1, y_counts_T1 = np.unique(leaves["T1"][1], return_counts=True) 2: cate_T2, y_counts_T2 = np.unique(leaves["T2"][1], return_counts=True) 3: 4: print(y_counts_T1) 5: print(y_counts_T2) 6: print(np.argmax(y_counts_T1)) 7: print(np.argmax(y_counts_T2)) 8: print(cate_T1) 9: print(cate_T2) 10: print("====") 11: print(cate_T1[np.argmax(y_counts_T1)]) 12: print(cate_T2[np.argmax(y_counts_T2)]) 81 -
8 2 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.21:
訪ؿ٭غס♣銨⡑ (⮔꿔) ㎫ 4.21 ךע NumPy ס unique ꫀ俙ס return_counts ؛وب٘٤ True מי㵅鉿鉿םזיַױն׆ס篙卸עյ1 ח潨סꂉ⡑כי겏阛 גطإٛյ2 ח潨סꂉ⡑כיאבס⡑ס⮂槁俙ֿ⮂ⲇױն אסג״ NumPy מֽׄ祔ⷃם겏阛ꫀ俙כ⻎坎מׇ槏闋ַגדׄכ虘 ַסךעםַ־כ䘼ַױն㎫ 4.21 ךע y_counts_T1 y_counts_T2 ֿאב劄㝕⡑⺅ؕ٤ظشؠت argmax 氠ַי⺅䕑յ׆ cate_T1 cate_T2 סؕ٤ظشؠتכי䭰㴻׆כך T1 ס♣銨⡑ס 0 כ T2 ס♣銨⡑ס 2 䕑׆כֿך׀ױն1 מחַיע T1 כ T2 מف ٚفٚמ䨾㺲׆ככםזגג״׆׆ךעלהס♣銨⡑כםם־ז גյכׇ槏闋ַגדׄגכ䘼ַױն 82 -
8 3 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.22:
訪ؿ٭غס♣銨⡑ס阾ꜗ ׆סֹמ阛砯ג♣銨⡑㎫ 4.22 סֹמ node_values כי阾ꜗ גכ脝ֻױն ׆׆ױךס⫐㵼ךյsplit_leafյthresholdsյnode_values ס┩חֿזג סךյ׆⩧מ◙廠鉿ֹ׆כמחַי♧┫牞霼יַ׀ױն◙廠מ ֽׄ勎ס㺤ꪛמֵגזיעאבס thresholds מֽׄ⮔㼜מ㸐יյ 4-2-5 硼ך牞霼ג leaves ⰺ䓜יסכ⻎坎ם⭦槏◙廠氠ס X מꈌ 氠ף虘ַסךֿյX ⮔ⰺיⰺ䓜ייַׂכئ٤وֿٜⰺ䓜י םַ訪ؿ٭غֵֿյ׆ס⺅䪒ַֿ㸴չ㝕㜟ךն׆ס㎇ꉌמֵגז יյ◙廠מַֽיעئ٤وٜס⮔ⰺךעםׂյTrue/False סؕ٤ظشؠت אבס訪ؿ٭غמⰺ䓜י偙ֿ祔ⷃךעםַ־כ䘼ױגն 1: init_idx = X[:, 0] > np.min(X[:, 0])-1 2: split_idx_left = X[:, 0] < 5. 3: split_idx_right = X[:, 0] >= 5. 4: 5: print(init_idx[:10]) 6: print(X[:10, 0]) 7: print(split_idx_left[:10]) 8: print(split_idx_right[:10]) 83 -
8 4 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.23:
韢槏悍砯Ώ ⮔ⰺמֵגזיע㎫ 4.23 סֹם⭦槏杅䖇ꓪסꫲ⡑מ㸐י鉿ֻף虘 ַךն2 鉿潨כ 3 鉿潨ס X ס⮬䭰㴻 (㎫ 4.23 ךע 0) כ嬟鼛俙㲻 (㎫ 4.23 ךע 5.) ס 2 חמ thresholds ס⡑ꈌ氠׆כך◙廠מֵגזיס 訪ؿ٭غס⮔ⰺֿ⺪茣ךն ׆ split_leaf מ㕈טַי 4-2-5 硼ס leaves כ⻎坎ם⭦槏鉿ֹכ 虘ַךֿյ訪ؿ٭غס⮴劻⡑ע⪢י True ךֵססյ訪ؿ٭غ⮔ⰺ 갾מ False ֿ⮂י匡ג״׆ס⺅䪒ַ脝ֻ䖩锡ֵֿױն 1: init_idx = X[:, 0] > np.min(X[:, 0])-1 2: split_idx_left = X[:, 0] < 5. 3: split_idx_right = X[:, 0] >= 5. 4: 5: split_right_1 = X[:, 0] < 5.2 6: split_right_2 = X[:, 0] >= 5.2 7: print(split_right_1[:10]) 8: print(split_right_2[:10]) 9: print("======") 10: print(split_idx_right[:10]) 84 -
8 5 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 11: print("======")
12: print(split_idx_right[:10]*split_right_1[:10]) 13: print(split_idx_right[:10]*split_right_2[:10]) ㎫ 4.24: 韢槏悍砯ΐ 㸐䗎כיע㎫ 4.24 סֹמյ訪ؿ٭غ㺤ꪛס갾מ偆ַ匛⚂䑑ך阛砯 ג True/False כ㺤ꪛ⯼ס True/False ס畤阛砯׆כך and 匛⚂כ ⻎坎ם⭦槏㵅槁׆כֿך׀ױն 85 -
8 6 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 1: init_idx
= X[:, 0] > np.min(X[:, 0])-1 2: split_idx_left = X[:, 0] < 5. 3: split_idx_right = X[:, 0] >= 5. 4: 5: split_right_1 = X[:, 0] < 5.2 6: split_right_2 = X[:, 0] >= 5.2 7: 8: idx = split_idx_right[:10]*split_right_1[:10] 9: X_10 = X[:10, :] 10: print(idx) 11: print(X_10[idx, :]) 12: print("=====") 13: print(X_10) 86 -
8 7 痥 4 畍 㹋鄲 4.2 寸㹀加ך㛇劤Ⳣ椚 ㎫ 4.25:
韢槏悍砯Α 阛砯גؕ٤ظشؠت X y מ氠ַ׆כךյ㎫ 4.25 סֹמ訪 ؿ٭غⷃ⛺ך X ⺅⮂׆כֿ⺪茣ךնױגյؕ٤ظشؠتⷃ⛺ך node_value ס⡑⺅⮂׆כך◙廠鉿ֹ׆כֿך׀ױն 87 -
8 8 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 4.3 ،ٝ؟ٝـٕ㷕统
4-3 硼ךעؓ٤ئ٤هٜ㳔肪ס㵅鍮מחַי牞霼鉿ַױն㕈勓溷מע 4-2 硼ך㵅鍮ג尴㴻勎ؠٚتⵊյאסؕ٤تذ٤ت鏿俙✑䧯⫙⯈ 氠׆כך㵅鍮ֿ⺪茣םג״յ4-2 硼סٜؓإٛثّס㵅鍮מ嬟כ嬟 鼛溷槏闋עַכ䘼ַױն♧┫յ4-3-1 硼ךעف؟٤ء (Bagging)յ 4-3-2 硼ךעٚ٤رّنؚٝتع (Random Forest)յ4-3-3 硼ךעⴢꏕه٭ تطؔ٤ء (Gradient Boosting) מחַיאב⺅䪒ַױն 4.3.1 غؘؚٝ 4-3-1 硼ךעف؟٤ء (Bagging) ס㵅鍮מחַי⺅䪒ַױն♧┫յ 4-2-1 硼ך⺅䪒זג iris dataset X כ y כי牞霼鉿ַױն ױյف؟٤ء脝ֻ갾מ䖩锡מםסֿյ ր沌םئ٤وٜמ㕈טַ י栃皑ם㳔肪㊭圸碎ցכַֹ׆כךն׆סֹם免ע㕈勓溷מעئ ٤وٜסؕ٤ظشؠتאב✑䧯ף虘ַכ脝ֻכ虘ַךն 1: np.random.seed(10) 2: 3: sample_idx = np.random.choice(np.arange(0, 100, 1), size=int(100*0.7)) 4: unique_idx, idx_counts = np.unique(sample_idx, return_counts=True) 5: print(sample_idx) 6: print("=====") 7: print(unique_idx) 8: print(idx_counts) 88 -
8 9 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 ㎫ 4.26:
ؕ٤ظشؠتס✑䧯 (ꓨ鏿ֵ) ㎫ 4.26 ע 100 ⠥סئ٤وٜ־ 70 ⠥ꓨ鏿ֵךئ٤وٛ٤ء갾ס ؕ٤ظشؠت气䧯כ脝ֻיַגדׄגכ䘼ַױնnp.random.seed(10) ע闋鞃סꌬ⻉┪յ篙卸ס⫙槁䙎䬎⟛ג״מ▸俙ס㎷㴻鉿ַױגն 1: sample_idx = np.random.choice(np.arange(0, X.shape[0], 1), \ 2: size=int(X.shape[0]*0.1)) 3: 4: print(sample_idx) 5: print("=====") 6: print(X[sample_idx, :]) 89 -
9 0 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 ㎫ 4.27:
iris ־סئ٤وٛ٤ء ㎫ 4.26 ע iris ־ 10 סئ٤وٜꓨ鏿ֵךئ٤وٛ٤ءג篙卸ך նؕ٤ظشؠت 88 ֿꓨ鏿יַסךյ[5.6 3. 4.1 1.3] ֿ◝㎇ئ٤وٛ ٤ءיַ׆כֿ牞霼ך׀ױն 尴㴻勎ס㕈勓⭦槏ֿؠٚتⵊיַףف؟٤ءעⷃמئ٤وٜ־ٚ ٤رّמ䬂⮂յ鏿俙㎇㵅鉿鉿ַյ4-2-6 ס♣銨⡑ס阛砯כ⻎锡꽝ך㝂 俙尴 (⮔꿔) ױגע䌐㐬 (㎇䊟) 阛砯յ◙廠⡑阛砯ױն 4.3.2 ٓٝتيؿٖؓأز 4-3-2 硼ךעٚ٤رّنؚٝتع (Random Forest) ס㵅鍮מחַי⺅ 䪒ַױն㕈勓溷מע 4-3-1 硼סف؟٤ءכׯ⻎ךֿյ杅䖇ꓪסٚ٤ رّךסꈷ䫘⻉י鉿ֹכַֹ掾ֿ沌םױն 90 -
9 1 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 1: np.random.seed(0)
2: np.random.choice(np.arange(0, X.shape[1], 1), \ 3: size=int(X.shape[1]*0.7), replace=False) ㎫ 4.28: 杅䖇ꓪסꈷ䫘 (ꓨ鏿ם) 杅䖇ꓪסꈷ䫘מֵגזיעꓨ鏿םַֹמꈷצגַסךյ㎫ 4.28 ךע np.random.choice ס؛وب٘٤ס replace False כױגն ױגյ㳔肪ס갾מꈷ䫘ג杅䖇ꓪע◙廠ס갾מ氠ַ䖩锡ֵֿסךյ㳔 肪㊭ס㜟俙כי媘יֽׂ䖩锡ֵֿױն 1: select_features = [] 2: 3: np.random.seed(0) 4: select_features.append(np.random.choice(np.arange(0, \ 5: X.shape[1], 1), size=int(X.shape[1]*0.7), replace=False)) 6: select_features.append(np.random.choice(np.arange(0, \ 7: X.shape[1], 1), size=int(X.shape[1]*0.7), replace=False)) 8: select_features.append(np.random.choice(np.arange(0, \ 9: X.shape[1], 1), size=int(X.shape[1]*0.7), replace=False)) 10: 11: print(select_features) 91 -
9 2 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 ㎫ 4.29:
杅䖇ꓪס阾ꜗכ◙廠免ס攍 㳔肪免מ❈氠ג杅䖇ꓪמחַיעյ㎫ 4.29 סֹמ׆כך◙廠免 מ攍ֿ⺪茣מםױնאבס㳔肪㊭ס◙廠⡑⺅䕑ג䔿ס⭦槏ע ⮔꿔ךֵ㎇䊟ךֵف؟٤ءכ⻎םסךאל곓ׂ脝ֻםׂי虘ַ ־כ䘼ַױն 4.3.3 ⺟ꂁـ٦أذ؍ؚٝ 4-3-3 鞃ךעⴢꏕه٭تطؔ٤ء (Gradient Boosting) ס㵅鍮מחַי⺅ 䪒ַױն⮔꿔מחַיע⡑סⰺ䓜ימꫀי鏿겧מםַסךյ ㎇䊟סײמחַי⺅䪒ַױն 92 -
9 3 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 ㎫ 4.30:
ⴢꏕه٭تطؔ٤ءס㵅鍮 ⴢꏕه٭تطؔ٤ءמחַיעؠٚتכיס㵅鍮אסױױ牞霼 偙ֿ־ַכ䘼גסךյ㎫ 4.30 ךע⯁㛿ס Jupyter نٜؒؕ ס⭦槏ׇ箩♃׆כמױגն㕈勓溷מע scikit-learn סؕ٤ذ٭ نؘ٭تמ⛣יַסךյscikit-learn כ⻎坎םْؕ٭ةךׇ牞霼ַגד ׄגכ䘼ַױն ױյclf ע尴㴻勎ס㵅鍮ؠٚتⵊגס־ؕ٤تذ٤ت✑䧯յ אה氠ַי㳔肪鉿םזיַ׆כֿ־־כ䘼ַױն׆סֹם 㵅鍮עؓ٤ئ٤هٜ㳔肪מַֽיעאל㜟זג需ךעםַכ䘼ַױ ֿյف؟٤ءכסꇙַע┫־ 3 鉿ךն㳔肪מ氠ַג X 氠ַי y_pred 阛砯յ׆ y_־䑛׀յ⫙䍲 y_מחַי㳔肪鉿ֹכַֹ嵣מם ױն ׆מյאױךמ㳔肪ג㳔肪㊭ס篙卸מ㕈טַי♧䔿ס㳔肪㊭ס 㳔肪鉿ֹه٭تطؔ٤ءס⭦槏㵅槁׆כֿך׀ױն׆׆ך㵅鍮מ 93 -
9 4 痥 4 畍 㹋鄲 4.3 ،ٝ؟ٝـٕ㷕统 ֵגזיյy_ y
ס deepcopy כםׄף⡑ֿ⻎免מ紬겏׆כך 劳ױׂםַ䮕ⳛ獏׆כמ峜䟨ךնױגյy_pred y_־䑛ׂ갾 מـؕق٭قْٚ٭ذס α ־ׄ׆כֵסךյقْٚ٭ذ⭦槏ֿ⺪ 茣םֹמ㵅鍮鉿ַױגն 94