ʻAʻole, ʻoiaʻiʻo, ʻaʻole wau koʻikoʻi. Pono e loaʻa ka palena i ka hiki ke maʻalahi i ke kumuhana. Akā no nā pae mua, hoʻomaopopo i nā manaʻo kumu a "komo" wikiwiki i ke kumuhana, hiki ke ʻae ʻia. E kūkākūkā mākou pehea e inoa pono ai i kēia mea (nā koho: "Machine learning for dummies", "Data analysis from diapers", "Algorithms for the little poder") ma ka hopena.
I ka lae. Ua kākau i kekahi mau polokalamu noi ma MS Excel no ka nānā ʻana a me ka hōʻike ʻike ʻana i nā kaʻina hana i hana ʻia ma nā ʻano aʻo mīkini like ʻole i ka wā e nānā ana i ka ʻikepili. ʻO kaʻikeʻana i ka manaʻoʻiʻo, ma hope o nā mea a pau, e like me ka'ōlelo a ka poʻe lawe i ka moʻomeheu, ka mea i hoʻomohala i ka hapa nui o kēia mauʻano (ma ke ala,ʻaʻole lākou a pau. ʻO ka "mīkini vector kākoʻo" ikaika loa, a iʻole SVM, kākoʻo vector mīkini ka mea i hanaʻia o ʻO ko mākou hoa pili ʻo Vladimir Vapnik, Moscow Institute of Management. 1963, ma ke ala! I kēia manawa, akā naʻe, aʻo ʻo ia a hana ma USA).
ʻEkolu faila no ka loiloi
1. K-ʻo ia hoʻi ka hui ʻana
ʻO nā pilikia o kēia ʻano e pili ana i ka "aʻo ʻole ʻia," inā pono mākou e puʻunaue i ka ʻikepili mua i kekahi helu o nā ʻano i ʻike mua ʻia, akā ʻaʻohe helu o nā "pane pololei"; pono mākou e unuhi iā lākou mai ka ʻikepili ponoʻī. . ʻO ka pilikia maʻamau o ka ʻimi ʻana i nā ʻano ʻano o nā pua iris (Ronald Fisher, 1936!), i manaʻo ʻia ʻo ia ka hōʻailona mua o kēia kahua ʻike, ʻo ia wale nō ke ʻano.
He maʻalahi ke ʻano. Loaʻa iā mākou kahi hoʻonohonoho o nā mea i hōʻike ʻia ma ke ʻano he vectors (sets o nā helu N). I nā irises, he mau helu kēia o nā helu 4 e hōʻike ana i ka pua: ʻo ka lōʻihi a me ka laulā o nā lobes waho a me loko o ka perianth, kēlā me kēia (). Koho ʻia ka metric Cartesian maʻamau e like me ka mamao, a i ʻole ke ana o ka pili ma waena o nā mea.
Ma hope aʻe, koho ʻia nā kikowaena puʻupuʻu (a ʻaʻole ʻole, e ʻike i lalo), a ua helu ʻia nā mamao mai kēlā me kēia mea i nā kikowaena cluster. Hōʻailona ʻia kēlā me kēia mea ma kahi ʻanuʻu hoʻololi i hāʻawi ʻia ma ke kikowaena kokoke loa. A laila e hoʻoneʻe ʻia ke kikowaena o kēlā me kēia puʻupuʻu i ka helu helu o nā koina o kona mau lālā (ma ka hoʻohālikelike me ka physics, kapa ʻia ʻo ia ʻo "center of mass"), a ua hana hou ʻia ke kaʻina hana.
Hoʻohui koke ke kaʻina hana. Ma nā kiʻi ma nā ʻāpana ʻelua e like me kēia:
1. Ka puunaue maalea mua o na kiko ma ka mokulele a me ka heluna o na puu
2. E wehewehe ana i nā kikowaena pūʻulu a me ka hāʻawi ʻana i nā kiko i kā lākou mau pūʻulu
3. Ka hoʻoili ʻana i nā hoʻonohonoho o nā kikowaena puʻupuʻu, e helu hou ana i ka pili ʻana o nā helu a hiki i ka paʻa ʻana o nā kikowaena. ʻIke ʻia ke alahele o ke kikowaena puʻupuʻu e neʻe ana i kona kūlana hope.
I kēlā me kēia manawa, hiki iā ʻoe ke hoʻonohonoho i nā kikowaena puʻupuʻu hou (me ka ʻole o ka hoʻopuka ʻana i ka mahele hou o nā helu!) ʻO ka makemakika, ʻo ia hoʻi, no ka hoʻonui ʻia ʻana o ka hana (ka huina o nā huina huinahā mai nā kiko a i nā kikowaena o kā lākou mau puʻupuʻu), ʻaʻole mākou e ʻike i kahi honua, akā he palena liʻiliʻi. Hiki ke hoʻopau ʻia kēia pilikia ma o kahi koho koho ʻole o nā kikowaena cluster mua, a i ʻole ma ka helu ʻana i nā kikowaena hiki (i kekahi manawa ʻoi aku ka maikaʻi o ke kau pololei ʻana iā lākou ma kekahi o nā kiko, a laila aia ka hōʻoia ʻaʻole mākou e nele. pūʻulu). I kēlā me kēia hihia, he infimum mau ka set finite.
(mai poina e hiki ke kākoʻo macro. Ua nānā ʻia nā faila no nā maʻi virus)
ʻO ka wehewehe ʻana i ke ʻano ma Wikipedia -
2. Hoʻopili ʻia e nā polynomial a me ka hoʻokaʻawale ʻikepili. Hoʻomaʻamaʻa hou
ʻO ka mea ʻepekema maikaʻi a kaulana hoʻi o ka ʻepekema data K.V. Ua wehewehe pōkole ʻo Vorontsov i nā ʻano aʻo mīkini e like me "ka ʻepekema o ke kaha kiʻi ʻana ma nā kiko." Ma kēia hiʻohiʻona, e ʻike mākou i kahi maʻamau i ka ʻikepili me ka hoʻohana ʻana i ke ʻano ʻāpana liʻiliʻi loa.
Hōʻike ʻia ke ʻano o ka hoʻokaʻawale ʻana i ka ʻikepili kumu i ka "hoʻomaʻamaʻa" a me ka "mana", a me ke ʻano e like me ke aʻo hou ʻana, a i ʻole "hoʻoponopono hou" i ka ʻikepili. Me ka hoʻopili pololei ʻana, e loaʻa iā mākou kekahi hewa ma ka ʻikepili hoʻomaʻamaʻa a me kahi hewa nui aʻe ma ka ʻikepili mana. Inā ʻaʻole pololei, hopena ia i ka hoʻoponopono pololei ʻana i ka ʻikepili hoʻomaʻamaʻa a me kahi hewa nui ma ka ʻikepili hoʻāʻo.
(He mea ʻike maopopo ʻia ma o nā helu N hiki i kekahi ke kaha kiʻi i hoʻokahi pihi o ka degere N-1, a ʻo kēia ʻano hana ma ka hihia maʻamau ʻaʻole e hāʻawi i ka hopena i makemake ʻia. )
1. E hoonoho i ka mahele mua
2. Māhele mākou i nā helu i ka "hoʻomaʻamaʻa" a me ka "mana" i ka ratio o 70 a 30.
3. Kahakiʻi mākou i ka pihi pili ma nā wahi aʻo, ʻike mākou i ka hewa i hāʻawi ʻia ma ka ʻikepili mana
4. Kahakiʻi mākou i kahi pihi pololei ma o nā wahi hoʻomaʻamaʻa, a ʻike mākou i kahi hewa nui ma ka ʻikepili hoʻomalu (a ʻaʻohe ma ka ʻikepili aʻo, akā he aha ke kumu?).
Hōʻike ʻia, ʻoiaʻiʻo, ʻo ia ka koho maʻalahi loa me ka mahele hoʻokahi i ka "hoʻomaʻamaʻa" a me ka "mana" subsets; ma ka hihia maʻamau, hana ʻia kēia i nā manawa he nui no ka hoʻoponopono maikaʻi loa o nā coefficients.
E ho'ā i nā macros no ka hana pono
3. Hoʻololi ka gradient iho a me ka dynamics o ka hewa
E loaʻa kahi hihia 4-dimensional a me ka regression linear. E hoʻoholo ʻia nā coefficient regression linear i kēlā me kēia ʻanuʻu me ka hoʻohana ʻana i ke ʻano iho gradient, ʻo ka mua ʻaʻole nā coefficient a pau. Hōʻike ka pakuhi ʻokoʻa i ka dynamics o ka hōʻemi hewa ʻana e like me ka hoʻoponopono ʻana o nā coefficients me ka pololei. Hiki ke nānā i nā kuhi 2-dimensional ʻehā.
Inā ʻoe e hoʻonoho i ka ʻanuʻu gradient he nui loa, hiki iā ʻoe ke ʻike i kēlā me kēia manawa e lele mākou i ka liʻiliʻi loa a hiki i ka hopena i kahi helu ʻoi aku ka nui o nā ʻanuʻu, ʻoiai ma ka hopena e hōʻea mau mākou (koe ke hoʻopaneʻe mākou i ka pae iho. nui - a laila e hele ka algorithm " i nā spades"). A ʻaʻole maʻalahi ka pakuhi o ka hewa e pili ana i ka ʻanuʻu iteration, akā "jerky".
1. E hana i ka ʻikepili, e hoʻonoho i ka ʻanuʻu gradient iho
2. Me ke koho pololei ʻana i ka ʻanuʻu gradient iho, hiki mākou i ka liʻiliʻi loa
3. Inā koho hewa ʻia ka ʻanuʻu gradient iho, ʻoi aku ka nui o ke kiʻekiʻe, "jerky" ka pakuhi hewa, ʻoi aku ka nui o nā ʻanuʻu.
и
4. Inā koho hewa mākou i ka ʻanuʻu iho ʻana, neʻe mākou mai ka liʻiliʻi loa
(No ka hana hou i ke kaʻina hana me ka gradient descent step values i hōʻike ʻia ma nā kiʻi, e nānā i ka pahu "reference data").
Wahi a ke kaiāulu i mahalo ʻia, ʻae ʻia ke ʻano maʻalahi a me ke ʻano o ka hōʻike ʻana i ka mea? He mea pono ke unuhi i ka ʻatikala ma ka ʻōlelo Pelekania?
Source: www.habr.com