- Ukuqhekeka kwe-QR kuthuthukisa ukuzinza kwezinombolo ku-OLS ngokugwema ukugoba kwenombolo yesimo se-matrix etholakala kuma-equation ajwayelekile.
- Ukuzindla kwendlu kunikeza indlela eqinile yokuguqula i-matrix yomklamo ibe yisimo esingunxantathu esiphezulu, okwenza kube lula ukufaka esikhundleni esiphumelelayo.
- Ukuphatha izinhlelo ezingenazo izikhundla kudinga amasu athuthukile njengokushintshashintsha kwamakholomu noma i-SVD ukuqinisekisa ikhambi elinembile lama-least-squares.
Wake wazibuza ukuthi kungani isofthiwe yezibalo yobungcweti ingafaki izinombolo kufomula ye-OLS yencwadi? Nakuba indlela yakudala ibukeka inhle ephepheni, empeleni iyiphupho elibi lokunemba kwekhompyutha . Uma ubhekene nedatha yangempela, ikakhulukazi ekubuyiseleni kwe-quadratic lapho izibikezeli zingahlobana khona kakhulu, indlela ejwayelekile yokuguqula ama-matrices ivame ukuholela emaphutheni asendle angenza imiphumela yakho ibe yize ngokuphelele.
Yilapho-ke ukuhlukaniswa kwe-QR kuqala khona ukusebenza, kusebenza njengenjini eyimfihlo ngaphansi kwe-hood yamathuluzi afana nomsebenzi we-R we-lm(). Esikhundleni sokwandisa ubunzima benkinga, le ndlela ihlukanisa i-matrix yokuklama ibe yizicucu ezimbili ezilula, okwenza indlela eya kuma-coefficients ibe lula kakhulu futhi izinzile kakhulu . Ake singene shí endleleni yokusebenza kokuthi lokhu kusebenza kanjani nokuthi kungani kuyindinganiso yegolide yokuxazulula izinkinga ze-least-squares.
Inkinga Ngezilinganiso Ezijwayelekile
Kusethaphu ejwayelekile ye-Ordinary Least Squares (OLS), sihlose ukunciphisa isamba sezinsalela eziyisikwele. Isixazululo sencwadi yi-equation evamile: β = (XᵀX)⁻¹Xᵀy . Inkinga enkulu lapha yigama elithi XᵀX. Uma uphindaphinda i-matrix nge-transpose yayo, empeleni ubeka isikwele inombolo yesimo . Inombolo yesimo ngokuyisisekelo iyisilinganiso sokuthi umsindo omncane kangakanani kudatha yakho okhuliswa ube yiphutha elikhulu ekuphumeni kwakho.
Uma i-matrix yakho yomklamo isivele "ithinta kancane" (okusho ukuthi inamakholomu acishe abe yi-collinear), ukwenza isikwele kulokho kungazinzile kungaholela emaphutheni aphansi adla ukunemba kwakho. Ungase ulahlekelwe izinombolo eziningana ezibalulekile, yingakho indlela ye-QR ikhethwa kakhulu—isebenza ngqo ne-X futhi igwema lokho kuphambuka okuyingozi phakathi kwe-matrix yomkhiqizo oxubile.
Ukuhlahlela Ukuhlukaniswa kwe-QR
Umqondo oyinhloko ukuhlanganisa i-matrix yomklamo u-X ibe yi-matrices ezimbili ezithile: u-X = QR. Lapha, u-Q uyi-matrix e-orthogonal , okusho ukuthi amakholomu ayo aqondile futhi anobude beyunithi eyodwa. I-matrix u-R ingunxantathu ophezulu , okusho ukuthi konke okungaphansi kwe-diagonal eyinhloko kungu-zero. Lokhu kuguqulwa kuyashintsha umdlalo ngoba i-matrix e-orthogonal kulula kakhulu ukuyiguqula—i-inverse yayo imane nje iyi-transpose yayo (Q⁻¹ = Qᵀ).
Uma sifaka lokhu enkingeni ye-least-squares, yonke into iwa ibe uhlelo olulula kakhulu: Rβ = Qᵀy . Njengoba i-R ingunxantathu ophezulu, akudingeki senze ukuguqulwa kwe-matrix okubizayo nokungazinzile. Esikhundleni salokho, sisebenzisa inqubo ebizwa ngokuthi i-back-substitution , sixazulula i-coefficient yokugcina kuqala bese siyixhuma kabusha ukuze sithole okusele. Iyashesha, ithambile, futhi iqinile ngokwezibalo.
Ukuzindla Kwabendlu vs. Gram-Schmidt
Kunezindlela ezimbalwa zokubala ngempela u-Q no-R. Kungenzeka ukuthi uke wabona inqubo ye-Gram-Schmidt ekilasini le-algebra eliqondile; kuyindlela enembile yokususa ama-projection ukuze wenze ama-vector abe yi-orthogonal. Kodwa-ke, inguqulo yakudala ingaba namaphutha okukhansela lapho ama-vector ecishe afane. Ngisho ne-Modified Gram-Schmidt (MGS) ingcono, kodwa ayikabi yiyo ehamba phambili.
Amalabhulali angempela afana ne-LAPACK asebenzisa i-Householder reflections . Cabanga uthatha i-vector bese uyibonisa ngaphesheya kwendiza ukuze uyivumelanise ne-axis ethile. Ngokusebenzisa uchungechunge lwalezi zibonakaliso, singakhipha ngokuhlelekile izakhi ezingaphansi kwe-diagonal ka-X. Le ndlela izinzile kakhulu ku-finite-precision arithmetic, okuqinisekisa ukuthi izakhiwo ze-orthogonal ze-Q zigcinwa ngisho nangedatha engcolile.
Ukubhekana Nokuntuleka Kwesikhundla kanye Nokungalingani Kwezikhundla
Kwenzekani uma idatha yakho "iphukile"? Kwezinye izimo, enye yamakholomu akho ingaba inhlanganisela ephelele yomugqa wezinye—lokhu kushoda kwezinga . Esixazululweni se-OLS esijwayelekile, lokhu kungenza i-matrix eyodwa kanye ne-inversion kungenzeki. Indlela ye-QR iphatha lokhu kahle, ikakhulukazi uma ihambisana ne -column pivoting.
Ukujika kuhilela ukushintshanisa amakholomu ngokusekelwe ekubalulekeni kwawo (isilinganiso sawo) ukuze kusunduzwe izingxenye "ezinolwazi" kakhulu ze-matrix ziye phezulu kwesobunxele. Uma i-matrix ingenaso isitifiketi ngempela, inqubo ye-QR izoyiveza ngokushiya o-zero ku-diagonal ka-R . Ezimweni ezinzima kakhulu, i-Singular Value Decomposition (SVD) iyisikhali esiphezulu, njengoba inganikeza ikhambi lesilinganiso esincane ngisho nalapho uhlelo lunganqunywanga ngokuphelele.
I-Logic Yokusebenzisa Okusebenzayo
Ukuze usebenzise lokhu, uqala ngokwakha i-matrix yakho yokuklama, kufaka phakathi ikholomu yalezo ze-intercept. Ube ususebenzisa ukuguqulwa kwe-Householder ukuthola u-R kanye nokubonakaliswa okuqongelelwe ukwakha u-Q. Uma usunalokhu, uphindaphinda u-Qᵀ ngevektha yakho yokuphendula u-y. Isinyathelo sokugcina umsebenzi we-backsolve , ophindaphinda kusukela emgqeni ongezansi ka-R kuya phezulu ukuze kuhlukaniswe i-coefficient ngayinye u-β.
Lonke leli phayiphi liqinisekisa ukuthi asikhulisi umsindo. Kungakhathaliseki ukuthi usebenzisa lokhu ku-Python usebenzisa i-NumPy noma usebenzisa ama-primitives angaphakathi ka-R, i-logic ihlala ifana: gwema i-XᵀX ngazo zonke izindleko futhi uvumele i-orthogonality ye-Q yenze umsebenzi osindayo. Yingakho amamodeli akho okubuyela emuva ehlala enembile ngisho noma izibikezeli zinobuqili futhi zihlobene kakhulu.
Ngokushintsha kusuka kuzibalo ezivamile ezingenalwazi kuya endleleni ye-QR-Householder, sishintshana ngobulula bomqondo ukuze sithole izinzuzo ezinkulu ekuthembekeni . Ukusetshenziswa kokuguqulwa kwe-orthogonal kuvimbela ukukhuliswa kwamaphutha ezinombolo, kuyilapho ukubuyiselwa emuva kunikeza indlela ephumelelayo eya ezilinganisweni zokugcina. Ukuqonda lokhu kuguquka kusuka kumbono kuya ekusetshenzisweni kwezinombolo yikho okuhlukanisa imodeli eyisisekelo kuleyo eqinile ngokwezibalo futhi elungele ukuhlaziywa kwedatha yebanga lokukhiqiza.
