Kungani i-MAP ne-MRR Yehluleka Ukuthola Izinga Losesho (nokuthi Yini Okufanele Uyisebenzise Kunalokho)

nimda December 25, 2025

0 3 6 minutes read

Kungani i-MAP ne-MRR Yehluleka Ukuthola Izinga Losesho (nokuthi Yini Okufanele Uyisebenzise Kunalokho)

ngokuvamile sebenzisa I-Mean Reciprocal Rank (MRR) futhi I-Mean Average Precision (MAP) ukuhlola izinga lamazinga abo. Kulokhu okuthunyelwe, sizoxoxa ngokuthi kungani (MAP) kanye (MRR) engahambisani kahle nokuziphatha kwabasebenzisi besimanje ekusezingeni losesho. Bese sibheka amamethrikhi amabili asebenza njengezinye izindlela ezingcono kokuthi (MRR) kanye (MAP).

Yini i-MRR ne-MAP?

I-Mean Reciprocal Rank (MRR)

IMean Reciprocal Rank ((MRR)) yizinga elimaphakathi lapho kuvela khona into yokuqala efanele.

$$mathrm{RR} = frac{1}{text{izinga lento yokuqala efanele}}$$

Ku-e-commerce, izinga lokuqala elifanele lingaba izinga lento yokuqala echofozwe ukuphendula umbuzo.

Usesho lwe-Amazon lwe-'burr coffee grinder'. Lapha, sithatha ukuthi into yesibili ingumphumela ofanele.

Kulesi sibonelo esingenhla, cabanga ukuthi into efanelekile yinto yesibili. Lokhu kusho:
$$mathrm{Reciprocal Rank} = frac{1}{2}$$
Izinga elilinganayo libalwa kuyo yonke imibuzo esesethi yokuhlola. Ukuze uthole imethrikhi eyodwa yayo yonke imibuzo, sithatha isilinganiso samazinga alinganayo ukuze sithole Kusho Isikhundla Esivumelanayo

$$mathrm{Mean Reciprocal Rank} = frac{1}{N}sum_{i=1}^N {frac{1}{text{Rank of First Relevant Item}}}$$

lapho (N) inombolo yemibuzo. Kule ncazelo, singabona ukuthi (MRR) igxile ekutholeni umphumela owodwa obalulekile kusenesikhathi. Ayilinganisi ukuthi kwenzekani ngemva komphumela wokuqala ohlobene.

I-Mean Average Precision (MAP)

I-Mean Average Precision ((MAP) ikala ukuthi isistimu izibuyisa kanjani izinto ezifanele nokuthi ziboniswa kusenesikhathi kangakanani. Siqala ngokubala Ukunemba Okumaphakathi (AP) kumbuzo ngamunye. Sichaza i-AP njenge
$$mathrm{AP} = frac{1}{|R|}sum_{k=1}^{K}mathrm{Precision@}k cdot mathbf{1}[text{item at } k text{ is relevant}]$$
lapho (|R|) kuyinombolo yezinto ezifanele zombuzo
(mathrm{MAP}) iyona isilinganiso se (mathrm{AP}) kuyo yonke imibuzo

I-equation engenhla ibukeka kakhulu, kodwa empeleni ilula. Ake sisebenzise isibonelo ukuyihlehlisa. Cabanga ukuthi umbuzo unezinto ezi-3 ezifanele, futhi imodeli yethu ibikezela ukuhleleka okulandelayo:

Rank: 1 2 3 4 5 
Item: R N R N R

(R = efanele, N = ayibalulekile)
Ukubala i- (MAP), sibala i-AP endaweni ngayinye efanele:

@1: Ukunemba = 1/1 = 1.0
@3: Ukunemba = 2/3 ≈ 0.667
@5: Ukunemba = 3/5 = 0.6

$$mathrm{AP} = frac{1}{3}(1.0 + 0.667 + 0.6) = 0.756$$
Sibala okungenhla kuyo yonke imibuzo futhi siyilinganisele ukuze sithole (MAP). Ifomula ye-AP inezingxenye ezimbili ezibalulekile:

Precision@k: Njengoba sisebenzisa Precision, ukubuyisa izinto ezifanele kusenesikhathi kuveza amanani aphezulu anemba. Uma imodeli ilinganisa izinto ezifanele kamuva, Precision@k iyancipha ngenxa ye-k enkulu
Isilinganiso Sokunemba: Silinganisa ukunemba phezu kwesamba senani lezinto ezifanele. Uma isistimu ingalokothi ibuyise into noma iyibuyisele ngale kokunqamuka, into ayinikezi lutho enambeni ngenkathi ibala ngedenominator, enciphisa (AP) kanye (MAP).

Kungani i-MAP ne-MRR Engalungile Ngezinga Losesho

Manje njengoba sesihlanganise izincazelo, masiqonde ukuthi kungani (MAP) kanye (MRR) zingasetshenziselwa ukukala imiphumela yosesho.

Ukuhlobana kugrediwe, hhayi kanambambili

Uma sibala (MRR), sithatha izinga lento yokuqala efanele. Ku-(MRR), siphatha zonke izinto ezifanele ngokufanayo. Awenzi mehluko uma into ehlukile ehambisanayo ivela kuqala. Eqinisweni, izinto ezahlukene zivame ukuba nokuhambisana okuhlukile.

Ngokufanayo, kokuthi (MAP), sisebenzisa ukuhambisana kanambambili- sivele sibheke into elandelayo efanele. Futhi, (MAP) awenzi mehluko kumphumela wokuhlobana wezinto. Ezimweni zangempela, ukuhlobana kubekwe ngokwezinga, hhayi kanambambili.

Item     : 1 2 3 
Relevance: 3 1 0

(MAP) kanye (MRR) zombili azinaki ukuthi into efanele iyinhle kangakanani. Behluleka ukulinganisa ukuhambisana.

Abasebenzisi Skena Imiphumela Eminingi

Lokhu kucace kakhulu kokuthi (MRR). Ekubaleni (MRR), siqopha izinga lento yokuqala efanele. Bese uziba konke ngemuva. Kungaba kuhle ukubheka, i-QA, njll. Kodwa lokhu kubi ngezincomo, ukusesha umkhiqizo, njll.

Ngesikhathi sokusesha, abasebenzisi abami kumphumela wokuqala ohlobene (ngaphandle kwezimo lapho kunempendulo eyodwa kuphela efanele). Abasebenzisi baskena imiphumela eminingi enikela kukho konke ukuhambisana kosesho.

I-MAP igcizelela kakhulu ukukhumbula

(MAP) iyahlanganisa
$$mathrm{AP} = frac{1}{|R|}sum_{k=1}^{K}mathrm{Precision@}k cdot mathbf{1}[text{item at } k text{ is relevant}]$$
Ngenxa yalokho, yonke into efanele ifaka isandla ekutholeni amaphuzu. Ukushoda kwanoma iyiphi into ehlobene kulimaza amaphuzu. Uma abasebenzisi benza usesho, abanantshisekelo yokuthola zonke izinto ezifanele. Banentshisekelo yokuthola izinketho ezimbalwa ezinhle kakhulu. (MAP) ukulungiselelwa kuphusha imodeli ukuze ifunde umsila omude wezinto ezifanele, noma ngabe umnikelo wokuhlobana uphansi, futhi abasebenzisi abalokothi baskrole kude kangako. Ngakho-ke, (MAP) igcizelela kakhulu ukukhumbula.

I-MAP Ibola Ngokuqondile

Cabangela isibonelo esingezansi. Sibeka into efanele ezindaweni ezintathu ezahlukene bese sibala i-AP

Izinga	Ukunemba@k	AP
1	1/1 = 1.0	1.0
3	1/3 ≈ 0.33	0.33
30	1/30 ≈ 0.033	0.033

Ukunemba Okumaphakathi kuwo wonke Amazinga ahlukene

I-AP ku-Rank 30 ibukeka kabi kune-Rank 3, ebukeka kabi kune-Rank 1. Isikolo se-AP siwohloka ngokuhambisana nezinga. Eqinisweni, i-Rank 3 vs Rank 30 ingaphezu komehluko we-10x. Kufana nokubonwa uma kuqhathaniswa nokungabonwa.

(MAP) izwela indawo kodwa ibuthakathaka kuphela. Akubonisi ukuthi ukuziphatha komsebenzisi kushintsha kanjani ngesimo. (MAP) izwela indawo nge-Precision@k, lapho ukubola okunezinga kuwumugqa. Lokhu akubonisi ukuthi ukunaka komsebenzisi kwehla kanjani emiphumeleni yosesho.

I-NDCG ne-ERR yiZinqumo Ezingcono

Ukuze uthole izinga lemiphumela yosesho, (NDCG) kanye (ERR) ukukhetha okungcono. Balungisa izinkinga (MRR) kanye (MAP) ehlushwa kuzo.

I-Reciprocal Rank elindelwe (ERR)

I-Reciprocal Rank elindelwe ((ERR)) ithatha imodeli yomsebenzisi we-cascade lapho umsebenzisi enza okulandelayo

Iskena uhlu ukusuka phezulu kuye phansi
Ezingeni ngalinye (i),
- Ngokunokwenzeka (R_i), umsebenzisi wanelisekile futhi iyama
- Ngamathuba (1- R_i), umsebenzisi uyaqhubeka ebheka phambili

(ERR) ibala i izinga elilinganayo elilindelekile lalesi sikhundla sokuma lapho umsebenzisi anelisekile khona:
$$mathrm{ERR} = sum_{r=1}^n frac{1}{r} cdot {R}_r cdot prod_{i=1}^{r-1}{1-{R}_i}$$
lapho (R_i) ethi (R_i = frac{2^{l_i}-1}{2^{l_m}}) futhi (l_m) iyinani eliphezulu lelebula elingenzeka.

Ake siqonde ukuthi (ERR) ihluke kanjani kokuthi (MRR).

(ERR) isebenzisa (R_i = frac{2^{l_i}-1}{2^{l_m}}), okuhambisana nesigaba, ngakho umphumela unganelisa kancane umsebenzisi
(ERR) ivumela izinto eziningi ezifanele ukuthi zinikele. Izinto zakuqala zekhwalithi ephezulu zinciphisa umnikelo wezinto zakamuva

Isibonelo 1

Sizothatha isibonelo sethoyizi ukuze siqonde ukuthi (ERR) kanye (MRR) zihluka kanjani

Rank     : 1 2 3 
Relevance: 2 3 0

(MRR) = 1 (into efanele isendaweni yokuqala)
(ERR) =
- (R_1 = {(2^2 – 1)}/{2^3} = {3}/{8})
- (R_2 ={(2^3 – 1)}/{2^3} = {7}/{8})
- (R_3 ={(2^0 – 1)}/{2^3} = 0)
- (ERR = (1/1) cdot R_1 + (1/2) cdot R_2 + (1/3) cdot R_3 = 0.648)
(MRR) uthi izinga eliphelele. (ERR) uthi akuphelele ngoba into ephakeme ehlobene ivela kamuva

Isibonelo sesi-2

Ake sithathe esinye isibonelo ukuze sibone ukuthi uguquko lwezinga luthinta kanjani (ERR) umnikelo wento. Sizobeka into ebaluleke kakhulu ezindaweni ezihlukene ohlwini bese sibala (ERR) umnikelo waleyo nto. Cabangela izimo ezingezansi

Izinga 1: ([8, 4, 4, 4, 4])
Izinga 2: ([4, 4, 4, 4, 8])

Masibale

Ukuhambisana l	2^l − 1	I-R(l)
4	15	0.0586
8	255	0.9961

Ihlanganisa u-R(l) kumalebula ahlukene ahlobene

Ukusebenzisa lokhu ukubala (ERR) okuphelele kuzo zombili izilinganiso, sithola:

Izinga 1: (ERR) ≈ 0.99
Izinga 2: (ERR) ≈ 0.27

Uma sibheka ngokukhethekile umnikelo wento ngesikolo sokuhlobana esingu-8, siyakubona lokho ukwehla komnikelo walelo themu ngu-6.36x. Uma inhlawulo ibiwumugqa, ukwehla bekungaba ngu-5x.

Isimo	Umnikelo wokuhambisana-into engu-8
Izinga 1	0.9961
Izinga 5	0.1565
Isici sokulahla	6.36x

Umehluko ekunikeleni ngokushintsha kwezinga

Inzuzo Eyengeziwe Yesaphulelo Ejwayelekile (NDCG)

Inzuzo Eyengeziwe Eyehlisiwe Ejwayelekile ((NDCG)) enye inketho enhle ekufanelekela kahle ukukala imiphumela yosesho. (NDCG) yakhelwe phezu kwemibono emibili esemqoka

Inzuzo: Izinto nge amaphuzu aphezulu ahlobene abaluleke kakhulu
Isaphulelo: izinto ezivela kamuva zibiza kancane kakhulu njengoba abasebenzisi abanaki kakhulu izinto zakamuva

I-NDCG ihlanganisa umqondo Wokuzuza Nesaphulelo ukuze udale amaphuzu. Ukwengeza, iphinda yenza amaphuzu ajwayelekile ukuze avumele ukuqhathanisa phakathi kwemibuzo eyahlukene. Ngokusemthethweni, inzuzo nesaphulelo kuchazwa ngokuthi

(mathrm{Gain} = 2^{l_r}-1)
(mathrm{Isaphulelo} = log_2(1+r))

lapho (l) kuyilebula yokuhlobana kwento endaweni (r) futhi (r) indawo lapho yenzeka khona.

Inzuzo inefomu lokuchasisa, eliklomelisa ukuhambisana okuphezulu. Lokhu kuqinisekisa ukuthi izinto ezinomphumela ophezulu wokuhambisana zifaka isandla kakhulu. Isaphulelo se-logarithmic sijezisa izinga lakamuva lezinto ezifanele. Kuhlanganiswe futhi kwasetshenziswa kulo lonke uhlu olusezingeni, sithola Inzuzo Eyengeziwe Eyehlisiwe:

$$mathrm{DCG@K} = sum_{r=1}^{K} frac{2^{l_r}-1}{mathrm{log_2(1+r)}}$$

ngohlu olubekiwe lwezinga (l_1, l_2, l_3, …l_k). (DCG@K) ukubala kuyasiza, kodwa amalebula ahlobene angahluka esikalini kuyo yonke imibuzo, okwenza ukuqhathanisa (DCG@K) kube ukuqhathanisa okungalungile. Ngakho sidinga indlela yokwenza kube jwayelekile (DCG@K) amanani.

Lokho sikwenza ngokwenza ikhompuyutha (IDCG@K), okuyinzuzo eqoqwayo enesaphulelo efanelekile. (IDCG) ubuningi obungenzeka (DCG) ngezinto ezifanayo, ezitholwe ngokuzihlunga ngokuhambisana ngohlelo olwehlayo.

$$mathrm{DCG@K} = sum_{r=1}^{K} frac{2^{l^*_r}-1}{mathrm{log_2(1+r)}}$$

(IDCG) imele izinga eliphelele. Ukwenza kube ngokwejwayelekile (DCG@K), sibala

$$mathrm{NDCG@K} = frac{mathrm{DCG@K}}{mathrm{IDCG@K}}$$

(NDCG@K) inezindawo ezilandelayo

Kuboshwe phakathi kuka-0 no-1
Iqhathaniseka kuyo yonke imibuzo
1 uku-oda okuphelele

Isibonelo: Okuhle vs Uku-oda Okubi Kancane

Kulesi sibonelo, sizothatha amazinga amabili ahlukene ohlu olufanayo bese siqhathanisa (NDCG) amanani abo. Cabanga ukuthi sinezinto ezinamalebula ahlobene esikalini esingu-0-3.
Izinga A

Rank     : 1 2 3 
Relevance: 3 2 1

Izinga B

Rank     : 1 2 3 
Relevance: 2 1 3

Sihlanganisa (NDCG) izingxenye, sithola:

Izinga	Inzuzo (2^l − 1)	Ilogi yesaphulelo₂(1 + r)	Ukunikela	B umnikelo
1	7	1.00	7.00	3.00
2	3	1.58	1.89	4.42
3	1	2.00	0.50	0.50

Izibalo ze-DCG zethemu ngayinye

I-DCG(A) = 9.39
I-DCG(B) = 7.92
I-IDCG = 9.39
I-NDCG(A) = 9.39 / 9.39 = 1.0
NDCG(B) = 7.92 / 9.39 = 0.84

Ngakho-ke, ukushintsha into efanelekile kude nesikhundla 1 kubangela ukwehla okukhulu.

I-NDCG: Ingxoxo Eyengeziwe

Isaphulelo esakha idinominetha yokubala (DCG) singokwemvelo ye-logarithmic. Ikhula kancane kakhulu kunomugqa.

$$mathrm{discount(r)}=frac{1}{mathrm{log2(1+r)}}$$

Ake sibone ukuthi lokhu kuqhathaniswa kanjani nokubola komugqa:

Izinga (r)	Linear (1/r)	I-Logarithmic (1 / log₂(1 + r))
1	1.00	1.00
2	0.50	0.63
5	0.20	0.39
10	0.10	0.29
50	0.02	0.18

I-Linear Decay vs Logarthmic Decay

(1/r) iyabola Ngokushesha
(1/log(1+r)) iyabola Kancane

Isaphulelo se-Logarithmic sijezisa ngokuhamba kwesikhathi amazinga anamandla kunesaphulelo somugqa. Umehluko phakathi kwerenki 1 → 2 mkhulu, kodwa umehluko phakathi kwerenki 10 → 50 mncane.

Isaphulelo selogi sinokuncishiswa kwe-marginal enciphayo ekujezisweni kwamazinga akamuva ngenxa yokuma kwe-concave. Lokhu kuvimbela (NDCG) ekubeni i-metric esindayo kakhulu lapho izinga 1-3 libusa umphumela. Inhlawulo yomugqa izoziba ukukhetha okuphusile kwehle kuya phansi.

Isaphulelo se-logarithmic siphinde sibonise iqiniso lokuthi ukunaka komsebenzisi kwehla kakhulu phezulu ohlwini bese kuncipha esikhundleni sokwehla ngokuhambisana nezinga.

Isiphetho

I-(MAP) kanye (MRR) zingamamethrikhi okubuyisa ulwazi oluwusizo, kodwa azifaneleki kahle kumasistimu wezinga wosesho lwesimanje. Nakuba (MAP) igxile ekutholeni wonke amadokhumenti afanelekile, (MRR) iphatha inkinga yezinga njengemethrikhi yendawo eyodwa. (MAP) kanye (MRR) zombili azinaki ukuhlobana kwezigaba kwezinto ekusesheni futhi zizithatha njengezinambambili: zifanelekile futhi azibalulekile.

(NDCG) kanye (ERR) zibonisa kangcono ukuziphatha kwangempela komsebenzisi ngokubala izikhundla eziningi, okuvumela izinto ukuthi zibe nezikolo ezingezona kanambambili, kuyilapho kunikeza ukubaluleka okuphezulu ezikhundleni eziphezulu. Kuzinhlelo zokukala zokusesha, lawa mamethrikhi azwela indawo awawona nje ukukhetha okungcono- ayadingeka.