ãã¯ãããŠã®ãã¿ãŒã³èªèããéç§°ãã¯ããã¿ãã®ç¬¬2ç« ã®è§£èª¬ã§ãã
13ããŒãžçšåºŠã®ããã«èªã¿è§£ãã®ã«æéãããªããããŸããã»ã»ã» èºããã€ã³ãã¯2ç¹ãã£ãŠãããŒãã¹ãã©ããæ³ã§æžãããŠããèšç®åŒãæåŸã®ãã€ã¢ã¹ã»åæ£åè§£ã§ãããã¯ããŒãªã³å±éãªã©æç³»äººéã«ãšã£ãŠã¯é£ããå
容ãå«ãã§ããŠæéãããããŸãã
æ¹ããŠæ©æ¢°åŠç¿ã®ããŒãã«ã®é«ããæããŸãããCourseraæ©æ¢°åŠç¿å
¥éã³ãŒã¹ãæ°åŠå匷ãããŠãããŠããã£ããšæ¹ããŠæããŸãã!
â»ã¯ããã¿å匷系ã¯èšäºãæç³»ç€ŸäŒäººãã¯ããã¿ã§æ©æ¢°åŠç¿ãæ°åŒå«ããŠçè§£ãã奮éèšãã«ãŸãšããŸããã
å 容
2.1. èå¥èŠåãšåŠç¿æ³ã®åé¡
2.1.1. èå¥èŠåã®æ§ææ³
èå¥èŠåã¯ãå ¥åããŒã¿$x$ããã¯ã©ã¹$C_i \in \Omega = \{ C_1, \ldots, C_k \}$ãžã®ååã§ãã
ãã£ããçè§£ãé£ãæãåºãŠããŸãããå ¥åããŒã¿$x$ã«èå¥é¢æ°ã䜿ããšã¯ã©ã¹ã«åé¡ãããããšããæå³ã
| èšå·ç | æå³ | äŸ(ããããã) |
|---|---|---|
| $C_i$ | 1ã€ã®ã¯ã©ã¹(åé¡) | ã°ãŒãšããã¯ã©ã¹ |
| $C_i\in \Omega$ | ã¯ã©ã¹$C_i$ã¯ã¯ã©ã¹éå($\Omega$)ãæ§æããå (èŠçŽ ) | ããããããšããã¯ã©ã¹éåã®æ§æèŠçŽ ã®1ã€ã«ã°ãŒããã |
| ${C_1, \ldots, C_k}$ | å (èŠçŽ )ã¯${C_1, \ldots, C_k}$ | ã°ãŒã»ãã§ãã»ã㌠|
| åå(mapping) | éåAã®åå (èŠçŽ )ãããããéåBã®å (èŠçŽ )ãå¯äžæå®ãããããªèŠåããããšããAããBãžã®ååããšèšã | ã5æ¬ã®æã®ç¶æ ãããããããã(ã°ãŒã»ãã§ãã»ããŒ)ãžã®åå |
â»$\Omega$ ã¯å€§æåã®ãªã¡ã¬
以äžãåèã«ããŸããã
代衚çãªèå¥èŠåã®æ§ææ³ã¯ä»¥äžã®4ã€(ã©ããåŸã§è©³ããåŠç¿ããã®ã§ä»ã¯ããŸã調ã¹ãªã)ã
| èå¥èŠåæ§ææ³ | å 容 | äŸ |
|---|---|---|
| äºåŸç¢ºçã«ããæ¹æ³ | ãã¿ãŒã³ç©ºéã«ç¢ºçååžãä»®å®ãäºåŸç¢ºçãæå€§ã®ã¯ã©ã¹ã«åé¡ | ãã€ãºã®æå€§äºåŸç¢ºçæ³ |
| è·é¢ã«ããæ¹æ³ | å ¥åãã¯ãã«$x$ãšåã¯ã©ã¹ã®ä»£è¡šãã¯ãã«ãšã®è·é¢ãèšç®ããäžçªè¿ã代衚ãã¯ãã«ã«åé¡ | æè¿åæ³ |
| 颿°å€ã«ããæ¹æ³ | 颿°$f(x)$ã®æ£è² /æå€§å€ã§ã¯ã©ã¹ã決ãã | ããŒã»ãããã³ååŠç¿åè·¯ãSVM |
| æ±ºå®æšã«ããæ¹æ³ | èå¥èŠåã®çåœã«å¿ããŠæ¬¡ã®èå¥èŠåãé æ¬¡é©çšããæ±ºå®æšã®åœ¢ã§ã¯ã©ã¹ã決ãã | - |
2.1.2. æåž«ä»ãåŠç¿
èå¥èŠåã¯å ¥åããŒã¿(ç¹åŸŽãã¯ãã«)ããã¯ã©ã¹ãžã®ååã $y=f(x)$ãšãã颿°ãçšããŠè¡šçŸãããèå¥èŠåã®åŠç¿ã¯ãã®$f()$ãåŠç¿ããŒã¿ãçšããŠæ±ºããããšã§ããã
ç·åœ¢èå¥é¢æ°
2ã¯ã©ã¹åé¡ã®ç·åœ¢èå¥é¢æ°ã¯ä»¥äžã®ããã«**ç·åœ¢é¢æ°(å ç©)**ãçšããŠè¡šçŸããèå¥ã¯ã©ã¹ã¯çµæ$y$ã®æ£è² ã§æ±ºãããå ¥åããŒã¿ãæ£ããã¯ã©ã¹ã«å¯Ÿå¿ããããã«ãã©ã¡ãŒã¿$w$ã調æŽã
y = f(x;w) = w_1 x_1 + \cdots + w_d x_d = w^T x
| èšå·ç | æå³ |
|---|---|
| $w$ | $w$ã¯weightã®ç¥ã§ãéã¿ãããã¿ã¡ãŒã¿ãšãåŒã¶ã |
| $f(x;w)$ | ã»ãã³ãã³ã®å($x$)ã¯å€æ°ãåŸ($w$)ã¯ãã©ã¡ãŒã¿ ç¹åŸŽ$x$ãå ¥åããã©ã¡ãŒã¿$w$ã䜿ã£ã颿°ã§åºå |
å ç©
ã宿°ãã¯ãã«ã®æšæºå
ç©ããšã¯ä»¥äžã®ããã«å®çŸ©ãããããã§ãã
â»$\mathbb R^n$ã¯å®n次å
æ°ç©ºéãæå³ãã
x = \left[
\begin{array}{rrr}
x_1\\
x_2 \\
\vdots \\
x_d
\end{array}
\right],\,\,
w = \left[
\begin{array}{rrr}
w_1\\
w_2 \\
\vdots \\
w_d
\end{array}
\right]
\in \mathbb R^n
\\ (w, x) = w_1 x_1 + \cdots + w_d x_d
å³å¯ãªç·åœ¢
$f(x) = b + wx$ ã¯å³å¯ãªæå³ã§ç·åœ¢ã§ã¯ãªãã
å³å¯ã«ç·åœ¢ã§ããå Žåã¯éãåããã®åç $f(ax)=af(x),,, f(x+y)=f(x)+f(y)$ãæç«ããå¿
èŠãããã宿°é
$b$ãããã®ã§ç·åœ¢ã§ã¯ãªãããã®ãããªé¢æ°ãã¢ãã£ã³é¢æ°ãšåŒã¶ã
ãã ãã$b=w_0x_0(x_0 = 1)$ãšèãããšç·åœ¢é¢æ°ã«ãªãã®ã§ã$b$(宿°é
)ããã£ãŠãç·åœ¢èå¥é¢æ°ãšåŒã°ããŠããã
æåž«ããŒã¿
åŠç¿ãããããã«å
¥åããŒã¿ãšãã®ã¯ã©ã¹ãæå®ããããŒã¿ã察ã«ããåŠç¿ããŒã¿ãå¿
èŠãã¯ã©ã¹ãæå®ããããŒã¿ãæåž«ããŒã¿ãšåŒã³ãŸãã2ã¯ã©ã¹ã§æ£è² ãšãããš$t \in \{-1, +1\}$ãã¯ã©ã¹éå$t$ã®å
(èŠçŽ )ã¯-1ãš+1ãšããæå³ã
â»$t$ã¯æåž«ããŒã¿(Teacher Data)ã®"t"ã£ãœãã
ã¯ã©ã¹æ°ã3ã€ä»¥äžããå Žåã¯ãããŒå€æ°è¡šçŸãçšããŠãäŸãã°4ã¯ã©ã¹ã®å Žåã¯$t=(0, 1, 0, 0)^T$ã®ãããªæåž«ããŒã¿ã«ããããã®ãããªæåž«ããŒã¿ãK察1笊å·å(1-of-K coding)ãšããã
åŠç¿ã«çšãããããã¹ãŠã®å
¥åãšæåž«ããŒã¿ã®å¯Ÿã®éåãæåž«ããŒã¿ã»ãããšããã$D_L$ã§è¡šãã
åŠç¿ã®ç®çã¯åŠç¿ããŒã¿ãæ£ããèå¥ã§ãã$w$ãæ±ããããšã§ãåŠç¿ããŒã¿ã1ã€ãã€äœ¿ã£ãŠ$w$ãå°ããã€ä¿®æ£ããŠããæ¹æ³ãªã©ãããã
åŠç¿ã«äœ¿çšããªãã£ã**ãã¹ãããŒã¿ã»ãã($D_T$)**ã§æ§èœè©äŸ¡ãããã
2.1.3. æåž«ä»ãåŠç¿ãšç·åœ¢ååž°
æåž«ããåŠç¿ã®ååž°ã«é¢ããŠã§ããCourseraæ©æ¢°åŠç¿å
¥éãªã³ã©ã€ã³è¬åº§ã®1é±ç® - ç·åœ¢ååž°ãšç·åœ¢ä»£æ°æ¹ããããããããå
容ãšããŠãã¡ãã§ååã§ãã

ãã®æ¬ã§ã¯ãåé¡ãäž»çŒãªã®ã§ååž°ã«é¢ããŠã¯å²æãããã§ãã
2.1.4. æåž«ãªãåŠç¿
å
¥åããŒã¿éã®è·é¢ãé¡äŒŒåºŠãçµ±èšçãªæ§è³ªã«åºã¥ããŠãã¯ã©ã¹ãèªåçã«çæ(ã¯ã©ã¹ã¿ãªã³ã°)ããæåž«ãªãåŠç¿ããã(èªå·±çµç¹ååŠç¿ãšãåŒã¶)ã
äžéšã®ããŒã¿ã®ã¿æåž«ãã€ããã»ãã¯æåž«ãªãã§åŠç¿ãè¡ã**圢質å°å
¥åŠç¿(transductive learning)**ãšããåŠç¿æ³ãããã
èšäºãåæåž«ããåŠç¿_Semi-Supervised LearningããèŠãã»ãã圢質å°å
¥åŠç¿ã ãã§ãªãããã倧ããªæŠå¿µã§ãããåæåž«ããåŠç¿ãããããããããŸãšãŸã£ãŠããŸãã
2.2. æ±åèœå
(åŠç¿ããŒã¿ã§ã¯ãªã)æªç¥ã®ããŒã¿ã«å¯Ÿããèå¥èœåãæ±åèœåãšããããã®èª€å·®ãæ±å誀差ãšããã
2.2.1. åŠç¿ããŒã¿ãšãã¹ãããŒã¿ã®äœãæ¹
å
šããŒã¿ãæåž«ããŒã¿ã»ãã$D_L$ãšãã¹ãããŒã¿ã»ãã$D_T$ã«åå²ããŸãã

以äžãçšèªã§ã($\epsilon$ã¯ãã€ãã·ãã³ããšèªã¿ãŸã)ã
| çšèª | 説æ | äŸ |
|---|---|---|
| 誀ãç$\epsilon(p_L, p_T)$ | $D_L$ã§åŠç¿ã$D_T$ã§ãã¹ããããšãã®èª€ãç | 1äžæã®ç¡¬è²šã®ãã¡8000æã®åŠç¿ããŒã¿ã»ããã䜿ã£ãŠåŠç¿ã2000æã§ãã¹ããããšãã®èª€ãç |
| æ¯éå£ | ãã¹ãŠã®ããŒã¿ | æµéãããã¹ãŠã®100å硬貚 |
| çã®ååž($p$) | æ¯éå£ã®$d$次å ç¹åŸŽååž | æµéãããã¹ãŠã®100å硬貚ã®ååž |
| åã(ãã€ã¢ã¹) | $D_L, D_T$ã®ååž$p_L, p_T$ãšçã®ååž$p$ãšã®ãã | æµéãã100åç¡¬è²šå šéšãšäœ¿çšãã1äžæã®$D_L, D_T$ã®ååžã®ãã |
| çã®èª€ãç($\epsilon(p, p)$) | çã®ååž$p$ã«åŸã£ãåŠç¿ããŒã¿ãšããŒã¹ãããŒã¿ã䜿ã£ããšãã®èª€ãç | æµéãã100å硬貚ã®ååžã§ã®ã¢ãã«ã®èª€ãç |
| åä»£å ¥èª€ãç($\epsilon(p_L, p_L)$) | åŠç¿ããŒã¿ã§ãã¹ãããå Žåã®èª€ãç | 8000æã®åŠç¿ããŒã¿ã»ããã䜿ã£ãŠåŠç¿ããå Žåã®åŠç¿ããŒã¿ã®èª€ãç |
以äžãåŠç¿ãšãã¹ãã«ããŒã¿åå²ããæ¹æ³ã§ãã
(1) ããŒã«ãã¢ãŠãæ³(holdoutæ³)
ããŒã¿å
šäœãäºåå²ããäžæ¹ãåŠç¿ã«äœ¿ã($p_L$)ãšä»æ¹ããã¹ãã«äœ¿ã($p_T$)ããã®ãšãã®èª€ãçã**ããŒã«ãã¢ãŠã誀ãç(holdout error)**ãšåŒã³ã$\epsilon(p_L, p_T)$ã§è¡šãã
$å代å
¥èª€ãç \leq çã®èª€ãç \leq ããŒã«ãã¢ãŠã誀ãç$ ã®é¢ä¿ãæãç«ã€ã
E_{D_L}\{\epsilon(p_L, p_L)\} \leq \epsilon(p, p) \leq E_{D_T}\{\epsilon(p_L, p_T)\} \\
â»ããã§ã®Eã¯Expectation(æåŸ å€)ã®é æå
(2) äº€å·®ç¢ºèªæ³(cross validationæ³)
äº€å·®ç¢ºèªæ³(cross validationæ³)ã¯ã亀差æ€èšŒæ³ãšãåŒã°ãããŒã¿ãmåã®ã°ã«ãŒãã«åå²ããŸãã1ã°ã«ãŒãããã¹ãããŒã¿ã»ããããã®ä»ãåŠç¿ããŒã¿ã»ãããšãmååŠç¿ãšãã¹ãã宿œã
äŸãã°3åå²($m=3$)ã«ããå Žåã¯äžèšã®ãšããã§ãã

誀èå¥çã®äºæž¬å€ã¯$i$çªç®ã®ã°ã«ãŒãã§ãã¹ããããšãã®èª€ãçã$\epsilon_{-1}$ã§è¡šããšä»¥äžã®éãã§ãã
\epsilon = \frac{1}{m} \sum_{i=1}^m \epsilon_{-i}
(3) äžã€æãæ³(leave-one-outæ³)
äº€å·®ç¢ºèªæ³ã®é¡äŒŒæ¹æ³ã§ãžã£ãã¯ãã€ãæ³ãšãåŒã°ããŠããŸããäº€å·®ç¢ºèªæ³ã¯ã°ã«ãŒãã§åå²ãã°ã«ãŒãã«å¯ŸããŠãã¹ããããŠããŸããããäžã€æãæ³ã§ã¯1ä»¶ã®ããŒã¿ã ãããã¹ãçšã«ããŠã1件以å€ã®ããŒã¿ã§åŠç¿ãããŸãã
ããŒã¿æ°åã ãåŠç¿ã»ãã¹ããç¹°ãè¿ã(100ä»¶ãã£ãã100ååŠç¿ãšãã¹ããå®è¡ãã)ã®ã§æéãããããŸãã
(4) ããŒãã¹ãã©ããæ³(bootstrapæ³)
Nåã®ããŒã¿ã«å¯ŸããŠNåã®åŸ©å
æœåºãããããšã§ãããŒã¿Nåã®ããŒãã¹ãã©ãããµã³ãã«ãäœãã

埩å
æœåºã¯æ¢ã«æœåºãããããŒã¿ãå«ããŠæœåºå
ãšãããããããŒã¿ãéè€ããå Žåããã(äžã®å³ã§ããããŒãã¹ãã©ãããµã³ãã«ã§ã®ããŒã¿#3ãéè€)ã
ããŒãã¹ãã©ãããµã³ãã«ããåŸããã誀èå¥çã$\epsilon(N^*,N^*)$ãšè¡šããããšããšã®ããŒã¿éåNããã¹ãããŒã¿ãšããŠåŸããããèå¥ç$\epsilon(N^*,N)$ãšã®å·®ãæ±ããã
bias = \epsilon(N^*,N^*) - \epsilon(N^*,N)
ããŒãã¹ãã©ãããµã³ãã«ãè€æ°(æäœã§ã50çšåºŠããã)çæããäžèšã®å¹³åãããã€ã¢ã¹ãæšå®ãã誀èå¥çã®äºæž¬å€$\epsilon$ãæ±ããã
â»$\overline{bias}$ã¯ãã€ã¢ã¹ã®å¹³åã
\epsilon = \epsilon(N, N) - \overline{bias}
Nåã®åŸ©å æœåºã§1床ãµã³ãã«ããã確ç$p'$
p' = 1-(1-\frac{1}{N})^N \approx 1-e^{-1} = 0.632
($1-\frac{1}{N})^N=e^{-1}$ã®éšåãè£è¶³ã
ãããããã€ãã¢æ°$e$ã¯$f(x)=a^x$ã®ææ°é¢æ°ã§$f'(0)=1$ãã€ãŸã$x=0$ã®åŸãã1ãšãªãåº$a$ã®å€ã極éã®èšç®åŒã«ãããšä»¥äžã®éãã
\lim_{h\rightarrow0}\frac{e^h-1}{h}=1
$h$ã$x$ã«ããŠã
\lim_{x\rightarrow0}\frac{e^x-1}{x}=1
ããã«$e^x-1=u$ãšãããš$e^x=1+u$ , $x=\log{(1+u)}$ãšãªãã$x\rightarrow0$ã®ãšã$u\rightarrow0$ãšãªãã®ã§ä»¥äžã«å€åœ¢(å³èŸºã¯$=1$ã§åºå®)ã
\lim_{x\rightarrow0}\frac{e^x-1}{x}=
\lim_{u\rightarrow0}\frac{u}{\log{(1+u)}}=
\lim_{u\rightarrow0}\frac{1}{\frac{\log{(1+u)}}{u}}=1
ãããã$\lim_{u\rightarrow0}\frac{\log{(1+u)}}{u}=1$ãšãªãã$u$ã$x$ã«çœ®ãæã$\lim_{x\rightarrow0}\frac{\log{(1+x)}}{x}=1$ãèšããã$\frac{1}{x}$ãåã«åºããŠå³èŸºã®$=1$ã®éšåã$\log{e}$ã«å€ããæåŸã«$\log$ã䞡蟺ãšãåãã
\lim_{x\rightarrow0}\frac{1}{x}{\log{(1+x)}}=
\lim_{x\rightarrow0}{\log{(1+x)}}^\frac{1}{x}=\log{e} \\
\lim_{x\rightarrow0}{(1+x)}^\frac{1}{x}=1
ããã«$x=\frac{1}{v}$ã«çœ®ãæãããš$x\rightarrow0$ã®ãšã$v\rightarrow \pm \infty$ãšãªãã®ã§ä»¥äžã«å€åœ¢ã
\lim_{v\rightarrow\pm \infty}(1+\frac{1}{v})^v=e \\
\lim_{x\rightarrow\pm \infty}(1+\frac{1}{x})^x=e
ããšã¯$x$ã$-N$ã«ããã°ã¯ããã¿ã«æžãããŠããåŒã«ãªããŸãïŒ
â»$N\rightarrow \pm \infty$ã¯çç¥
(1-\frac{1}{N})^N \approx e^{-1}
P15èæ³šèšèŒéšåã®èšç®åŒ
\begin{eqnarray}
& &
(x+y)^n \\[5pt]
&=&
{}_n \mathrm{ C }_n x^n
+{}_n \mathrm{ C }_{n-1} x^{n-1} y
+{}_n \mathrm{ C }_{n-2} x^{n-2} y^2
+\cdots \\[5pt]
& & \cdots
+{}_n \mathrm{ C }_{k} x^k y^{n-k}
+\cdots
+{}_n \mathrm{ C }_1 x y^{n-1}
+{}_n \mathrm{ C }_0 y^n
\end{eqnarray}
äžèšã䜿ã£ãŠ$x = 1$ $y=-\frac{1}{N}$ãšããŠ$(1-\frac{1}{N})^N$ãå±éããŸããããã§è泚ã®å·ŠåŽã®å®æã§ãã
(1-\frac{1}{N})^N =
{}_N \mathrm{ C }_{N} 1^N +
{}_N \mathrm{ C }_{N-1}1^{N-1}(-\frac{1}{N})^1 +
{}_N \mathrm{ C }_{N-2}1^{N-2}(-\frac{1}{N})^2 + \cdots \\
= 1 + N(-\frac{1}{N}) + \frac{N(N-1)}{2!} \frac{1}{N^2} - \cdots \\
ãã®åŒã¯e^{-1}ã®ãã¯ããŒãªã³å±éããåŒãšåãã§ã($N\rightarrow \pm \infty$ã«ããŠããã®ã§åã)ã
e^x=1+x+\dfrac{x^2}{2!}+\dfrac{x^3}{3!}+\dfrac{x^4}{4!}+\cdots \\
e^{-1}=1-1+\dfrac{1}{2!}-\dfrac{1}{3!}+\cdots
éšåçã«$\frac{N(N-1)}{2!}$ãåãåºããŠå°ãè£è¶³ãããšä»¥äžã®ããã«å±éã§ããŸã($N\rightarrow \pm \infty$ã®å Žå$\frac{1}{N}=0$ãªã®ã§)ã
ããã§è泚éšåãã¹ãããªããŸããïŒ
\lim_{N\rightarrow \pm \infty}\frac{N(N-1)}{2!} \frac{1}{N^2} =
\lim_{N\rightarrow \pm \infty}\frac{1-\frac{1}{N}}{2!} =
\frac{1}{2!}
以äžã¯åèã«ãããªã³ã¯ã§ãã
ç°¡åãªææ³æ¯èŒ
åææ³ã®åŠç¿ã»ãã¹ãåæ°ã¯ä»¥äžã®éãã
| ææ³ | åŠç¿ã»ãã¹ãåæ° |
|---|---|
| ããŒã«ãã¢ãŠãæ³ | 1å |
| äº€å·®ç¢ºèªæ³ | ã°ã«ãŒãæ°å |
| äžã€æãæ³ | ããŒã¿æ°å |
| ããŒãã¹ãã©ããæ³ | ä»»æã®æ°(50åçšåºŠä»¥äžãšèšãããŠãã)+1 "+1"ã¯å ããŒã¿ã«å¯ŸããåŠç¿ã»ãã¹ã |
Cross-Validation and Bootstrap for Accuracy Estimation and Model SelectionãComparing the Bootstrap and Cross-Validationã«äº€å·®æ€èšŒæ³ãšããŒãã¹ãã©ããæ³ã®æ¯èŒãæžãããŠããŠãéåžžã«è峿·±ã(ãã©èŠçšåºŠã ãã亀差æ€èšŒæ³ã®æ¹ãåªããŠããïŒ)ã
2.2.2. æ±åèœåã®è©äŸ¡æ³ãšã¢ãã«éžæ
誀ããç®æšããå°ãããªããªãå Žåã«èå¥é¢æ°ãå€ãããèå¥é¢æ°ãå€ããããã©ã¡ãŒã¿ãå€ãããããããã¹ãããŒã¿ã«å¯Ÿãã誀ããæãå°ãããªããã©ã¡ãŒã¿ãéžæããæ¹æ³ãã¢ãã«éžæãšããã
ååž°(颿°è¿äŒŒ)ã®äŸã ãšã以äžã®åŒã§èããããšãã§ããã
| å 容 | åŒ | åè |
|---|---|---|
| ãµã³ãã«ããŒã¿é¢æ° | $f(x)=0.5+0.4*\sin(s \pi x) + \epsilon =$ $h(x)+\epsilon$ |
å¹³å0ã忣0.05ã®æ£èŠååžã«ãã€ãº $\epsilon$ãä¹ã£ã颿° |
| p次å€é åŒã«ããè¿äŒŒæ²ç· | $y(x;a)=a_0+a_1x+a_2x^2+\cdots+a_px^p$ $a=(a_0,a_1,\cdots,a_p)^T$ |
$y(x;a)$ã¯ç¹åŸŽ$x$ãå ¥åããã©ã¡ãŒã¿$a$ã䜿ã£ã颿° |
| å¹³åèªä¹èª€å·®(MSE: Mean Square Error) | $\int f(y(x;D)-h(x))^2 p(x) dx = $ $E\{ (y(x;D)-h(x))^2\}$ |
çå®ã®é¢æ°ã𿩿¢°åŠç¿ã«ããã€ãã£ãã¢ãã«é¢æ°ã®çµæãšã®å·®ã®èªä¹ å€é åŒãã©ã¡ãŒã¿$a$ã¯åŠç¿ããŒã¿ã»ãã$D$ãæ±ºãããšäžæã«æ±ºãŸãã®ã§$y(x;D)$ãšããŠãã $p(x)$ã¯$p(x)=1$ã®ç¢ºçå¯åºŠé¢æ° |
| å¹³åèªä¹èª€å·®(è€æ°ããŒã¿ã»ãã) | $MSE_D=E_D\{ (y(x;D)-h(x))^2\}$ | æ©äŒåŠç¿ã®å Žåã¯ããŒã¿ã»ããã«äŸåãããããããŒã¿ã»ããã«å¯Ÿããå€ã ãšããããšãæç€º |
MSEã«é¢ããŠã¯ã**ãã€ã¢ã¹ãšåæ£(ããªã¢ã³ã¹)**ã«åè§£ã§ããŸããèšèã®å®çŸ©ã¯äžèšã®Wikipediaåç §ã
ãã€ã¢ã¹ïŒåãïŒ
åŠç¿ã¢ã«ãŽãªãºã ã«ãããŠã誀差ã®ãã¡ãã¢ãã«ã®ä»®å®ã®èª€ãã«ç±æ¥ããåããã€ã¢ã¹ã倧ããããããšã¯ãå ¥åãšåºåã®é¢ä¿ãé©åã«æããããŠããªãããšãæå³ããéå°é©åããŠããã
ããªã¢ã³ã¹ïŒåæ£ïŒ
誀差ã®ãã¡ãèšç·ŽããŒã¿ã®æºããããçããåãããªã¢ã³ã¹ã倧ããããããšã¯ãæ¬æ¥ã®åºåã§ã¯ãªããèšç·ŽããŒã¿ã®ã©ã³ãã ãªãã€ãºãåŠç¿ããŠããããšãæå³ããéå°é©åããŠããã
ãã€ã¢ã¹ãšåæ£ã¯ãã¬ãŒããªãã®é¢ä¿ã«ããããã®é¢ä¿ããã€ã¢ã¹ã»åæ£ãã¬ãŒããªããšåŒã³ãŸãã
ãã€ã¢ã¹ã»åæ£åè§£ã«ã€ããŠ
(y(x;D)-h(x))^2 = (y(x;D)-E_D{y(x;D)}+E_D{y(x;D)}-h(x))^2
åŒãå±éããŠãããŸãã
(y(x;D)-E_D\{y(x;D)\}+E_D\{y(x;D)\}-h(x))^2 = \\
(y(x;D)-E_D\{y(x;D)\})^2 +
(E_D\{y(x;D)\}-h(x))^2 + \\
2(y(x;D)-E_D\{y(x;D)\})(E_D\{y(x;D)\}-h(x))
$MSE_D=E_D\{ (y(x;D)-h(x))^2\}$ã®ããã«å®éã¯æåŸ å€ããšã£ãŠããã®ã§ãæçµé ã®$2(y(x;D)-E_D\{y(x;D)\})(E_D\{y(x;D)\}-h(x))$ã®éšåã¯ãŒãã«ãªããŸãããªããªãæçµé ã®ååéš$(y(x;D)-E_D\{y(x;D)\})$ãæåŸ å€ããšããšãŒãã«ãªãããã§ãã
E_D\{(y(x;D)-E_D\{y(x;D)\})\} = E_D\{y(x;D)\}-E_D\{y(x;D)\} = 0
ããã§åŒãå°ãåçŽåã§ããŸããã
MSE_D=E_D\{ (y(x;D)-h(x))^2\} \\
= (y(x;D)-E_D\{y(x;D)\})^2 + (E_D\{y(x;D)\}-h(x))^2
ãã®ååéšå$(E_D\{y(x;D)\}-h(x))^2$ããã€ã¢ã¹ã§ãã¢ãã«ãäºæž¬ããå€ãšå®éå€ãšã®ä¹é¢ã§ãã
åŸåéšå$(y(x;D)-E_D\{y(x;D)\})^2$ãããªã¢ã³ã¹ã§ãããŒã¿ã»ããã«äŸåããã¢ãã«éã®åæ£ã§ãã
以äžã¯çè§£ããäžã§åèã«ãããªã³ã¯èšäºã§ãã
- ãã€ã¢ã¹-ããªã¢ã³ã¹åè§£ïŒæ©æ¢°åŠç¿ã®æ§èœè©äŸ¡: ãã€ã¢ã¹ - ããªã¢ã³ã¹åè§£ã«ã€ããŠäžçªãããã
- æ©æ¢°åŠç¿ã®ãã€ã¢ã¹ããªã¢ã³ã¹åè§£ãå€å žçãªçµ±èšåŠããäžå¯§ã«è§£èª¬ãã: çµ±èšåŠãå«ããŠãã€ã¢ã¹ããªã¢ã³ã¹åè§£ã解説
- ãçµ±èšçã«ææãã ãã§ã¯è¶³ããªãã¯ã±ïŒãã€ã¢ã¹ïŒããªã¢ã³ã¹åè§£ã®ã¯ãªã: ãã€ã¢ã¹ãšããªã¢ã³ã¹ã®ã€ã¡ãŒãžå³ããããããã
- ãã€ã¢ã¹ãšããªã¢ã³ã¹ãçè§£ãã: ãã€ã¢ã¹ãšããªã¢ã³ã¹ã®Pythonã³ãŒããã
- æ¡ä»¶ä»ãæåŸ å€ïŒåæ£ã®æå³ãšæåå ¬åŒ: æ¡ä»¶ä»ãæåŸ å€ã«ã€ããŠã®åºæ¬
â»$\hat{a}$ã®ããã«*^*ãæåã®äžã«ã€ãããšãããããããšåŒã³ãŸããæšå®å€ã®ããšã
ãã¯ããã¿ãå匷系èšäºãªã³ã¯
| ç« | é ç® | æé(h) | é£æåºŠ | åŠãã ããš |
|---|---|---|---|---|
| 1 | ã¯ããã« | 8.9 | ç¹åŸŽã®å, ç¹åŸŽç©ºé, 次å ã®åªã | |
| 2 | èå¥èŠåãšåŠç¿æ³ã®æŠèŠ | 12 | ããŒã«ãã¢ãŠãæ³,äº€å·®ç¢ºèªæ³ ,äžã€æãæ³ ,ããŒãã¹ãã©ããæ³ ,ãã€ã¢ã¹ã»åæ£ãã¬ãŒããªã, éåŠç¿ | |
| 3 | ãã€ãºã®èå¥èŠå | 14.8 | ãã€ãºèå¥èŠå, ROCæ²ç· | |
| 4 | 確çã¢ãã«ãšèå¥é¢æ° | 18 |
|
å¹³åãã¯ãã«, å ±åæ£è¡å, æšæºå, ç¡çžé¢å, çœè²å, æ£èŠååž, æå°€æšå® |
| 5 | kæè¿åæ³(kNNæ³) | 8 |
|
æè¿åæ³, ãããã€å¢ç, kNN |
| 6 | ç·åœ¢èå¥é¢æ°(åå) | 30.4 |
|
æ£èŠæ¹çšåŒ |
| 6 | ç·åœ¢èå¥é¢æ°(åŸå) | ååã«æéã¯å«ã |
|
ãã£ãã·ã£ãŒã®ç·åœ¢å€å¥é¢æ°, å€å¥åææ³, ããžã¹ãã£ãã¯ååž° |
| 7 | ããŒã»ãããã³ååŠç¿èŠå | 13.5 |
|
å€å±€ããŒã»ãããã³, 誀差éäŒææ³, ã·ã°ã¢ã€ã颿° |
| 8 | ãµããŒããã¯ãã«ãã·ã³ | 14.7 |
|
ã«ãŒãã«ããªãã¯, Μ-SVM |
| 9 | éšåç©ºéæ³ | 15.4 |
|
äž»æååæ, ç¹ç°å€åè§£, CLAFICæ³, ã«ãŒãã«äž»æååæ, ã«ãŒãã«éšåç©ºéæ³ |
| 10 | ã¯ã©ã¹ã¿ãªã³ã° | 8.4 |
|
è·é¢ã®å ¬ç, ãã³ã³ãã¹ããŒè·é¢, K-平忳, èåæ³, æ··åæ£èŠååžã¢ãã« |
| 11 | è奿©ã®çµã¿åããã«ããæ§èœåŒ·å | 11.6 |
|
ããŒããªãŒã©ã³ãå®ç, æ±ºå®æš, ãã®ã³ã°, ã¢ãããŒã¹ã, ã©ã³ãã ãã©ã¬ã¹ã |