Slide 33
Slide 33 text
機械学習技術のコアは数学であるが
㻚㻜㻜㻓㻔
㻔㻜㻜㻜㻓㻚
㻕㻓㻓㻓㻓㻔
㻕㻓㻓㻓㻓㻚
㻕㻓㻓㻔㻓㻔
㻕㻓㻓㻔㻓㻚
㻕㻓㻓㻕㻓㻔
㻕㻓㻓㻕㻓㻚
㻕㻓㻓㻖㻓㻔
㻕㻓㻓㻖㻓㻚
㻕㻓㻓㻗㻓㻔
㻕㻓㻓㻗㻓㻚
㻕㻓㻓㻘㻓㻔
㻕㻓㻓㻘㻓㻚
㻕㻓㻓㻙㻓㻔
㻕㻓㻓㻙㻓㻚
㻕㻓㻓㻚㻓㻔
㼗㼌㼐㼈
crease and decrease in the number of
ncerning a fact
that a common-sense fact is found on the
ase in a similar manner as time passes. The
and expired states are represented by a uni-
ibution. In total, the temporal distribution
nce of a statement on the Web is modeled as
ibution.
matical formulation is as follows. We repre-
Recognition Model using a mixture distribu-
ans the probability that web page about a
be created at time t. It is expressed as a lin-
n of a Gaussian distribution N(t; µ, σ2) with
an exponential distribution f(t) with weight
= α1N(t; µ, σ2) + α2f(t) (1)
i: index for distributions (i ∈ {1, 2}).
αi
: weight for distribution i.
λ: parameter for the exponential distribution.
µ: mean vector for the Gaussian.
σ2: variance for the Gaussian.
φi: parameter vector (αi, λ, µ, σ2).
pi(xk
|φi): probability of xk
by distribution i.
Φ: parameter vector for the mixture model.
p(xk
|Φ): probability of xk
by the mixture model.
select initial estimated parameter vector Φ
until Φ converges to Φ do
Φ ← Φ
for each i do
initialize Ψi
, Mi
, Si
for each k do
ψik
← αipi(xk|φi)
p(xk|Φ)
Ψi
← Ψi
+ ψik
Mi
← Mi
+ ψik
xk
if i = 1 then
Si
← Si + ψik
(xk
− µ)2
αi
← Ψi
n
if i = 1 then
µ ← Mi
Ψi
, σ2 ← Si
Ψi
if i = 2 then
λ ← − Ψi
Mi
return Φ
This algorithm is based on the calculation in Appendix
A.
数理的内容の学習は，初学者にはハードルが高い
35