Loading...

선형대수학 기본 용어 -중급자편 4-

1. normal matrix conjugate transpose $A ^{H}$에 대하여 $AA ^{H} =A ^{H} A$를 만족시키는 행렬 $A$를 normal matrix라고 부른다. 모든 원소가 실수인 행렬이라면 $AA ^{T} =A ^{T} A$인 행렬 $A$를 normal matrix라고 부른다. 2. orthogonal matrix 벡터 $x=$의 norm이라는 것은 $$\left \| x \right \| = \sqrt{x\cdot x} = \sqrt{\sum_{i=1}^{n}x_{i}^{2}}$$ the construction of the norm of a vector is motivated by a desire to extend the intuitive notion of the len..

선형대수학 기본 용어 -중급자편 3-

1. adjugate matrix 주어진 square matrix A의 모든 원소를 대응하는 cofactor로 바꾸고 transpose한 행렬을 말합니다. 즉 $A _{nn} = \left \{ a _{ij} \right \}$에 대하여 $a _{ij}$의 cofactor $c _{ij} =(-1) \left \| M _{ij} \right \|$로 치환하여 만든 행렬 $C _{nn} = \left \{ c _{ij} \right \}$의 transpose $C ^{T} =adjA= \left \{ c _{ij} \right \} ^{T}$를 adjugate matrix라고 부릅니다. 이 행렬이 중요한 이유는 $A _{nn}$의 inverse matrix를 구하게 만들어줍니다. 즉 $A _{nn}$의 역..

2021. 11. 4. 21:20

선형대수학 기본 용어 -중급자편 2-

1. eigenvalue 행렬 $A$에 대하여 등식 $Au= \lambda u$을 만족시키는 어떤 실수 $\lambda$를 $A$의 eigenvalue라 부르고 이에 대응하는 벡터 $u$를 eigenvector라고 부릅니다. $A _{nn}$의 eigenvalue는 n개가 존재하는데 각각의 eigenvalue에 대하여 대응하는 eigenvector는 무수히 많을 수 있습니다. $Au= \lambda u$를 생각하면 eigenvector $u$는 선형변환 $A$에 의해 변환을 하더라도 단순히 길이만 변하거나 방향이 반대만 되는 벡터를 의미합니다. 1) A의 eigenvalue의 곱은 A의 determinant와 같습니다. $$det(A)= \prod _{i=1} ^{n} \lambda _{i}$$ 2) A..

2021. 11. 3. 21:27

선형대수학 기본 용어 -중급자편 1-

square matrix의 어떤 특성을 나타내주는 하나의 scalar value로 mapping하는 함수를 말합니다. 구체적으로 determinant가 0이 아니라는 것은 주어진 square matrix가 invertible이라는 것과 동치가 됩니다. 행렬 $A=\begin{pmatrix} a_{11} & a_{12} & \cdots & a_{1n}\\ a_{21}& a_{22}& \cdots & a_{2n}\\ \vdots & \vdots & \vdots & \vdots \\ a_{n1}& a_{n2}& \cdots & a_{nn} \end{pmatrix}$가 주어질 때 기호로 $$det(A)=\left | A \right |=\begin{vmatrix} a_{11} & a_{12} & \cdots &..

2021. 11. 3. 20:57

CatBoost 모형

그 이름 Cat가 categorical feature를 뜻하는데 categorical 변수에 최적화되어있다고 논문에서 주장하고 있습니다. “ Two critical algorithmic advances introduced in CatBoost are the implementation of ordered boosting, a permutation-driven alternative to the classic algorithm, and an innovative algorithm for processing categorical features “ 논문에서 언급하는 ordered boosting은 일반적인 boosting이 모든 데이터 row에 대해 gradient 업데이트 과정을 거쳤다면 Catboost는 다음..

Light GBM 모형

LightGBM은 그 이름 Light에서도 알 수 있듯이 computational cost를 줄이기 위해 여러 가지 고급기술?을 도입하였다고 합니다. 그 중 하나인 Gradient based one sided sampling에 대해 직관적으로 이해해봅시다. 위에서 gradient boosting 기법을 다시 한번 생각해보면 각 data row마다 residual을 계산하여 learning rate를 이용한 예측값을 갱신하는 과정이 있었습니다. 이것은 마치 data row가 가지는 gradient로부터 gradient descent를 하는 과정을 연상시키죠. 그런데 Gradient based one sided sampling은 이름에서도 알 수 있듯이 gradient를 기반으로 필요없는 데이터는 버리겠다는..