Home

Welcome to the stockER wiki!

The model I used

*Added one more convolution layer and fully connected layer, each.

---

Various Sets

1. Training Set: Used to train the model

2. Test Set: Used for test

3. Validation Set: Used before test(validation check)

---

Training Strategies

1. k-fold Cross-Validation

2. LOOCV(Leave-One-Out Cross Validation):

• Advantage: Less bias
• Disadvantage: Too many calculations (NFL)

3. Bootstrapping: sampling randomly from set D to D', m times. (n(D) = n(D') = m)

• So-called, out of bag prediction(~36.8% may not be sampled --> used for the testing phase, not training)
• Used if data set is hard to classify into groups & fewer data sets

4. Recall and Precision

• Precision: True Positive / (True Positive + False Positive): 양성이라 판정했을 때의 진짜 양성일 능력
• Recall: True Positive / (True Positive + False Negative): 양성을 양성이라 할 능력

PR_Graph PR Graph's Area under P and R may be the discrimination factor but uses an F1 rate instead.

5. Bias-Variance-Decomposition BVD_explanation

• Error = bias**2 + var + eps(noise)

6. t검정 내용 추가 요망...

---

Linear Models

**Projection onto the hyperplane <-> least square method <-> Euclidian distance minimization **

1. Multivariable Linear Regression: Least Square Method <-> Projection onto a hyperplane in Euclidian space f(x) = Wx + B ~ y

2. Log Linear Regression: f(x) = Wx + B ~ ln(y)

3.Logisitic Regression: Uses sigmoid function as a surrogate function of unit step function f(x) = Wx + B ~ ln(y/(1-y))

• W and B determined easily by Maximum Likelihood Theorem

LDA - Linear Discriminant Analysis: Minimizing the covariance of smaple sets

1. LMM(Lagrange Multiplier Method)
2. SVD(Singular Value Decomposition)
3. GRQ(Generalized Rayleigh Quotient)

Class Imbalance

Too much positive or negative value compared to a negative and positive value, respectively
Thus,

1. change system's threshold --> Rescaling method

2. Undersampling: getting rid of too much data

3. Oversampling: adding data

#MULTILABEL LEARNING???

---

dddd