The term 1
1.1 Generalization
The learned ability to model for new samples called generalization
1.2 over-fitting - the key obstacle
Learner can fit all of the data sample, that some of the training sample its own characteristics as a potential general nature of all samples , resulting in a reduced ability of generalization phenomenon.
- Excessive function which assumes the characteristic variable
- Can only be alleviated, can not be eliminated
- Common cause factors: the ability to learn to have a good
1.2.1 Solution
- Effect of human or rounding algorithm using characteristic variables is not high
- Regularization: wherein all the variables but retained the size reduction parameter
1.3 error rate
Misclassification number of samples as a percentage of the total sample, set a small number of samples == category as a positive category ==
1.3.1 precision P (Precision)
Positive results in the machine predictable result of the user really need the proportion of eg: judge people get cancer, people really get cancer rate
1.3.2 recall / recall R (Recall)
The proportion of share of results of results users really need the machine successfully predict eg: all people get cancer, and can determine the ratio of people get cancer
- And P mutually exclusive
1.3.3 weigh the precision and recall
By changing the determination threshold value, a high threshold, a low P high R; low high low threshold P R
1.3.4 BEP
Value equal precision and recall
1.3.5 using the harmonic mean is determined based on precision and recall bad algorithm
Each statistical variables reciprocal reciprocal arithmetic average. Higher value to a lower weight
\ [\ frac {1} { F1} = \ frac {1} {2} (\ frac {1} {P} + \ frac {1} {R}) \ tag { 1.1} \]
\[ F_1=2\frac{PR}{P+R}\tag{1.2} \]
- Results constant is smaller than the arithmetic mean
- \ (F_1 \ in [0, 1] \) , the higher the better
1.4 chi-square distribution
The square of the standard normal distribution
1.5 Boolean data
I.e., logical data type value of 0 or 1