[Machine learning algorithm and learning _2_ theoretical articles] 2.2 M_ _ logistic regression classification

First, the principle set forth

Algorithm type: supervised learning classification algorithm _

Input: numeric or nominal type (required nominal type hot encoded)


   Solving regression binary classification manner, by introducing a Sigmoid function y intermediate values ​​the actual value of y is mapped onto two categories.

Second, the algorithm selects


Third, the algorithm process


 X is a function 1.Sigmoid range is (-∞, + ∞), y is the range (0,1) is a monotonically increasing function;

2. predicted value y> 0.5 class 1 <class 0 to 0.5, y value may also be interpreted as a probability as 0 and class 1;

3. Similarly using the "least squares" concept, to obtain the best fit equation, to give the objective function;

4. to the objective function is minimized, need called " gradient descent "algorithm, which process is substantially as follows: on the hyperplane a similar mountains, from any point of view, computing partial derivatives, advances a certain distance along the partial derivatives of the negative direction (referred to as" learning rate "), until changes after the initial point of difference from the mobile small (referred to as "convergence") so far.

Metrics: least squares objective function: the least squares objective function solution: gradient descent

Fourth, the characteristics of

Advantages: simple, easy to understand and implement; computational cost is not high, fast, low storage resources.

Disadvantages: easy underfitting, classification accuracy may not be high. Outliers and missing values sensitive

Fifth, the code API


Guess you like

Origin www.cnblogs.com/everda/p/11347959.html