1, machine learning tasks
- Machine learning task is to predict the mark from the property, namely determination of probability
2, discriminant model
- For example no, discriminant model can be obtained in accordance with markers, which can be directly determined out
- For dichotomous tasks, actually get a score, when the score is greater than the threshold, compared with positive class, otherwise it is a like a
- As shown in FIG left, is actually directly determined the boundaries
- Discriminant model is called "discrimination" model, because it is based on "discrimination"
- Common discriminant model
- Linear regression models
- Support vector machine ( SVM )
- Logistic regression ( LR )
- Neural Networks ( NN )
- Gaussian process ( Gaussian Process )
- Conditional Random CRF
- CART (Classification and regression tree)
- Boosting
3, the model formula
- For example no, need to model formula is obtained between the joint probability distribution of different labels, the probability of winning big
- As shown in the right side of FIG, and the absence of any boundary, for example no (red triangles), the sum of two joint probability distribution (two classes), compare, as a predicted high probability category
- Generation model is called "generation" model, because it is based on joint probability forecast
- The joint probability can be interpreted as the probability of the "Generate" distribution of the sample (or called basis)
- Specifically,
- Known machine learning, from a set of candidate selected from one of
- The sample may have, (the X-| Y_2) , (the X-| Y_3) , ...... ,
- The actual data is how to "generate" depends on
- Then the final results of the election predicted "generate" the greatest probability that
- Familiar Naive Bayes all know, for input, you need to obtain several joint probability, then the larger one is to predict results
- Common generative model
- Discriminant analysis: Gaussian discriminant model
- Naive Bayes ( Naive Bayes )
- Gaussian mixture model ( Gaussians )
- K neighbors ( KNN )
- Hidden Markov model ( the HMM )
- Bayesian networks
- Sigmoid belief networks ( S igmoid Belief Networks )
- MRF ( Markov Random Fields )
- Deep belief network DBN
- Latent Dirichlet Allocation ( LDA , Latent Dirichlet Allocation )
- Multi-expert model ( at The Mixture of Experts Model )
4, analog Case
- Determine a sheep or goat sheep
- Discriminant model:
- Learning from historical data to model
- This feature extraction sheep to predict this sheep is a sheep or goat
- That can give the sheep a sheep directly from the probability of feature
- Generative model:
- According to first learn the characteristics of goat model of a goat
- It features a learning model sheep sheep
- Extracting features from sheep, a goat model into what is the probability seen, put sheep seen what is the probability model
- That is a big probability model class sheep belongs to this category
- That is, the generative model for each class to have a try, that is, the last resulting most probable
Reference blog:
https://www.zhihu.com/question/20446337 (machine learning "model determination," and "generative model" What is the difference)
https://blog.csdn.net/u010358304/article/details/79748153 (Model VS generated discriminant model)
https://www.nowcoder.com/questionTerminal/e7ac0572b29a490da333d2c7ff8623ac?orderByHotValue=0&done=0&pos=1&mutiTagIds=631&onlyReference=false (discriminant model formula model)