Easy-to-understand Machine Learning

 (Preface)I wrote a machine learning ticket yesterday. Let's write one today. This book is mainly used for beginners and is very basic. It is suitable for sophomores and juniors. Of course, it is also applicable if you have not read machine learning before your senior or senior. Machine Learning is essential for both research on intelligence and other things. You can see that GFW uses machine learning. Do you have to learn more about it.
(Full-text structure)In fact, I thought about it. Learning a subject, listing a bunch of books, and commenting on it is actually of little use to beginners. He doesn't know what it is, you can only leave a bunch of cold lines for a comment. So first, let's talk about what machine learning does? What is the basic content.
(What is this product ?)Machine learning, also known as data mining and pattern recognition, has many definitions. In other words, what machine learning needs to do is to have some data (such as your friends and their speeches). We need to process the data, we hope to get the information we want from the data (for example, which of these friends are connected to you ). From the above example, we can see that machine learning is actually a imitation of human intelligence and the only way to achieve human and higher intelligence.

(What are these items ?) What does he have in general?

 

(Difficult machine learning theory, mathematics)

Part 1: underlying theory of machine learning: There are some underlying theories of machine learning, such as reasoning and planning, approximate calculation theory, regularization, Promotion Theory, core method, and of course well-known statistical machine learning theory. This part is not intended for beginners.

IThese theories are actually summarized in practice and cannot be understood without basic machine learning methods;IIThe mathematics required by these theories is very high. You should not mention anything about advanced mathematics, these theories should at least have the general knowledge of functional analysis, the general knowledge of optimization theory, the general knowledge of matrix theory, the mathematical foundation of higher probability theory, random analysis, and so on.3.I think these theories are meaningless for most people who just want to use machine learning methods. You just want to use machine learning. These theories are all entertaining to you.

(The main method of machine learning, I can see it in mathematics)
Part 2: Machine Learning Methods: This part should be learned by beginners and must be learned.
(According to the data format, the general situation is given.) The data processing can be divided into the following parts:
1. Supervised Learning: That is, your data has been processed very well, and you can clearly identify which data is used.
2. unsupervised learning: Your Data has started in Taiyuan. It's all a pile of data and you don't know which one is the case.
3. Semi-Supervised Learning: Supervised learning is good, but data requirements are too high. unsupervised learning is not very effective, but has low requirements on data. So let's make a compromise and mark a part of the data first, then, we use supervised learning to label other unlabeled data. If the accuracy produced by the algorithm is below a certain parameter, we will give this data to people.
4. Migration Learning: For example, if we have a method for similarity analysis between books, can this method be used for the analysis of online users for everyone? This is migration learning.
5. Enhanced Learning: Learn based on environment feedback.

6. "various random learning": In fact, there are still many wonderful learning methods, not the above five more common learning methods, they are generally a variety of wonderful ideas, coupled with a variety of love mathematical theory derived from. Because it is not very mature, so beginners do not have to worry too much about this part.

(What is the data we generally process ?) 
The data we process is generally table-like. To put it bluntly, that is, each piece of data is a vector. (a few days ago, I saw a child saying that the vector has a direction. I think it is speechless, = !, Now you can't even imagine the direction of vectors in physics. Although vectors can be imagined as geometric shapes, this is the basis of analysis, but do not stick to ry ). Since each piece of data is a vector, it is clear that all data constitute a vector space. This vector space must have certain abstract imagination capabilities, not onlyEuropean SpaceAnd possiblyTopology... Now you will know why.

That knows the above content. Let's take a look at the most basic and core methods of machine learning. Of course, if you read more books than this, don't be surprised. I just listed the very common and common methods.

(The most commonMachine Learning Method = Basic Method * Expansion Method * application field)
 (Basic method) 
1. Association Analysis: Now the data is a sales record. We need to find out which products are often bought together. At this time, there will be two main methods: the Apriori method, mainly pruning, AIS and stem are familiar to him. Among them, stem is the association analysis algorithm used for SQL language, and FP-growth is mainly used to build a tree to accelerate the algorithm through this structure; vertical Association mining and array methods are also available.
2. Decision Tree: There is an application called mind-reading. He keeps asking you to provide information about the people you want, deduct them, and finally find the people you want. This application seems magical. To put it bluntly, you can use a decision tree to create one. A decision tree is a tree with conditions on each side of the tree. The root node is the Start Node and the leaf node is the result node, continuously move to the corresponding Tree node based on the information on the edge until the leaf node gives the result. This is the decision tree. Decision trees are a type of algorithms, including ID3 and C4.5.
3. Sensor: Do you still remember the vector space I just mentioned? Every vector can be expressed as a point in space. If we can find a straight line, divide all points into two parts, one of which is Class A and the other is Class B. Then we have another point in the future. We only need to look at which side of the line he is in, we can directly determine his category. Sensor is a type of algorithm. There are too many algorithms and they are not enumerated one by one.
4. Support Vector Machine: An upgraded version of the sensor. Anyone who has studied functional theory knowsA complete inner product space is a Hilbert space.The kernel method can be used in the Hilbert space. SVM uses the maximum interval principle and Kernel Method to Improve the sensor so as to get better results. SVM is a type of algorithm.
5. Feedback Neural Networks: An upgraded version of the sensor. Sensor is a linear function. If multiple linear functions are nested with each other and use nonlinear dynamics to provide Complex Vector Space Surface descriptions, we will get better results than sensor. =! Question: What is the combination of support vector machine and Feedback Neural Network?
6. Neural Networks:Actually, neural networks include Feedback Neural Networks. The reason for putting forward Feedback Neural Networks alone is that they are used too much and inherit the sensor. However, neural networks are very, very, very rich, and very complex algorithms. I will try to classify them into several classes, mainly including:Hierarchical network,Latency Neural Network,Coupled Neural Network,Self-organizing Neural Network,Recursive Neural Network(It is a bit like a latency neural network, but it is somewhat different in terms of continuous and discrete quantities, which can be achieved by continuous analog circuits ),Radial Basis Function Network(This is actually a regularization network. Generally, the RBF network is used to feed back the neural network T-regularization ),Integrated neural network,Fuzzy Neural Network,Boltzmann Machine(A network using the annealing algorithm ),Probabilistic Neural NetworkAnd so on. Of course, there is also a neural field theory that requires the knowledge of differential ry, which is the basic theory of machine learning and can be ignored by beginners. Of course there are still people trying to design neural network computers that beginners can ignore. Of course, neural networks are amazing. neural networks can be used for learning about PCA, ICA, lda (linear discriminant analysis), and LDA (hidden geographic distribution.
7. Statistical Judgment Method: It is the statistical decision theory designed based on the statistical theory. In fact, statistical decision is a very practical theory, and many of these methods are applied in various machine learning fields, such as minimizing the maximum loss, sequential decision, and parameter estimation. Naive Bayes is one of them. This is also a type of algorithm.
8. Bayesian Network: A theory supported by reasoning and planning theories.
9. Sequence Analysis Method: Analyzes the learning of a sequence. A language is a sequence of words, so it is a hidden Markov method.
10. Logistic Regression: If you have studied ecology, you are no stranger to logical equations and logistic regression. In fact, this sensor is actually a matter of urine. He and hidden Markov models can be used to learnMaximum entropy principle. In fact, the maximum entropy principle can be introduced by using the Poisson-Laplace equation in the variational method under the information theory. This is also an exercise after Duda's pattern classification.
11. Clustering Method: We have a bunch of data, and we want to know which of them is one type. It is also a type of method, commonly used: K-means, hierarchical clustering, density distribution clustering, model clustering, graph clustering algorithms (including ant colony clustering ).
12. Data Processing MethodsFor example, PCA, Lda, and ICA.

13. Others:Sorry, although I have written some, I think the basics seem to be more than that, and it seems difficult to sum up. But the main thing is the above. Please add.

(Expansion method) 
1. Online:As we know, the current data is constantly updated. However, because the data is huge, we cannot update each time, so we re-calculate it, so there is a way for the algorithm to increment is calledOnline. The basic methods can all be used to find their online methods.
2. Distributed and parallel: This provides distributed and parallel methods for all of the above basic methods for big data.
3. Corrected overfitting Method: Since most of the above basic methods have the issue of over-fitting, let's just put it bluntly.It is to fit the noise in the data.To make the learning effect worse. The expected information is y = x + 1, and the current information is y = (x ^ 100 + 1) /(x ^ 99 + 1) + 1 .. Obviously, the latter is too precise, but the effect is not good. Most of the above basic methods can be corrected using the corrected overfitting method. WhereRegularizationIs a better method.
4. Various mathematical methods: That's right. You're not mistaken. All kinds of mathematics are messy. For example, if fuzzy mathematics is entered in disorder, a bunch of new methods are generated: fuzzy SVM and fuzzy neural networks. Another example is the comprehensive mathematical chaos, business space and Granular computing. Another example is Li Qun: Li Qun machine learning. For example, what is Manifold Learning when the differential ry is disordered. I think it's a pleasure to watch all of these events.

 (Application field) 

1. Apply it to a graph and change itGraph Mining.
2. Apply it to databases and data warehouses, and changeData Mining.
3. Apply it to social networks and changeNetwork Science.
4. Apply it to natural language processing and changeStatistics Natural Language Processing(Errors are often used as entertainment ).
5. You should be in your field .... A change .... ....
Part 3: Machine Learning application: I will not talk about it much.
 (This book is here !)
In fact, I think it doesn't matter if you have a general understanding of machine learning. According to the above, Baidu posted some blogs, which may be more suitable for you. Or you can find some papers. I would like to recommend a few entry points here.
In fact, I think that when you search for machine learning in Amazon, I should have read all the books on the first page, most of the books on the second page, and the third page, too ..... You can also do this...
Once written in this regard book: http://blog.renren.cOM/blogs/389867835/847240971? Bfrom = 1, 011300082 

 

If you want to recommend:

1. Duda's pattern classification. In fact, I personally preferModern Pattern RecognitionThis book, but because the Chinese people are very emotional, it is true that many books in China are copied seriously, basically the purpose of publishing a book is to publish a book, it is not for people to see. However, there are also many good books in China!
2. Modern Pattern Recognition (version 2)It is really a good book. In my personal opinion, I won't doubt that Duda is perfect.
3. Mitchell's machine learning.
4. The first and second books require a lot of mathematics. If you want to buy a book or borrow a book, please read the description of the mathematical knowledge section in the preface. Otherwise, you will only have a hard time completing the mathematics. In fact, the third book does not need much mathematics, but it involves too little content and involves some simple machine learning theories. This is not as goodAn introduce
To machine learning)It can be said that this book uses the least mathematical knowledge and tells some content, which is also very thin and nice-looking.
Actually, that's enough. Why? Because there are a lot of people asking me which books I want to read (because I think I have read books and papers all day long ).
But they repeatedly asked, it seems that you recommend a few of them, they looked at or didn't look at it, and they asked you to recommend it again...
In my heart, learning which new subject is not smooth sailing. You think you have no foundation before. You can use 200-plus pages of machine learning books for one afternoon to kill them... So loving.

To put it bluntly, books are not important. What matters is your determination. Do you want to learn this subject! If you are holding a half-hanging mentality, you don't need to read it if you are decisive, because the half-hanging determination won't understand any subject! If you want to learn it, stick to it, read it repeatedly, and check it repeatedly...

Original article:Http://www.douban.com/group/topic/36273899/