What are the initial knowledge of machine learning algorithms? Machine learning is undoubtedly now an important part of the field of data analysis, all things engaged in it work in the field of people in peacetime work will be more or less the use of machine learning algorithm. There are many algorithms for machine learning, but there are two types of big aspects: one is learning, the other is the similarity of algorithms.
Learning Style:Depending on the type of data, there are different ways to model a problem. In the field of machine learning or artificial intelligence, people will first consider the algorithm's learning style. In the field of machine learning, there are several main ways of learning. It is a good idea to classify the algorithm according to the way of learning, which allows people to consider the best possible results by choosing the most suitable algorithm based on the input data when modeling and algorithm selection. The main learning methods and learning models of the algorithm are as follows: Supervised learning: input data is called training data, they have known labels or results, such as spam/non-spam or stock prices for a certain period of time. The model's parameter determination needs to pass a training process, in which the model will ask for predictions, and when the predictions do not match, the changes need to be made. Unsupervised learning: The input data is not labeled or has a known result. Build the model by speculating on the structure that exists in the input data. Examples of such problems relate to the learning of union rules and clustering. Examples of algorithms include the Apriori algorithm and the K-means algorithm. Semi-supervised learning: the input data consists of tagged and unmarked components. While the right predictive model already exists, the model must be able to predict and organize the data by discovering the underlying structure. Such issues include classification and regression. Typical algorithms include the generalization of some other flexible models that make assumptions about how to model unlabeled data. Intensive learning: input data is provided to the model as an incentive from the environment, and the model must respond. Feedback does not come from the training process as supervised learning, but rather as an environmental penalty or reward. Typical problems are system and robot control. Examples of algorithms include Q-Learning and sequential differential learning (temporal difference learning).
Algorithmic SimilarityAccording to the function and form similarity of the algorithm, we can classify the algorithm, for example, tree-based algorithm, neural network based algorithm and so on. Of course, the scope of machine learning is very large, and some algorithms are difficult to classify into a certain category. For some classifications, the same classification algorithm can be used for different types of problems. Here, we try to classify commonly used algorithms in the easiest way to understand them.
regression analysisRegression is a modeling method that determines the amount of prediction error of a model and then iteratively optimizes the relationship between variables by this amount. Regression method is the main application of statistics, which is classified as statistical machine learning. This is confusing because we can use regression to refer to a class of problems and a class of algorithms. In fact, regression is a process. Here are some examples: normal least squares logistic regression stepwise regression multiple adaptive spline regression (MARS) local polynomial regression fitting (loess)
An instance-based approachAn instance-based learning model models decision-making issues based on examples that are considered important in the training data or that are necessary for the model. This approach typically builds a sample database and then compares the new data to the database based on a similarity metric to find the most matching item and finally make predictions. Thus, the case-based approach is also known as the "winner-take-all" approach and memory-based learning. This approach focuses on the representation of existing instances and the measurement of similarity between instances. K Nearest Neighbor Algorithm (kNN) learning vector quantization (LVQ) self-organizing mapping (SOM)
Regularization MethodThis is an extension of another method (usually a regression analysis method) that punishes a model with a high degree of complexity and tends to promote a better, simpler model. I've listed some of the regularization methods here because they are popular, powerful, and usually just simple improvements to other methods. Ridge regression Lasso Algorithm (LASSO) elastic Network
Decision Tree LearningThe decision tree method models the decision-making process, which is based on the actual values of the attributes in the data. Decisions are forked on a tree structure until a particular record can be predicted. In the problem of classification or regression, we use data to train decision trees. Classification and regression tree algorithm (CART) iterative binary Tree 3 generation (ID3) C4.5 algorithm Chi-square automatic Interactive view (CHAID) Single-layer decision tree random forest multivariate adaptive spline regression (MARS) Gradient Propulsion Machine (GBM)
Bayesian algorithmBayesian methods are those that explicitly apply Bayesian theorems to classification and regression problems. Naive Bayesian algorithm aode algorithm Bayesian reliability Network (BBN)
Kernel function MethodThe most famous method of kernel function is the popular support vector machine algorithm, which is actually a series of methods. The kernel function method is concerned with how to map the input data to a high dimensional vector space, in which some classification or regression problems can be easily solved. Support Vector Machine (SVM) radial basis function (RBF) linear discriminant analysis (LDA)
Clustering MethodJust like regression, clustering represents both a class of problems and a class of methods. Clustering methods are generally divided according to the modeling method: Centroid-based or hierarchical structure. All methods use the intrinsic structure of the data to classify the data into the most common category. K-Mean maximum expectation algorithm (EM)
Association Rule LearningAssociation rule learning is a class of algorithms for extracting rules that best explain the relationships between variables in the observed data. These rules can find important and commercially useful associations in a large cube and are then further exploited. Apriori Algorithm Eclat algorithm
Artificial Neural networkArtificial neural networks are algorithms that are inspired by the structure and/or function of biological neural networks. They are commonly used in regression and classification problem of pattern matching method, but in fact this huge subclass contains hundreds of algorithms and algorithms of deformation, can solve various types of problems. Some of the classic popular methods include (I've separated deep learning from this class): Perceptron reverse propagation algorithm Hopfield Neural Network Adaptive Mapping (SOM) Learning vector quantization (LVQ)
Deep LearningThe deep learning approach is a modern and improved version of artificial neural networks using inexpensive and redundant computational resources. Such methods attempt to build much larger and more complex neural networks, as mentioned earlier, many of which are based on very limited tagging data in large data sets to solve semi-supervised learning problems. Restricted Boltzmann machine (RBM) Depth Belief network (DBN) convolutional neural network Cascade Automatic Encoder (SAE)
Dimensionality Reduction MethodAs with clustering methods, the dimensionality reduction approach attempts to summarize or describe the data using the intrinsic structure of the data, and the difference is that it uses less information in an unsupervised manner. This is useful for visualizing high-dimensional data or simplifying data for subsequent supervised learning. Principal component Analysis (PCA) Partial least squares regression (PLS) Salmon Mapping Multidimensional Scale analysis (MDS) projection pursuit
Integration MethodThe integration method is composed of several weaker models, which are trained independently, and their predictions are integrated in some way to produce a general forecast. Much effort is focused on choosing what type of learning model to use as a sub-model and how to integrate their results. This is a very powerful technology, and therefore very popular. Propulsion technology (boosting) self-exhibition Integration (Bagging) Adaptive Propulsion (ADABOOST) Cascade generalization Strategy (Blending) gradient Propulsion (GBM) random forest
Ready to learn Directory
