First, the definition of machine learning
- Arthur Samuel definitions (1959): Machine learning is the study so that the computer can be programmed without displaying computer given discipline learning ability. (Machine Learning: Field of study that gives computers the ability to learn without being explicitly programmed.)
- Tom Michell definition (1998): T to complete the task, the computer learn from experience E, P and to measure performance, and to improve performance by empirical P. E (A computer program is said to learnfrom experience E with respect to some task Tand some performance measure P, if itsperformance on T, as measured by P, improveswith experience E.)
For example, we need to automatically classify messages as spam (Spam) and regular mail (Not Spam), this time the task T refers to the message classification this work experience is history E-mail data, which contains the message content and whether labeled as spam, the performance of P refers to the correct classification number, the higher the more accurate.
Second, the machine learning classifier
Machine learning is divided into supervised learning (Supervised Learning) and unsupervised learning (Unsupervised Learning). In addition, reinforcement learning (Reinforcement Learning) and recommendation system (Recommender Systems).
1. Supervised Learning (Supervised Learning)
In supervised learning, the data corresponding to the correct answer has been given.
For regression problems (Regression) corresponding to the value, and gives data is also given (the correct answer). For example, house prices predicted, given the price and housing area, our task is to predict prices by area, at this time given the training data includes the housing area and the price, which price is the correct answer. Machine learning algorithms need to find a curve (or line) used to fit the data to obtain a model that housing area accepts as input, the output of the model forecast prices. Regression model is to predict the value of the data is continuous. Figure:
In addition to regression problems, as well as classification (Classfication) . For example, we want to predict whether breast cancer, the training set shows the relationship between tumor size and whether cancer, cancer of whether these data called a tag (Tag). Typical features of classification is output to discrete values (Discrete valued output). 0 represents an output of the algorithm cancer, cancer represents an output. Figure:
2. unsupervised learning (Unsupervised Learning)
For supervised learning, the training set contains the data (feature vector) and the correct answer (or tag value). Without supervised learning the main problem is that when the data does not give the correct answer correspondence, we need to grouping (or classification), called clustering (Clustering). As shown below:
Its wide range of applications, such as news aggregators on the network, when the program to search multiple news, many of these news reports are the same thing, this time using a clustering algorithm on the news can be classified as a class, the formation of topics. For example, it reported that Huawei news by clustering algorithm is shown as a class. In addition, its application as well as social network analysis (analyze what people are linked and classified), the market is divided (divided into different types of customers based on customer data).
In addition to clustering, as well as cocktail party algorithm, PCA (Principal Component Analysis, Principal Component Analysis) and other unsupervised learning algorithm.