Machine learning Cornerstone Note 3--When you can use machine learning (3)

3 Types of Learning
3.1 Learning with Different Output Space Y

The method of machine learning is categorized from the angle of the output spatial type.

1. Two-dollar classification (binary classification): The output label is discrete, two-class.

2. Multivariate classification (Multiclass classification): The output label is discrete, multi-class. The dualistic classification is a special case of multivariate classification.

3. Regression (Regression): The output is a continuous value.

4. Structural Learning (structured learning): The output is a structure. The difference between structural learning and classification is that the output of structural learning is not explicitly classified.

Summary:

To do a problem:

Answer:

3.2 Learning with Different Data Label yn

The method of machine learning is categorized from the point of view of data markers.

1. Supervised learning (supervised learning): training assumptions with tagged data.

2. Unsupervised learning (unsupervised learning): training assumptions with unlabeled data.

Common examples: clustering, density estimation, anomaly detection, and so on.

3. Semi-supervised learning (semi-supervised learning): Because of the large number of unmarked data and the cost of tagging, the data part of the training hypothesis (usually a small amount) is marked.

Common examples are: face recognition, efficacy prediction, and so on.

4. Intensive learning (reinforcement learning): The label of the training hypothesis data is "implicit" and usually does not directly indicate what is correct. Popularly speaking, is the input data to the system, if the output of the system does not match the expected output, the "penalty" system, if the output is close to the expected, "reward" system, so as to achieve the purpose of adjusting the system, optimize the learning effect.

Common example: the advertising system (through whether the user clicks on ads to adjust the display of ads: the current user clicks the ads are interested in the current user, then the next time the computer will be more similar to the theme of ads) and so on.

Summary:

To do a problem:

Answer:

3.3 Learning with Different Protocol f⇒ (xn, yn)

The method of machine learning is classified from the perspective of learning strategy.

1. Bulk Learning (Batch learning): sample One-time batch input to the learning algorithm, can be called by the image of the cramming learning, thus obtaining a fixed hypothesis. Is the most common machine learning strategy.

2. elearning (Online learning): Note the difference from Bulk learning: The assumption of online learning is constantly being adjusted according to the sample.

2.1 The current assumptions passively accept a new sample and then readjust the model parameters based on the true value and the predicted value.

The 2.2 is continuously 2.1 until all the samples have been completed. In time, we use samples to correct the model and optimize it.

Example: PLA and intensive learning.

The difference between bulk learning and online learning on message classifications:

3. Active learning (Active learning): attention and online learning are different. A kind of semi-supervised learning. For uncertain instances, the learning algorithm can actively ask the current instance of the tag, get feedback, adjust the system, continue to learn.

The above 3 learning strategies differ:

Summary:

To do a problem:

Answer:

3.4 Learning with Different Input Space X

The method of machine learning is categorized from the angle of the input space.

1. Specific features (concrete Features): Each dimension of the feature has a practical and concrete natural meaning, which is extracted manually and contains human intelligence.

Example:

2. Original feature (Raw Features): Each dimension of a feature has a simple, natural meaning that requires a machine or manual conversion to a more specific meaning.

Example: Recognize handwriting, enter just a simple pixel matrix in the image.

3. Abstract Features: Each dimension of a feature appears to have no natural meaning. Further feature transformations, feature extraction, and feature construction are required.

Example: various scoring systems (film scoring, etc.), given the user and object, get the user's rating of the object. First extract the identity of the UserID, and extract the characteristics of each song itemid, and then use these characteristics to learn.

The difficulty of learning is from large to small: Abstract features > Original features > Specific features.

The original feature and abstract feature all require the re-processing of feature engineering (Feature Engineering). Discrete features generally need only simple selection.

Summary:

To do a problem:

Answer:

Summarize:

Machine learning Cornerstone Note 3--When you can use machine learning (3)