Ten instances of machine learning problems]

What is machine learning? The answer to this question can be referred to the authoritative definition of machine learning, but in fact, machine learning is defined by the problems it solves. Therefore, the best way to understand machine learning is to observe some instances.

First, let's look at some examples of machine learning problems that we know and understand in real life. Then, we will discuss the classification of standard machine learning problems (naming systems) and learn how to identify a standard case. The significance of this is to understand the types of problems we face, so that we can think about the required data and the algorithms that can be tried.

Ten instances of machine learning problems

Machine learning problems are everywhere. They constitute the core or difficult part of the network or desktop software that is used daily. The suggestion on Twitter "Would you like to try it?" and Apple's Siri speech comprehension system are examples.

The following are ten examples of what machine learning is.

• Spam Detection: identify which emails are spam and which are not based on the emails in the mailbox. This model can be used to classify spam and non-spam. We should be familiar with this example.
• Credit card fraud detection: identify which transactions are operated by the user and which are not based on the user's credit card transactions within one month. Such a decision model can help the program refund those fraudulent transactions.
• Digit Recognition: identifies the numbers represented by each Handwritten Character Based on the handwritten zip code on the envelope. This model can help the program read and understand the handwritten zip code, and classify letters by location.
• Speech recognition: determine the specific requirements of a user from the user's discourse. This model can help the program to automatically fill in user requirements. The iPhone with the Siri system has this function.
• Face recognition: based on a large number of digital photos in the album, you can identify the photos that contain a person. Such a decision model can help programs manage photos Based on faces. Some cameras or software, such as iPhoto, have this function.
• Product recommendation: based on a user's shopping record and lengthy Favorites list, identify the products that the user is really interested in and willing to purchase. Such a decision model can help the program provide suggestions to customers and encourage product consumption. Log on to Facebook or googleplus and they will recommend users that may be associated with you.
• Medical Analysis: predict what the patient might suffer Based on the patient's symptoms and an anonymous patient data database. Such a decision model can be used by a program to provide support for professional medical practitioners.
• Stock trading: based on the existing and previous price fluctuations of a stock, determine whether the stock should be created, held, or reduced. Such a decision model can help programs support financial analysis.
• Customer Segmentation: based on the user's behavior pattern during the trial period and past behavior of all users, identify which users will be converted into the payment users of the product and which will not. Such a decision model can help the program to intervene in the user, to persuade the user to make payment earlier or better participate in the product trial.
• Shape Identification: You can use hand-drawn images on the touch screen and a known shape database to determine the shape you want to depict. Such a decision model can help the program to display the ideal version of the shape to draw clear images. The iPhone app instaviz can do this.

These ten instances demonstrate what a good idea is about machine learning. There is a special collection to record those historical examples. One example is a decision that requires modeling. It brings benefits to an industry or field for the effective automatic modeling of the decision.

Some problems are the most difficult problems in artificial intelligence, such as natural language processing and machine vision. Others are also difficult, but they are also a classic machine learning problem, such as spam detection and credit card fraud detection.

Think about your interaction with online or offline software over the past week. You can easily predict 10 or 20 machine learning instances that are directly or indirectly used.

Types of machine learning problems

Through the examples of the above machine learning problems, you must be aware of some similarities. This kind of skill is very valuable, because it is good at looking at the nature of phenomena, so that you can efficiently think about the data you need and the types of algorithms you can try.

There are some common machine learning classifications. The following categories are typical examples of the problems we encounter when studying machine learning.

• Classification: Mark data, that is, classify it into a certain category, such as spam/non-spam (email) or fraud/non-fraud (credit card transaction ). Decision modeling aims to mark New unlabeled data items. This can be seen as identifying issues and modeling differences or similarity between groups.
• Regression: The data is labeled as a real value (such as a floating point number) rather than a label. Simple examples include time series data, such as stock prices that fluctuate over time. The decision for this modeling is the new unpredicted data estimate.
• ? Clustering: data is not labeled, but data can be grouped based on similarity and other measures of the natural structure of data. You can refer to the list of the above 10 examples: Managing photos Based on faces rather than names. In this way, the user has to name the group, such as iPhoto on Mac.
• Rule Extraction: data is used as the basis for extracting proposal rules (premise/result, also known as if. These rules may, but not all point to each other. This means that these methods can be used to identify statistically persuasive relationships between data attributes, however, it is not necessary to make predictions. There is an example of how to find out the relationship between beer and diapers (this is a private regulation of data mining, and both expectations and opportunities are true or false ).

When you think that a problem is a machine learning problem (for example, a decision-making problem that requires modeling from data), you can then think about what types of problems can be borrowed directly, or, what kind of results do users or needs expect.

Resources

Few resources are available to list machine learning problems in the real world. Maybe they are there, but I didn't find them. I found some cool resources for your reference:

The annual "humies" Award: it is an award that grants computing results comparable to human algorithms. These algorithms only work on data or payment functions, so they can be so creative that they violate patents. It's amazing!

Artificial Intelligence effect: There is a concept: as long as the artificial intelligence program has achieved good results, it will not be regarded as artificial intelligence, but only as technology, and then be used daily. This concept also applies to machine learning.
AI competition: This competition involves very difficult issues in the AI field. If these problems can be solved, it will be a powerful example of Artificial Intelligence (the kind of real artificial intelligence that science fiction imagined ). Computer Vision and natural language processing are examples of Artificial Intelligence competitions. They are also considered as classification of specific areas of machine learning.

Ten questions about machine learning in 2013: there are some wonderful answers to this Quora question, one of which lists the rough categories of actual machine learning questions.

We have discussed some common examples and types of machine learning problems in the real world. Now, we have information about whether a problem is a machine learning problem. We can select some elements from the Problem description to determine whether it belongs to the classification type, returns to the ray series, or belongs to the rule extraction type.