What is machine learning?
What is machine learning? The answer to this question can refer to the authoritative definition of machine learning, but in reality, machine learning is defined by the problem it solves. Therefore, the best way to understand machine learning is to observe some examples.
First look at some real-life examples of well-known and understood machine learning problems, then discuss the classification (naming system) of standard machine learning problems, and learn how to identify which standard case a problem falls into. The point of this is that, knowing the type of problem we are facing, we can think about the data we need and the algorithms we can try.
Ten Examples of
Machine Learning Problems Machine learning problems are all over the place, and they form a core or difficult part of everyday web or desktop software. Examples include the "Want to try it out" suggestion on Twitter and Apple's Siri voice understanding system.
Here are ten real examples of what machine learning really is:
- Spam detection : Identify which is spam and which is not based on the messages in the mailbox. Such a model can be programmed to help categorize spam and non-spam. We should all be familiar with this example.
- Credit card fraud detection : Based on the user's credit card transactions within a month, identify which transactions are performed by the user and which are not. Such decision-making models can help programs refund those fraudulent transactions.
- Number recognition : According to the handwritten zip code on the envelope, identify the number represented by each handwritten character. Such a model could help programs read and understand handwritten zip codes and sort letters by geographic location.
- Speech Recognition : From a user's utterance, determine the specific request made by the user. Such a model can help a program to be able to and try to automatically populate user requirements. The iPhone with Siri system has this function.
- Face Recognition : Based on the numerous digital photos in an album, identify those photos that contain a certain person. Such decision-making models can help programs manage photos based on faces. Certain cameras or software, such as iPhoto, have this capability.
- Product recommendation : According to a user's shopping record and a long list of favorites, identify which products the user is really interested in and willing to buy. Such decision-making models can help programs advise customers and encourage product consumption. Log in to Facebook or GooglePlus, and they'll recommend users who might be connected to you.
- Medical Analysis : Predict what the patient might be sick with based on the patient's symptoms and an anonymous database of patient profiles. Such decision-making models can be programmed to support medical professionals.
- Stock trading : According to the current and past price fluctuations of a stock, it is judged whether to open a position, hold a position or reduce a position for this stock. Such decision models can help programs provide support for financial analysis.
- Customer segmentation : Based on user behavior patterns during the trial period and the past behavior of all users, identify which users will become paying users of the product and which will not. Such a decision-making model can help the program to intervene to persuade users to pay earlier or better participate in product trials.
- Shape identification : According to the user's hand drawing on the touch screen and a known shape database, determine the shape the user wants to draw. Such a decision model can help the program display an ideal version of the shape to draw a clear picture. The iPhone app Instaviz does just that.
These ten examples show a good idea of what a machine learning problem can look like. There is a dedicated anthology of those historically significant examples. One example is a decision that needs to be modeled, and the efficient automatic modeling of that decision brings benefits to an industry or field.
Some problems are some of the most difficult problems in artificial intelligence, such as natural language processing and machine vision (to deal with problems that people can easily deal with). Others are difficult, but they are also classic machine learning problems, such as spam detection and credit card fraud detection.
Think about your interactions with online or offline software over the past week. You can easily guess at ten or twenty examples of machine learning that are directly or indirectly used.
Types
of Machine Learning Problems With the examples of machine learning problems above, you must already be aware of some similarities. This skill is valuable because being good at seeing the essence of phenomena allows you to think efficiently about the data you need and the types of algorithms you can try.
Regarding machine learning, there are some common classifications. The following categories are typical for most of the problems we encounter when studying machine learning.
- Classification : Labeling data, that is, classifying it into a class, such as spam/not spam (mail) or fraud/not fraud (credit card transactions). Decision modeling is to label new unlabeled data items. This can be thought of as a discrimination problem, modeling differences or similarities between groups.
- Regression : Data are labeled with real values (such as floating point numbers) instead of a label. Simple and easy-to-understand examples are time series data, such as stock prices that fluctuate over time. This modeling decision is to estimate values for new unpredicted data.
- Clustering : Data is not labeled, but can be grouped based on similarity, as well as other measures of natural structure in the data. Take one from the list of ten examples above: Manage photos based on faces, not names. This way, the user has to name the group, like iPhoto on the Mac.
- Rule extraction : Data is used as the basis for extraction of proposed rules (premise/consequence, aka if). These rules, which may but are not always directional, mean that these methods can find statistically convincing relationships between attributes of the data, but not all are necessary to relate to what needs to be predicted. There is an example of finding out the relationship between buying a beer or buying a diaper, (this is the folk ordinance of data mining, true or not, articulating expectations and opportunities).
When you consider a problem to be a machine learning problem (such as a decision problem that needs to be modeled from data), then think about what problem types can be directly borrowed, or what results users or requirements can expect, and vice versa.
Resources
Few resources list real-world machine learning problems. Maybe they are there, but I didn't find it. I still found some cool resources for your reference:
- Annual "Humies" Awards : These are awards given to algorithms that compute results that rival humans. These algorithms can be so creative that they just work on data or paid functions to violate patents. Amazing!
- Artificial intelligence effect : There is a concept that as long as the artificial intelligence program achieves good enough results, it is no longer regarded as artificial intelligence, but only as technology, and then used daily. This concept also applies to machine learning.
- Artificial Intelligence Competition : This competition involves very difficult problems in the field of artificial intelligence, and if these problems can be solved, it will be a powerful case for proving artificial intelligence (the kind imagined in science fiction, real artificial intelligence). Computer vision and natural language processing are both examples of artificial intelligence competition problems, and they are also treated as domain-specific classifications of machine learning problems.
- Top 10 Machine Learning Questions of 2013 : This question on Quora has some really great answers, one of which lists a rough classification of actual machine learning problems.
Above we discussed some common examples and types of real-world machine learning problems. Now, we have the information to talk about whether a problem is a machine learning problem, and be able to pick out some elements from the problem description to tell if it is a classification type, a regression type, or a rule extraction type.
Do you know some real world machine learning problems? Share your thoughts in the comments.