There are five main ways to recommend algorithms:
Content-based recommendation: This is a recommendation based on users' personal interests. According to the historical behavior of a single user, the preference degree of content characteristics is calculated, and then the content that matches the user's preference is recommended.
Collaborative filtering algorithm: This is a group-based recommendation. It is recommended based on users' similarity, * * * content and clustering users into different groups based on demographic characteristics. (Explanation: There are two common collaborative filtering algorithms. One is user-based, that is, calculating the similarity between users. If A and B share the same interests, then A likes movies, and B probably will, too. The other is project-based, that is, calculating the similarity between projects. If movie C is similar to movie D, people who like movie C may also like movie D.)
Extended recommendation: based on user interest points, content categories and other extensions. You like historical information, and I'll give you information about archaeology and treasure hunting.
New hotspot recommendation: timeliness and hotspot recommendation based on global content. (When the product lacks user data and content data at the initial stage, the content distribution efficiency is very low. The effect of using content-based recommendation algorithm is not significant, but using some hot topics can gradually refine the user's portrait and precipitate the content through the user's personal behavior (like, comment, browse and collect) while ensuring a certain amount of traffic, so as to prepare for the later personalized recommendation).
Environmental characteristics: recommendation based on region, time and scene. Zhihu advertises dental clinics and weddings in your city.
Each algorithm has different effects and tastes better when combined, so many companies recommend feed in the form of "algorithm matrix". (We'll talk about this later. )
Advantages:
Content quality audit, community governance (abuse, tearing), recommending goods, reducing labor operating costs.
Continue to recommend the feed you are interested in, enhance the stickiness of users, and further increase the commercialization potential.
Let the user's killing time needs be better met and improve the user experience.
Disadvantages:
1. There is a technical error in the algorithm itself or the people behind it-as long as the algorithm is written by people, there must be a probability of error. For example, the angry smart speakers of German residents in the early morning and the out-of-control Uber self-driving car are all caused by bugs in the program. This method we overcome is actually relatively simple. However, for another algorithm that artificially calculates consumers, sometimes we may be powerless, such as the phenomenon of big data killing. Whether it is true or not, such problems are often difficult to identify, thus increasing the difficulty of supervision; (You can't see the word "money" in the vibrato video, only "Q" can be seen instead)
2. The algorithm ignores the human part-there is still a huge gap between the current artificial intelligence and the real understanding of human feelings and behaviors. Facebook reminds you that the essential reason behind the blessings of the deceased relatives is that AI can't really understand what death means to human beings; Therefore, man-machine combination (platform manual participation, user reporting and other independent measures) is needed, and algorithms cannot be used alone.
3. Algorithm training data bias-At present, the basic logic of artificial intelligence is to construct an appropriate machine learning model, then train the model with a large amount of data, and then use the trained model to predict new data. There is a very important premise here, such as the importance of input data. For example, Tay, a bad Microsoft robot, has a problem because the input data itself is biased. If the real world data itself is biased, then the prediction results will be biased.
First draw a conclusion: the algorithm will not lead to "information cocoon room"
An important premise of the judgment that "social media and algorithm recommendation lead to information cocoon room" is that we will only click on familiar and recognized content, and constantly let the machine deepen its impression on us: it turns out that they only like to watch these!
But in reality, this premise is too simplistic and even wrong.
At the individual level, we have all kinds of reading motivations. Influenced by all kinds of cognitive biases, we may tend to click on certain content, but it is by no means limited to what we agree with.
On the social level: we have social relationships on most apps, and we actively choose the accounts we pay attention to, which have an important impact on the content we can access. A person who has a certain social relationship on the APP is unlikely to fall into a narrow vision.
On the technical level: In the classification of algorithms, it is said that each algorithm has its advantages and disadvantages, so many companies recommend feed in the form of "algorithm matrix". But in the eyes of the general public, algorithm = content-based recommendation algorithm, ignoring that "content-based recommendation algorithm" is only one kind of algorithm, and other types of algorithms will also be used by products.
Enterprise level: No store manager wants customers to pay attention to the same category of goods every time they come to the store. Narrowing users' interest is not a good choice for commercialization.
Game:
If the recommendation is too strong, the attention will be weak. Tik Tok's immersive interaction and content-based algorithm recommendation are sharp tools to kill time. Recommended feed brushes are fun. Can you pay attention to the feed again?
* * * Health:
The algorithm has shortcomings, and attention can make up for it or gain something. Recommended feed ignores people's "sociality". In Zhihu, for example, the concerned content producers convey value to us, so we need a way to understand the output content of those dozens or hundreds of concerned objects. The circle of friends meets the information needs of our snooping, and so does it. (In addition, the process can be inferred from the results, and it will be clear more quickly if we have Bili Bili, Zhihu, Tik Tok and Aote. )