UGC Community operations: how to quickly and efficiently handle offending information

Intermediary transaction http://www.aliyun.com/zixun/aggregation/6858.html ">seo diagnose Taobao guest cloud host technology Hall

Jingdezhen is a civilized and harmonious country, and we always have to keep the Internet environment clean. But for the UGC Community products, content is created by users, the platform can only be found in time to deal with. If the filthy information is the network police stare, may not call you any hello, directly unplug the network cable.

So, in Jingdezhen dry products, you have to learn one thing: how to quickly and efficiently handle the violation information.

UGC Community Product Series 1: Handling of offending content

In accordance with the content of the transmission link, strangled in the cradle is the most insurance, that is, after the submission of automatic processing. But there is always a slip through the way or can not be killed, then only in the spread of the solution.

How to identify and automatically process after submission

How to automatically identify and deal with, you need to understand the content contained in the information carrier elements and the behavior of the published content model.

Information carrier, content can be divided into text, pictures, audio, video, which the word processing means the most mature, that is, keywords, URLs and other special text information filtering, pictures, audio, video, malicious URL link address recognition. This process time is very short, the system can only through the word library or two-dimensional model to do simple operations and give a hint of success, can not do too many operations. such as BBS, CMS adopt this filtration method treatment.

How to identify and deal with communication

Behavior patterns, the location of the user (IP identification, geographical identification), access (login behavior mode stay time, whether immediately after the departure), content publishing behavior (release frequency, submission speed how), account data (nickname, mailbox whether there is no rule; Account whether the same batch of registration, etc., social relations (whether with the vest account, whether there is a real social relationship exists, whether to participate in interaction), through these dimensions to establish models for user filtering (whether to disable the account), content filtering. This process requires a powerful system of operations, so after the release of processing. For example, search and ask the question of a few minutes after the release of the issue is cleared, is the use of this method.

The machine can not recognize the content of processing, such as some text content in the keyword metaphor (ru, charade, etc.), split (such as loading 13) and other processing methods, the machine can not be recognized, the use of user reporting function. The system filters the content by collecting user feedback. For example, Sina Weibo most of the content of the reporting function.

But there are special circumstances, the content of the release of the time does not violate the rules, but after the release of the bad effect was sentenced to the content of the violation. This is usually done by the system administrator for immediate manual cleanup. such as Sina Weibo kai-Fu Lee event ban three days.

Dealing with illegal content

On the violation of content, the general use of the foreground to clean up the background data retention practices, serious, clean up the data. Or it can be made visible only by the publisher and not by other users. Or the content of the bottom of the ranking is not rely on the previous display, let the violation of the content directly submerged.

After reading you will find that the process of dealing with the offending content, in fact, and the recommendation engine works very similar. The first layer of filtering requires a quick response, similar to the recommendation engine cold start, the second layer, and the third layer of filtering, it is necessary to use off-line data model and online recessive behavior model (System of user interaction generated data feedback analysis) and explicit behavior model (other users reported feedback data analysis) combined with analysis filtered out.

Although this article is a bit hard to write, but still good understanding. The main share is a way of thinking about how to deal with such problems. Product students in the specific work practice, can be combined with their own products to find the dimension to establish a filtering level model to use.