Ant-Suit-West Pavilion: Technical Challenge and solution of intelligent Finance

The following is a wonderful video collation:

At present, more and more applications of intelligent technology, such as micro-credit, insurance, payment, wind control, wealth and so on, the financial services also put forward more challenges, such as: Time sensitive, massive data, business diversity, system risk, strong security, automation and so on. In the fields of image/voice, NLP, machine learning, inference and decision, the application of intensive learning, unsupervised learning, graph inference, migration learning and other techniques are expected to achieve fast processing and real-time confrontation in large-scale data.

Depth Learning + diagram: Systematic risk prediction and monitoring

For the security of the user funds, the user account, equipment and the merchant three end to protect. The traditional wind control technology is based on the rules and strategies to achieve. As the number of cases increases, more and more rules are added, and traditional models are more difficult to meet the current requirements. The Ant-Suit is a tree model for untrusted transactions to further determine whether the account is stolen. At the same time, the use of GBDT+DNN to further improve the theft account model, the current increase of 10% detection rate. Take Alipay, for example, to make more than 10 million transactions faster and more accurate through risk checking every day. This is very beneficial to the system itself, the company cost, the promotion of user security.

Here's another example of an application of the graph Learning model: Junk account identification

Business-related network data, through the Structure2vec depth network technology (Structure2vec can be based on a small number of annotation data to determine whether the user is a good person or a bad person) to the quantitative representation of the graph, and then optimize the target according to business characteristics. When the user registers, uses the user, the equipment association to construct the graph, and determines whether the account is the garbage account. This can be the registration of garbage accounts to prevent and control, reduce the back-end risk base, stabilize the market indicators, greatly improve the overall quality of the account. Compared with the Node2vec and rules, the effect of Structure2vec is obvious.

In some spam messages, some Chinese character machines are not recognized (such as "bank" written "Burgundy" and "must"). In order to try to solve this problem, we can use the stroke information of Chinese characters, break these words into single words, divide them into strokes, use IDs to represent these strokes, generate n-ary stroke information, and then generate Chinese word vectors. This method has a good recognition effect on words, which can handle the malicious information entered by users to some extent and ensure the security of content.

Read the original