Method and device for training and using relational network embedding model

Abstract

The embodiment of the present description provides a method for training an embedded model of a relational network, which includes obtaining a marked node from the relational network, and each marked node has a corresponding label value to indicate the credit level of the corresponding user. Then, determine the node characteristics of the marked node, the neighbor node set, and the edge characteristics of each connection edge with each neighbor node. Based on node characteristics, edge characteristics and the first parameter set, determine the node embedding vector of the primary iteration of each calibration node, and the edge embedding vector of the primary iteration of each connected edge; then perform multi-level vector iteration to determine the multiple of each calibration node Node embedding vectors for level iterations. Further, based on the node embedding vector of the multi-level iteration of the calibration node, and the prediction parameter set, the predicted value of the node is determined; finally, each parameter value is adjusted to minimize the loss function, wherein the loss function is based on the predicted value of each calibration node and Determined by the tag value.

Description

technical field [0001] One or more embodiments of this specification relate to the fields of relational networks and graph embedding, and in particular, relate to training an embedding model of a relational network, and a method and device for analyzing a relational network using the model. Background technique [0002] Credit risk is one of the basic risks faced by financial enterprises. How to effectively control credit risk is the most concerned issue for financial risk managers. With the acceleration of economic globalization and financial liberalization, as well as the rapid development of Internet finance, the risks faced by the financial market are becoming increasingly complex. Risk control is the foundation of the development of a financial enterprise, and the risk assessment model is naturally the top priority of enterprise research. A good assessment model can reduce the number of bad debts of the company and improve the company's profitability, and at the same ti...

AI Technical Summary

Problems solved by technology

But in fact, the user data obtained by all commercial organizations is incomplete. For example, e-wallet platforms often cannot obtain the data of users' deposits and loans in banks.

Therefore, for an individual user who does not have rich data, he may not be able to enjoy due services due to his low credit evaluation score, or he may enjoy more lending rights because of his high score, which will cause certain risks to the enterprise

[0004] The traditional risk control model cannot fully utilize the value of data. The traditional model is mainly based on feature mining. According to the characteristics of different business mining and risk correlation, it uses models such as decision tree, logistic regression and instance-based deep learning to carry out risk analysis. Evaluate

Since these models only use the characteristics of each sample and cannot integrate other characteristics, they cannot fully utilize the value of data for comprehensive credit evaluation

Images