Method and device for risk prediction

Abstract

The invention provides a method for risk prediction, which comprises the steps: carrying out the covering processing of key information in a first sample document set without labels, and carrying out the pre-training based on the covered sample document set, and obtaining one or more pre-training models; constructing a risk prediction model according to the one or more pre-training models, and using a second sample document set with a label for training the risk prediction model, wherein the risk prediction model comprises the one or more pre-training models, a transformer layer and an output layer; and predicting a target document by using the risk prediction model to obtain a risk prediction result corresponding to the target document. According to the scheme of the invention, the frontier unsupervised pre-training technology can be migrated and applied to risk control modeling, and a better risk modeling effect is obtained.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to a technical solution for risk prediction. Background technique [0002] User risk prediction refers to the use of machine learning or deep learning methods to predict the future risks of users, such as the prediction of users' future repayment ability. In the existing technology, some documents issued by authoritative organizations are likely to become the most important basis for risk prediction. For example, in the field of Internet finance, the personal credit report of the central bank is frequently queried in daily business. The credit report reflects a person's background and Credit behavior history and other information, background information includes age, education, gender, occupation history, address history, provident fund payment history, etc. Credit behavior history mainly includes details of credit history, report query history and other information, bec...

AI Technical Summary

Problems solved by technology

[0003] For document-based risk prediction, the current industry mainly adopts artificial statistical methods for feature mining and deep-level sequence modeling using DNN (Deep Neural Networks, deep neural network) models. Artificial statistical feature mining is very Time-consuming, the obtained features are relatively shallow, and the modeling effect is limited, while the ordinary DNN model is prone to overfitting, and the risk prediction ability of the model will decline over time, and one of the difficulties in document understanding is that it may be full of Irregular or even contradictory information is obtained, and the DNN model is sensitive to noise in the data. The risk prediction ability of the final model is not ideal due to the noise in the training data, and the training stability is poor, and the hyperparameters are sensitive

Images