Vehicle loan application fraud identification method and device
A recognition method and car loan technology, applied in character and pattern recognition, instruments, business, etc., can solve the problem of not having the risk of mining customer fraud, large solution space, increasing model training time, etc., to save audit labor costs and fraud. Disposal cost, improvement in prediction accuracy, and effect of reducing the number of solution spaces
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0035]This example provides a parameter tuning method for the xgboost model. The xgboost model has 11 hyperparameters, the hyperparameter n_estimators is the number of tree models, the hyperparameter max_depth is the maximum depth of the tree model, and the hyperparameter min_child_weight is the smallest sample weight among child nodes And, the hyperparameter gamma is the minimum value of the loss reduction required for leaf nodes to branch, the hyperparameter subsample is the sampling rate of each subtree sample, the hyperparameter colsample_bytree is the sampling rate of each subtree feature, and the hyperparameter reg_alphaL1 is the regularization weight. The parameter reg_lambdaL2 is the regularization weight, the hyperparameter max_delta_step is the maximum moving step, the hyperparameter scale_pos_weight is the positive sample weight, and the hyperparameter learning_rate is the learning rate.
[0036] Traditional parameter tuning methods such as grid search, random search...
Embodiment 2
[0041] The present invention is improved on the basis of Embodiment 1. Since there is mutual influence among the 11 hyperparameters, different parameter tuning sequences will affect the parameter tuning results. Usually, several hyperparameters that have the greatest impact on the model effect are n_estimators, max_depth , min_child_weight, amma, subsample, colsample_bytree, and learning_rate. In the field of auto finance anti-fraud, since fraud samples are usually very few, the problem of sample imbalance is faced. At this time, the values of the two hyperparameters scale_pos_weight and max_delta_step play a significant role, so the present invention proposes to focus on adjusting The above 9 hyperparameters. After trial and error, this embodiment selects an optimal training sequence according to the characteristics of the xgboost algorithm, such as figure 1 shown.
[0042] In this embodiment, i=7, that is, seven adjustments are required to complete the adjustment of 11 h...
Embodiment 3
[0052] The present invention limits the candidate values of hyperparameters on the basis of Embodiment 1 or 2. Each hyperparameter has a corresponding value range. The general value range of each parameter is that n_estimators is an integer and n_estimators∈[200,600 ], Max_depth is an integer and Max_depth∈[2,8], Min_child_weight is an integer and Min_child_weight∈[1,10], Gamma is a real number and Gamma∈[1e-2,1], Subsample is a real number and Subsample∈[0.5,1 ], Colsample_bytree is real and Colsample_bytree ∈ [0.5, 1], Reg_alpha is real and Reg_alpha ∈ [1e-2, 1], Reg_lambda is real and Reg_lambda ∈ [1e-2, 1], Max_delta_step is real and Max_delta_step ∈ [0 , 1], Scale_pos_weight is a real number and Scale_pos_weight ∈ [1, 100], Learning_rate is a real number and Learning_rate ∈ [1e-2, 0.5].
[0053] It can be seen that the value range of each hyperparameter is very large. According to its own characteristics and the characteristics of the xgboost algorithm, this embodiment ...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


