Research On User Credit Ranking Based On Robust Ordinal Regression

With the development of social media sites,user credit grading is widely used in various fields of the Internet as an effective means to improve the accuracy of risk management and achieve differentiated intelligent recommendations.Especially,multi-category user credit evaluation has attracted substantial attention from many mobile application developers and operators.In practical application,the objectives of user credit evaluation are:(1)Anomaly detection and risk early warning;(2)Personalized information and service recommendation for privileged users.The above two goals still remained up-to-date challenges.The traditional methods of user credit evaluation and analysis only give each user a separate credit score by constructing the binary-classification or rank model,this method does not consider the global information of all users,and therefore it is difficult to obtain the potential value of applications.After fully analyzing the basic information and user behavior data,the above problem is transformed into an ordinal regression problem based on user credit grade ranking.In this paper,a robust ordinal nonlinear discriminant analysis model is proposed to get an effective and efficient credit grade evaluation method.The main contents are as follows:(1)In practical application,the sample size of users with different credit ranking is imbalanced distribution,and the traditional method does not consider this problem which leads to overfit or underfit learning.To solve this problem,an algorithm of user credit ranking based on structured non-linear ordinal regression is proposed.Firstly,The model generates an adaptive local weight matrix based on the current samples distribution,which solves overfit and underfit learning caused by the imbalanced distribution of different ordered samples;Then,the penalty constraint of ordered inter-categories is established to optimize the projection direction to avoid the noises of inter-categories and enhance the robustness of the model.This paper regards ROC curves and AUC values as evaluation metrics of imbalanced data classifiers,and transforms the sampling of actual Internet application user information into feature numerical extracting.Experimental results verify the feasibility and effectiveness of the ordinal regression mode,and this method achieves better sorting effect compared with the existing popular methods.(2)Different user categories contain hard-negative samples with rich potential classification information.If such samples are sampled to generate a classifier,the classification effect will be greatly improved.In this paper,an algorithm of user credit grading with triplet loss-based sampling is proposed.Firstly,The model induces the triplet metric constraint to obtain hard-negative samples that well represent the latent ordered class information to solve the problem of poor sorting effect due to fragile performance when KDLOR deals with imbalanced distribution of user data;Then,this algorithm is improved by identifying and evading noises in triplets to obtain hard-negative samples.It solves the problem that the noise samples and the hard-negative samples are similar and difficult to distinguish in the classification boundary,which enhances ordinal regression algorithms robustness and effectiveness.As above,the same actual Internet application user information is used to experiment,and the ROC curves and AUC values are used as evaluation metrics.Theory and experiment results proved that the sorting effect of the proposed model is better than that of the existing popular algorithms.