Research On Internet Consumer Finance Risk Control Based On Improved DeepFM

Consumption is the "ballast stone" of stable economic growth and one of the "troika" driving economic growth,while consumer finance is an important tool to unleash the potential of domestic demand and promote consumer spending.Based on the importance of consumer finance,various financial institutions have started to actively expand their consumer finance business to promote the development of the consumer market and meet the changes in consumer demand.In the context of today’s big data era,the Internet consumer finance business is growing rapidly with its advantages of efficiency,accuracy,real time and low cost,but along with the rapid development,many risks and challenges have emerged.Since Internet-related businesses are processed online with opaqueness,and there are also imperfections in personal credit systems,borrowers’ credit risk is high,so improving the predictive capability of credit risk control models has become the focus of building credit risk control systems.Extracting complex interaction features from user’s behavioral data in order to mine more effective information is the key to improve the prediction effect.User features are generally high-dimensional and sparse,and the combination relationship between features is not obvious.Traditional credit models cannot automatically combine features through algorithms,and need to manually construct interaction features based on experience in the feature engineering stage,but manually constructing complex interaction features not only consumes a lot of labor but also cannot guarantee the validity of the constructed features.The Deep FM model,which was initially applied to click-through rate prediction,can automatically perform feature interactions by model characteristics.However,the model suffers from incomplete extraction of low-order interaction features and can only learn implicit combinations of high-order features.In this paper,we improve the Deep FM model and propose the AMDFM model,which introduces an attention mechanism to learn the weights of different combinations of features when extracting low-order feature interactions,and introduces a multiheaded self-attention mechanism module for high-order feature interactions,so that the model can learn not only implicit high-order feature combinations but also explicit high-order feature combinations,which improves the predictive capability of the model.improves the predictive capability of the model.In this paper,we empirically investigate the proposed AMDFM model based on the Lending Club dataset.The effectiveness of the proposed model is demonstrated in terms of model hyperparameter impact,model performance,model interpretability and feature interaction layer module ablation experiments.In summary,a credit risk control model based on improved Deep FM is constructed in this paper.According to the idea of mining interaction features,by introducing the traditional attention mechanism to improve the low-order feature interaction module and the multi-headed self-attention mechanism to enrich the high-order feature extraction in the feature interaction layer of the model,the learning ability of the model for interaction features is enhanced,so that the model obtains better prediction performance.