Research And Implementation Of Auto Insurance Claim Prediction System Based On Improved Factorization Machine-based Deep Neural Network

Auto insurance is an important component of non-life insurance,and the profitability of auto insurance pushes forward a immense influence on insurance companies.Therefore,it is an important task to establish an accurate auto insurance claim prediction model.As technologies such as "Internet+" and big data enter the insurance field,traditional Generalized Linear Models show limitations,Machine Learning algorithms have begun to attract researchers' attention and have achieved rapid development.Aiming at the strong correlation between variables in the field of auto insurance,this thesis proposes an end-to-end prediction model and its improved version of auto insurance claim probability and amount,and develops an auto insurance claim prediction system based on the proposed model.The specific research contents of this paper are as follows:This thesis firstly introduces the research background and significance of the subject,the research status at home and abroad,and the research content and chapter arrangement of this thesis.Then the theoretical basis of the model and algorithm related to the subject is introduced.Then,based on the requirement of establishing the interaction between features in the field of vehicle insurance,a vehicle insurance claim prediction model named Deep Featurewise Interaction Factorization Machine(Deep Fw IFM,for short)is proposed,which combines Feature-wise Interaction Factorization Machines and Deep Neural Network.The model transforms high-dimensional sparse features into low latitude dense vectors through the embedding layer,and then inputs them into DNN and Fw IFM components to capture the interaction of high-order and low-order features.In the experiment of claim possibility prediction,AUC and Log Loss are used to compare the effect of Deep Fw IFM with the other six models,and MSE and MAE are used as evaluation indexes in the claim amount experiment.The experimental results show that Deep Fw IFM has better prediction effect.Further,an attention module Spatial and Channel Squeeze & Excitation(scSE)is added between the embedded layer of Deep Fw IFM model and the DNN input layer to construct a new vehicle insurance claim prediction model Squeeze Deep Feature-wise Interaction Factorization Machine(SDeepFwIFM).scSE uses a method called Squeeze-Excitation to calibrate the features,so as to strengthen the important features and suppress the non important features.The experimental results show that SDeepFwIFM achieves better performance compared with other four models.Finally,this thesis combines Python,Flask and My SQL to implement a vehicle insurance claim prediction system.The design and implementation procedure of each module of the system are described in detail,and the running effect of the system is displayed.