ttnBLSTM-CNN并行模型构建与银行用户流失预测研究

In the spatio-temporal model combining machine learning and cellular automata (CA), it is very important to solve the problem that the simulation accuracy of the key minority land classes is too low due to the imbalance of data distribution. Different data balance strategies and sampling algorithm schemes are deMarkov-MLP-CA Spatio-Temporal Dynamic Modeling And Comparison Analysis Of Equilibrium Strategies -Taking The Three Gorges Reservoir Area As An Examplesigned which is based on the Markov-MLP-CA spatio-temporal dynamic model and taking the Three Gorges Reservoir area as an example. And the Markov-MLP-CA simulation results under different schemes are compared and analyzed. The results show: (1) When the equilibrium degree of the training data set increased from 0.64% to 7.65%, 18.38%, 23.06% and 100% respectively, the KAPPA of the wetland which belongs to the minority land classes increased from 26.19%, 33.69%, 36.57%, 36.86% and 42.05% respectively. And the KAPPA of the shrub land also increased correspondingly. (2) After balancing the training data, the accuracy of the minority land classes has been improved in varying degrees.(3) The model which is coupled with Markov-MLP-CA and SMOTE-Tomek sampling algorithms has the following advantages: the total kappa is 0.8404, the volatility of kappa in different regions is the lowest (49.08%), and the value of Macro-F1 is the highest (0.7219). This study considers: (1) By improving the equilibrium degree of training data and sampling algorithm, the simulation accuracy of the minority land classes can be improved, the volatility of Kappa index can be reduced, and the overall performance of the model can be improved.(2) KAPPA, KAPPAindex volatility and Macro-F1 value should be considered in the model performance evaluation.(3) In comparison, the model coupled with Markov-MLP-CA and SMOTE-Tomek sampling algorithm has better simulation performance.