Abstract:

The capacity assessment model of community resilience governance is conducive to establishing the long-term mechanism for grassroots prevention and control under major epidemics. However， there are some problems with community governance capacity assessment models. Firstly， the numbers of samples collected are usually hundreds or thousands， which belongs to small samples in machine learning or deep learning， easily leading to underfitting and overfitting of model training. Secondly， the samples are divided into simple and difficult samples， and the traditional assessment models do not have enough ability to recognize difficult samples. Then， feature extraction methods rely on manual experience， which are difficult to learn how experts select features for assessment. Finally， the parameters of the deep learning models rely on expert experience， which are challenging to select the optimal parameters scientifically. The above problems affect the performance of model in evaluating the capacity of community resilience governance.Therefore， the model in the following ways is improved. Firstly， the model is augmented with data using density peak clustering to improve adaptive oversampling （DPAS） to treat the target class as a minority class sample and the remaining other classes as the majority class， for augmenting the sample size of the data. Secondly， Googlenet using transfer learning （TL） is improved to increase the effectiveness of feature extraction under small samples conditions. In addition， multi-class focal loss（MFL） is used instead of multi-class cross-entropy loss（MCL）is used to enhance the focus on difficult samples in model training， for improving the identification of difficult samples. Meanwhile， the hyperparameters of the optimized model using multi-objective slime mould algorithm （MOSMA） are guaranteed to be optimized under multi-objective conditions. Finally， the model performance is validated using the collected community sample dataset， comparing with different baseline models， multi-objective optimization methods and data enhancement models. Experiments demonstrate that the proposed model outperforms the other models. The structure and parameter settings are justified using ablation experiments and sensitivity analysis.A novel idea about community resilience governance capacity is provided， which can be directly applied to community resilience governance capacity under major epidemics. In addition， the results of this study have practical implications that can be extended to other areas of resilient governance assessment.

Key words: transfer learning （TL）, multi-objective slime mould algorithm （MOSMA）, data augmentation, community resilience governance, major epidemic