考虑易用性和可解释性的自组织映射-即时学习风险预测框架

A Risk Prediction Framework Based on Self-Organizing Mapping and Just-in-Time-Learning Considering Usability and Interpretability

为提高风险预测系统的易用性和可解释性,提出基于自组织映射网络(SOM)改进的即时学习(JITL)风险预测框架。首先,应用SOM对数据样本进行聚类,并对聚类特征进行解释。进而,通过基于聚类结果的样本选择算法构建待测数据的相似样本集,在线上调用作为基学习器的支持向量机(SVM)进行建模并输出风险预测结果。最后,采用一个交通事故数据集对风险模型的性能进行测试,检验其精度、易用性和可解释性。结果表明:采用SOM-JITL策略的SVM模型,受试者工作状况曲线面积指标达到0.720,相比不使用该策略的传统SVM模型提高17.5%,精度较高;SOM-JITL模型构建所需参数调节工作少,具有较好的易用性;此外,SOM聚类结果准确识别出处于交通拥堵等高风险场景,与现实场景一致,具有可解释性。综上,SOM-JITL策略能有效提高基学习器的性能,达到精度、可解释性和易用性的平衡,有助于以低成本大规模推广风险预测系统。

To enhance the usability and interpretability of risk prediction system,a traffic risk prediction framework based on just-in-time learning(JITL)improved via self-organizing mapping(SOM)is proposed.Firstly,SOM is applied for clustering the data samples and interpreting the clustering features.Then,a sample selection algorithm based on clustering results is used to construct a similar sample set for the data to be tested,and the support vector machine(SVM),which is the base learner,is invoked online to model and output the risk prediction results.Lastly,the model performance is tested using a traffic flow-crash dataset to evaluate interpretability and accuracy.The results show that the area under receiver operating characteristic curve of the SVM model using the SOM-JITL strategy reaches 0.720,which is 17.5%higher than that of the traditional SVM model without the strategy.The SOM-JITL requires less parameter adjustment,and has better usability.In addition,the clustering results of the SOM-JITL accurately identify high-risk scenarios,such as traffic congestion,which is consistent with realistic scenarios and has interpretability.In summary,the SOM-JITL can effectively enhance the performance of the base learner,and endow the model with balance among accuracy,interpretability and usability,facilitating the cost-effective and large-scale deployment of risk prediction systems.