Multivariate conditional Granger causality analysis based on feature selection and Backward Time-lagged Variable Selection algorithm

With the development of sensor technology, time series data gradually exhibits high-dimensional and multivariate characteristics. Therefore, accurately revealing the potential causal relationships among variables has become crucial for accurate modeling of time series. Due to the limitations of traditional Granger causality models in effectively applying to multivariate time series data, this paper proposes an improved Backward Time-lagged Variable Selection (FS-BTS) optimization method based on Isolation Forest and Maximum Relevance Minimum Redundancy algorithm, and combines it with conditional Granger causality to accurately reveal the implicit topological structure among variables. Specifically, the Isolation Forest algorithm can remove outliers in multivariate time series data, significantly suppressing the influence of outliers on modeling the causal relationships among variables. The introduction of the Maximum Relevance Minimum Redundancy algorithm allows FS-BTSCGCI to adaptively select a subset of features for the target variable, eliminating redundancy among variables. Unlike traditional Backward Time-lagged Variable Selection methods that struggle with eliminating redundancy among variables, this paper improves the Backward Time-lagged Variable Selection method and combines it with conditional Granger causality to enable FS-BTSCGCI to effectively select the optimal explanatory variables for the target variable from multivariate time series data. Compared to other traditional causal analysis models, FS-BTSCGCI has significant advantages in accurately identifying Granger causality relationships among variables. Furthermore, this paper conducts simulation experiments on benchmark and real datasets to validate the effectiveness of the proposed method.