LightGBM Bidirectional Feature Selection Algorithm for Time Series Prediction

In order to improve the prediction accuracy of high-dimensional complex time series, a LGB-BS Feature selection algorithm based on LightGBM is proposed. The algorithm first uses LightGBM algorithm to build an iterative lifting Tree model for the original feature set, and measures the importance of features. Then, two feature importance metrics are used to combine two-way selection strategy to eliminate irrelevant and redundant features and select the optimal feature subset, Finally, the LSTM neural network is used to predict the original feature set and the optimal feature subset, respectively. The experimental results show that compared with the original feature set, the LGB-BS algorithm can effectively improve the prediction accuracy of high-dimensional complex time series by selecting the optimal feature subset, and improve the problem of poor feature input in time series prediction models.