Ensemble Method And Application For Multi-target Regression Via Label-specific Features

In recent years,multi-target regression learning has received extensive attention in many fields such as ecology and economics,and it has increasingly important potential application value in the real world.Multi-target regression requires multiple targets to be predicted simultaneously for a single instance,so its main challenge comes from modeling the underlying relationships between input and output target variables and exploring the correlation between targets.Although research on multi-target regression has developed rapidly,most multi-target regression methods are based on the same input space to predict all output targets.However,each target may have its own independent feature space,and the best input features for each target may not be the same.On the other hand,most multi-target regression methods often use a single method to learn and predict all targets when dealing with complex input-output relationships.However,in fact,the relationship between the input space and the target variable may be complex.Aiming at the above problems,this thesis proposes a multi-target regression method via sparse integration and label-specific features(SI-LSF),which uses the correlation between targets to improve the overall prediction accuracy of the algorithm by constructing label-specific features for each target.Moreover,in order to handle the complex relationship between input and output,a sparse aggregation is used with various regression methods.The main research work of this thesis is as follows:1.In order to explore the correlation between targets and improve the prediction accuracy,in this thesis,new features called label-specific feature(LSF)that associated with each target are extended on the original input space in the boosting way on a single-target stacking framework.The effectiveness of label-specific feature is demonstrated by comparative experiments.2.In order to model the complex input and output relationships,based on the label-specific feature,this thesis proposes an ensemble method based on the LSF by using a sparse aggregate function to select different types of regression methods for integrated prediction of targets.The effectiveness and flexibility of sparse integration is demonstrated by a comparison experiment with 18 data sets.3.In order to verify the applicability of the proposed algorithm to practical applications,this thesis selects the actual scenario of supply chain demand forecasting to model the forward and long-term merchandise sales,so as to provide accurate and timely supply chain demand data foundation.In order to verify the validity of the model,this thesis conducts comparative experiments on 18 data sets with four classical multi-target regression methods,which fully proves the accuracy and effectiveness of the algorithm.At the same time,using the model to the supply chain demand forecasting scenario for Saudi Arabia's market,the applicability of the algorithm to practical applications is proved.