Research And Application Of Stacking Algorithm Based On Multiple Meta Models

In recent years,with the vigorous development of the Internet,people have generated massive amounts of data on the Internet,and machine learning has been widely used in many scenarios on the Internet.In theoretical research,Ensemble learning is an important research direction in machine learning,the Stacking algorithm has been widely researched and applied.However,the current research on the Stacking algorithm mainly focuses on the selection of the base model and the meta model and the adjustment of parameters,and rarely studies the structure.This paper studies the Stacking algorithm and improves its structure,and proposes a Stacking Algorithm Based on Multiple Meta Models.Compared with the original Stacking algorithm,the improved algorithm uses multiple meta-models and combines the output of the meta-models through a combination strateg..We apply the Stacking Algorithm Based on Multiple Meta Models to a text-based classification task and house price prediction,the correctness and validity of the algorithm are verified.The overall structure of the thesis is divided into three parts: first is the narrative of the research background and significance,second is the elaboration of the original Stacking algorithm and the improved algorithm,and finally the improved algorithm is applied to the classification and regression problems.Among them,the application of algorithms to classification and regression problems is the core content of this article,both of which include stages such as feature engineering,model training,model evaluation and model prediction.But feature engineering,integration strategies and evaluation indicators are different.The main work of the thesis includes:1)In regression and classification problem,the effectiveness of the Stacking Algorithm Based on Multiple Meta Models was verified on two data sets.In regression experiments,the root mean square errors are 0.145 and 0.5814 respectively;in classification experiments,the accuracy rates are 0.8341 and 0.9391,which are better than other methods.2)It is verified that the performance of the algorithm does not increase with the increase of the number of meta-models,but the individual performance and differences of each meta-model should be considered comprehensively.In the classification problem,the difference between the various models is measured by the ”disagreement measure”,and the two models with lower differences are used as the benchmark to prove that the models have higher differences.3)The training time and test time of each model are tested on each data set,which proves that the improved algorithm obtains better results with less time cost.4)The prediction performance of the algorithm under the condition that the meta model and the base model are independent of each other is tested.It is proved that the Stacking Algorithm Based on Multiple Meta Models does not require the base model and the meta model to be different.As long as the base model and the meta model are different from each other,and the performance of each model is close,you can get a better effect.5)Compared with other mainstream ensemble learning algorithms,it is proved that the improved algorithm is superior to other mainstream ensemble algorithms such as Bagging,Ada Boost and Random Forest.