Multi-Feature Data Trend Prediction Based on Fusion of Machine Learning Algorithms

In this paper, the research on data trend prediction is carried out through a variety of machine learning algorithm models around the construction and analysis of multi-feature models. First, the study constructs a feature disambiguation classification model, constructs feature relationships with the help of neural network algorithms, determine the calculation method of vectors in each layer, then corrects the connection weights and calculate the global error to realize the disambiguation of the features of the algorithmic model. Second, the evaluation model is constructed based on model fusion, using the feature selection method based on penalty term (combining LASSO regression, logistic regression and SVM model) and the feature selection method based on tree model (Random Forest, GBDT model, XGBoost model, LightGBM) to select features, and using the information gain and the rate of information gain to select the information and construct the CART tree; and the fusion classification model using Stacking integrated learning is used to construct the classification model. The fusion classification model of Stacking integrated learning is used to evaluate the feature normalization model, and Xgboost, Catboost, LR and Lightgbm algorithms are chosen as the base classifiers in the first layer, and Lightgbm algorithm is used in the second layer. Finally, the feature normalization model elimination scheme is determined by calculating the planar parameter increment, fusion evaluation model, etc. Comparison and analysis of the AUC of the algorithms such as GATboost, XGBoost, Lightgbm, LR, etc., and selection of algorithms with better optimization effect; combined with under-sampling and over-sampling techniques to deal with the imbalance samples, and after analysis, it is determined to use SMOTE sampling.