Research On Unsupervised Meta-learning Classification Algorithm

In the face of a large number of data sets and unlabeled data,how to extract features from these data and classify them correctly has always been a topic worthy of people’s research.Deep learning and reinforcement learning are also developing constantly,but there are still shortcomings in generalization ability and learning small sample data sets.Unsupervised meta-learning provides an example,which can well solve the above problems.For the hyperparameter optimization problem,this paper modified the model-independent meta-learning parameter optimization problem under the background of unsupervised meta-learning and realized a hyperparameter optimization method.By calculating the entropy weight of features,this method avoids subjective initialization of hyperparameters and quantitatively analyzes the change rule of features and the relationship between features and parameters,to find the global optimal solution and reduce the space consumed by each step of hyperparameter updating,thus improving the efficiency of finding the global optimal solution.At the same time,given the diversification of similarity measurement methods and the large consumption of similarity measurement matrix operation among prototype network samples,the calculation method of extension distance is proposed,which is compared with the calculation method of Euclid distance and improves the time efficiency of matrix operation.The two algorithms in this paper adopt different optimization methods.For the model-independent meta-learning,the entropy weight method is used to calculate the feature matrix to enhance the parameter updating process,which improves the optimization process of the algorithm,and the maximum improvement is 2.1%.For the meta-learning based on metric space,the extension distance is used to replace the Euclidean-style distance to calculate the similarity,and the point-to-interval measurement is realized.The best group improved by 2.36%.