Parameter tuning method and system of machine learning algorithm

Abstract

The invention discloses a parameter tuning method and system of a machine learning algorithm, which have the advantages of automation, rapidness, high efficiency and universality and belong to the field of machine learning. The method comprises the steps of S1, inputting a user data set and a to-be-adjusted parameter algorithm, and extracting features of the user data set; S2, finding a neural network of a to-be-adjusted parameter algorithm in a neural network database, inputting the features extracted in the step S1 into the neural network, and outputting parameters of the to-be-adjusted parameter algorithm by the neural network; wherein the neural network database comprises neural networks established and trained according to various machine learning algorithms, and the specific construction method comprises the following steps: constructing a knowledge base suitable for a data set of the machine learning algorithms; extracting features of each data set in the knowledge base; findingout optimal parameters for operating each machine learning algorithm on each data set; and establishing a neural network of each machine learning algorithm, and training the corresponding neural network by using the features of the data set and the corresponding optimal parameters as a training set to obtain a trained neural network.

Description

technical field [0001] The present invention belongs to the field of machine learning. Background technique [0002] Existing parameter tuning techniques related to machine learning algorithms mainly include grid search and genetic algorithm. Among them, grid search is to set some candidate values ​​for each parameter. When looking for the optimal combination of multiple parameters, each Combinations of candidate values ​​are traversed, multiple models are trained, and the combination of parameters in the model with the best performance is taken as the result. For example, the invention patent with the publication number CN107844837A "Method and System for Algorithm Parameter Tuning for Machine Learning Algorithms". [0003] The genetic algorithm selects a certain number of candidate solutions, abstracts them into chromosomes, selects some better solutions among the candidate solutions through fitness evaluation, and allows them to freely combine and mutate to generate the n...

AI Technical Summary

Problems solved by technology

[0004] The disadvantage of using grid search for parameter tuning is that the candidate values ​​of the parameters are artificially set and cannot be fully "automated"; the selection of candidate values ​​depends on experience, requires users to have relevant background knowledge, and has not been fully utilized. User data sets; many models need to be trained, which takes a lot of time, especially when the number of parameters is large

[0005] The disadvantage of using genetic algorithm for parameter tuning is that because of the need to consider the solution to specific problems and the abstract relationship between chromosomes (and the process of encoding and decoding), the programming of genetic algorithm is more complicated; the genetic algorithm itself contains many hyperparameters, Such as crossover rate and mutation rate, and the selection of these parameters seriously affects the quality of the solution, and the selection of these parameters is mostly based on experience; the characteristics of the user data set cannot be used in time, so the search speed of the algorithm is relatively slow. A more accurate solution requires more training time; the algorithm is dependent on the selection of the initial population; the potential capacity of the parallel mechanism of the algorithm has not been fully utilized

[0006] Therefore, the shortcomings of the existing machine learning algorithm-related parameter tuning methods are: the user needs to set some candidate values ​​​​in advance, it cannot be "automated", and the user needs to have relevant background knowledge; the "traversal" approach is time-consuming.

Images