Learning method for weights in weighted classifier model based on given user preferences

Abstract

The invention belongs to the technical field of machine learning and data mining, and discloses a learning method for weights in a weighted classifier model based on given user preferences. The methodcomprises the following steps of randomly generating a plurality of weight vectors as individuals in one population; for each weight vector, calculating predicted values of the weighted classifier model for all sample categories according to an expression, and then calculating a classification error total cost according to a prediction result and a cost matrix,such as the sum of classification error costs of all mistakenly classified samples; selecting a weight vector with relatively low total classification error cost from the weight vector set according to a certain proportion, and performing encoding, crossing and mutation genetic operation on the selected vector to generate a new weight vector set; repeating the second step to the third step on the newly generated weight vector in theprevious step until the condition of iteration stopping is achieved; wherein the weight vector obtained after decoding enables the total cost of the classification error to reach the minimum value.

Description

technical field [0001] The invention belongs to the technical field of machine learning and data mining, and in particular relates to a weight learning method in a weighted classifier model based on a given user preference. Background technique [0002] At present, the existing technologies commonly used in the industry are as follows: in real life, there are often such classification problems, that is, different classification errors will cause different costs, and this type of problem is called cost-sensitive classification problem. For example, in a bank fraud detection system, misclassifying a normal user as a fraudulent user may bring about unnecessary verification procedures or cause the bank to lose a customer; Usually it will bring a large amount of losses to the bank. Cost-sensitive classification problems are often associated with imbalanced training data distributions. Imbalanced data refers to the fact that the number of samples in the data set distributed on d...

AI Technical Summary

Problems solved by technology

The distribution of weights will greatly affect the performance of weighted classifiers

However, usually the weights in the weighted classifier model are unknown, and there is no effective way to find the appropriate weights to minimize the total cost of the classifier's classification error

[0008] To sum up, the existing problems in the existing technology are: usually the weights in the weighted classifier model are unknown, and no effective method has been found to find the appropriate weights to minimize the total cost of classification errors of the classifier

[0010] In the usual method for solving cost-sensitive problems, the weights of large-class samples and small-class samples are given artificially. The researchers simply give the set value of the weights, and do not design the scheme according to the optimization method.

Images