Selective up-sampling combined method for weighted ensemble classification prediction of unbalanced data flows

Abstract

The invention relates to the technical field of data mining, and discloses a selective up-sampling combined method for weighted ensemble classification prediction of unbalanced data flows. The method comprises the following steps of: screening minority class samples of history data blocks according to a similarity, and selecting samples closest to the current training data block in the aspect of concept; synthesizing the selected samples into new samples in a decision boundary area so as to selectively implement up-sampling; and carrying out weighted ensemble classification on the new sample by adoption of a probability distribution relevancy-based weight distribution strategy. According to the method, the minority class sample information is effectively increased through selecting history data with high similarities and synthesizing new data at the boundary area, so that the decision domain of the minority class is enlarged; and meanwhile, in order to adapt the dynamic data with concept drift and use an ensemble classification thought, the probability distribution relevancy-based weight distribution strategy is designed, so that the overall classification precision is enhanced. Experiment results show that the method is capable of effectively improving the minority class identification rate and the overall classification performance, and has the advantage of better processing the unbalanced data flows.

Description

technical field [0001] The invention relates to the technical field of data mining, in particular to an unbalanced data flow weighted integrated classification prediction method combined with selective upsampling. Background technique [0002] With the rapid development of communication technology, a large amount of dynamic data has appeared in many application fields, such as business transaction analysis, intrusion detection and industrial control. Since the data in the data stream presents a large amount and comes continuously, it cannot be loaded into the memory at one time, which makes it difficult to apply the traditional mining method effectively. The classification of data streams has always been one of the important research topics in the field of data stream mining. There are two difficulties in learning and classifying data streams: First, due to the continuous change of data over time, it will inevitably lead to the randomness of the concepts contained in the dat...

AI Technical Summary

Problems solved by technology

Due to the large amount of data in the data stream and the continuous arrival, it cannot be loaded into the memory at one time, making it difficult to apply traditional mining methods effectively.

The classification of data streams has always been one of the important research topics in the field of data stream mining. There are two difficulties in learning and classifying data streams: First, due to the continuous change of data over time, it will inevitably lead to the randomness of the concepts contained in the data. What is changing with time, that is, concept drift, how to effectively adapt to the drift change has become a hot and difficult point in data mining in recent years. Second, dynamic data is often accompanied by unbalanced category distribution. For example, fraudulent transactions in commercial transaction data may only account for 1% or even less. Since the information of the majority class occupies a dominant position, traditional classifiers tend to lean towards the majority class during training and prediction, so they cannot effectively deal with Minority class samples are used for learning, and in many fields, the recognition rate of minority classes is often more important, so the performance of traditional classifiers is often not satisfactory

However, since Naive Bayes is based on the assumption that all features are independent and there is no interdependence, its posterior probability cannot accurately measure the degree of correlation between historical samples and current concepts in data with complex distributions.

In addition, these two algorithms only use historical data for minority class upsampling, and do not generate new sample information, resulting in limited sampling effect

Images