| With the rapid development of Internet technology,information is growing faster and faster.Data mining has become the research hotspot,and unbalanced data classification is one of the important branches.Complex datasets should be pre-processed before data mining,and the discretization of continuous attributes is one of the important methods in data preprocessing.However,there are most discretization methods,which have a default prerequisite that classes should be roughly balanced distributed.If these algorithms are applied directly to the unbalanced data set,the algorithm will focus on the negative class with the majority of samples,but it is easy to ignore the few classes that we are more concerned about.That leads to discretization scheme not good enough,thus the later result will suffer.What’s more,the increasing data size also pose a challenge to the running speed of the algorithm.The use of GPU parallel computing is a new trend to solve these problems.Therefore,the main content of this paper is about how to discretize the unbalanced data effectively and accelerate the running speed of the algorithm.In order to solve the discretization of continuous attributes of unbalanced datasets,we propose a new discretization algorithm named CARU,which based on the relationship between the Classes and Attributes.After analyzing the two-dimensional quanta matrix of classes and attributes,and fully considering the distribution of data,we combine several discretization standards to define a new discretization criterion called CARU,which is the base of CARU algorithm.The algorithm gets a good discretization scheme by choosing the best breakpoints from the candidate set of breakpoints,which is better beneficial for the later classifying learning.Moreover,the influence of combining the discretization method and SMOTE sampling technique in the preprocessing of the unbalanced data was alse analyzed.The CARU was compared with 5 well-known discretization methods on 36 unbalanced datasets.The results obtained were contrasted through non-parametric statistical tests,which show that our proposal outperforms many of the others.And different discretization methods combined with SMOTE sampling technology,CARU algorithm have more promotion compared with other algorithms,since the CARU algorithm minimizes the loss of information by choosing the reasonable breakpoints.Aiming at speeding up the speed of CARU algorithm in dealing with large-scale data,this paper implements parallel CARU algorithm based on GPU computing platform.By analyzing several calculation key steps in CARU algorithm,such as the sorting of attribute values,the calculation of the discretization criter ia,the choice of breakpoints with the largest criterion value,and the discretizatio n between different attributes,the parallel CARU algorithm is designed using the advantages of GPU parallel computation,which can be extended to multiple GPU devices.Finally,the comparison between GPU-based parallel CARU algorithm and CPU-based serial CARU algorithm shows that when the data size is large,the running speed of the former is greatly improved.Under the best circumsta nces,the speedup of dual GPU can be 6.7 times as fast as 24 CPU threads’. |
