Study On KELM Based On Particle Swarm Optimization Strategy And Its Application

Kernel Extreme Learning Machine(KELM)is an improved ELM algorithm based on Extreme Learning Machine(ELM)with the implementation of kernel function.KELM not only inherits the advantages such as good generalization performance and fast learning speed of ELM,but also can get the least square optimization solution owing to the introduction of kernels,thus KELM gains a more stable and superior performance.After the analysis of KELM,it is known that although the introduction of kernels in KELM replaces the random initialization of hidden layer parameters in ELM,besides the inherent regularization parameter selection problem of KELM itself,the existence of kernels also makes some KELM parameters need to be jointly tuned.Therefore,the setting of different parameters will affect the performance of KELM to some extent.Aiming at the selection of sensitive parameters in KELM,this paper focuses on providing a KELM parameter selection method based on particle swarm intelligence optimization strategy,and validating its effectiveness through the application of drug-target interaction prediction and domain adaptation learning.The main work is presented as below:In this paper,the particle swarm optimization strategy is applied to optimize the parameters of KELM.On the basis of traditional particle swarm optimization(PSO),this paper also introduces two algorithms,i.e.quantum particle swarm optimization(QPSO)and particle swarm optimization(PSO)based on QPSO with chaotic sequence searching technology,namely chaotic quantum particle swarm optimization(CQPSO).The experimental results show that CQPSO not only possesses the superior global optimization ability of QPSO algorithm,but also avoids premature convergence of particles in the process of KELM optimization with the aid of chaotic sequence technology.The prediction of drug-target interaction relationship has been the keypoint of new drug development,especially the selection of predictive models.At present,machine learning algorithms such as support vector machine and random forest have been applied for drug-target interaction prediction,and have achieved good prediction results.In this paper,KELMs are applied to predict drug-target interaction.KELM,PSO-KELM,QPSO-KELM and CQPSO-KELM models are selected to predict drug-target interaction on publically available data sets of drug-target interaction pairs.The predicted results and corresponding values of certain evaluation metrics were obtained by employing well-trained prediction models on test sets.The experimental results indicate that the accuracy of the proposed prediction model is improved by 6.87%,recall rate by9.44%,specificity by 5.7%,accuracy by 3.1%,F1 by 5.96% and MCC by 14.24%.Computer vision research is an important field of artificial intelligence.Domain Adaptation is a transfer learning method widely used in computer vision studies.This paper presents a KELM based Domain Adaptation learning algorithm(DAKELM),which is composed of KELM as classifier,weighted balanced feature matching and instance weighting.The performance of DAKELM is validated and analyzed by experiments on common data sets of computer vision.Experiments show that the average classification accuracy of DAKELM is improved by 2.73% compared with WBDA,and the running time of this algorithm is also shortened by about 10 s.When we analyzed the parameters of DAKELM algorithm,it was found that there were some sensitive parameters in this method which could affect the experimental results.Therefore,three optimizational algorithms,PSO,QPSO and CQPSO,were introduced into the proposed DAKELM algorithm to optimize the selection of sensitive parameters.The experimental results show that compared with other similar algorithms,the average classification accuracy of DAKELM strategy improved by 3.45%-4.78% under the effect of particle swarm optimization.