Spectral Clustering Algorithms For Case-based Reasoning

Case-based reasoning is a mature way that uses past knowledge to solve problems with highly similar features and learns how to solve them.When retrieving cases,for a huge case base,the classic KNN retrieval algorithms need to match all cases when dealing with linear problems.Therefore,there are problems of high time cost and low efficiency.For this reason,most studies now cluster the entire case base to form clusters with different features.Spectral clustering algorithm,as a clustering algorithm based on spectrogram theory,is different from traditional clustering algorithms.Not only does it have no requirements for the spatial distribution features of the sample set,but the clustering result is still globally optimal,especially for the sample set.The spatial distribution feature is non-convex.However,the scale parameters of traditional spectral clustering algorithms are global and cannot be personalized according to the spatial distribution features of samples.Since then,although there have been many improved spectral clustering algorithms,most of the improved algorithms still have the disadvantages of parameter sensitivity,difficulty in selecting the nearest neighbor K value,and susceptibility to outliers.Based on the above analysis,this article mainly does the following work:(1)In view of the defect that the spectral clustering algorithm still uses experience to select the nearest neighbor value,it is vulnerable to outliers.an adaptive spectral clustering algorithm based on angle and modulus between samples(AMD?SC)is proposed.The algorithm uses the reciprocal of the angle between samples as the scale parameter,and uses the modulus between samples as an auxiliary,Take advantage of the feature attribute information of the case,expand the differentiation of element weights between different clusters,and achieve efficient clustering performance.Eventually,the experiment on the UCI public data set proves that the AMD?SC has good performance and high robustness.(2)Aiming at the shortcomings of traditional KNN algorithm in processing massive case bases that the amount of calculation is huge and traditional clustering algorithms cannot effectively cluster non-convex samples,the AMD?SC is used to achieve clustering,and the genetic particle swarm optimization is used to solve the K most similar cases of the target case in the cluster.Eventually,its proves by simulation that for linear problems,its Predictions precision of the improved CBR retrieval algorithm have tiny error and the algorithm performance is excellent.(3)Aiming at the problem of two search case databases in some target cases,the optimal principle retrieval strategy is proposed.Eventually,Its proves by simulation that After the improved CBR retrieval algorithm cooperates with the optimal principle retrieval strategy,its Predictions precision is further improved,and the retrieval strategy is effective.