小数据集下基于DRKDE-ICSO的BN结构学习

A BN Structure Learning Based on DRKDE-ICSO in Small Data Sets

为了解决在小数据集条件下进行数据拓展时产生数据高度相似的问题,提出了基于降维核密度估计的小数据集拓展方法,从而得到较为准确的拓展数据。另外,针对鸡群优化算法求解效率低下和收敛性不足的问题,提出改进的鸡群优化算法进行结构学习:在雄鸡的位置更新公式中引入莱维飞行,使鸡群算法具有更强的跳跃能力;采用指数递减的动态调节惯性权重,以加速局部搜索和提高收敛速度;通过引入最优个体引导策略,增加找到较优位置的概率。实验结果表明,所提算法在小数据集条件下,BIC评分、准确率及汉明距离等指标均优于MCMC算法、BPSO算法、CSO算法、ADLCSO-I算法和SA-ICSO算法。

In order to solve the problem of highly similar data in the condition of small data set expansion,the dimensionality reduced kernel density estimation method is utilized for expanding the small data set,obtaining more accurate expanded data.In addition,in order to solve the problems of low efficiency and weak convergence of CSO,an improved ICSO is proposed to learn the structure:L vy flight is introduced into the position update formula of rooster to make the algorithm jump further;the dynamic adjustment inertia weight with exponential decline is adopted to hasten local search and augmenting convergence speed;by introducing the most advantageous individual guidance approach,the likelihood of discovering the ideal position is increased.The experimental results show that the proposed algorithm is superior to the MCMC algorithm,the BPSO algorithm,the CSO algorithm,the ADLCSO-I algorithm and the SA-ICSO algorithm in terms of BIC score,accuracy and Hamming distance under conditions of small data set.