基于滑动窗口均值先验的非同构动态贝叶斯网络转换点检测算法

Sliding Window Prior Knowledge-Based Algorithm for Changepoint Detection in Non-homogeneous Dynamic Bayesian Networks

为了放宽动态贝叶斯网络中的同构假设,提出非同构贝叶斯网络.基于此种情况,文中提出结合先验知识的可逆跳转的马尔可夫链蒙特卡洛算法(APK-RJ-MCMC).算法基本假设为如果一个时间点左右窗口中数据均值间的欧氏距离越大,则这个时间点作为转换点的可能性越高.基于上述假设,可得到关于每个时间点作为转换点可能性的粗略估计,将其作为先验知识调控可逆跳转的马尔可夫蒙特卡洛采样技术(RJ-MCMC)采样转换点时的生成、消除、转换动作的提议概率之比,进而调节状态跳转时的接受概率.在人工数据集和基因数据集上的实验表明,相比其它算法,APK-RJ-MCMC在转换点检测上具有更高的检测后验概率.

To relax the homogeneity assumption of dynamic bayesian networks （DBNs）, the non-homogeneous DBNs is proposed. In this paper, an improved reversible-jump Markov chain Monte Carlo （RJ-MCMC） algorithm is put forward by integrating the prior knowledge about the sliding window, namely APK-RJ-MCMC. The basic assumption of APK-RJ-MCMC is that the bigger the distribution distance between the backward window and the forward window of a time point is, the higher the probability of the time point as a changepoint becomes. Based on the above assumption, the rough probability of each time point as a changepoint is obtained. And it is considered as prior knowledge to guide birth, death and shiftmoves in adjusted. proposed algorithm RJ-MCMC algorithm during the changepoint sampling. Finally, the accept probability is thus Experimental results on both the synthetic data and the real gene expression data show that the APK-RJ-MCMC has a higher posterior probability and better AUC scores than the traditional does in changepoint detection.