[3Xin2] Poster session 1

[3Xin2-48] An attempt to improve accuracy by updating prior distribution in variational EM algolithm of clustering by mixed Normal Inverse Gaussian Mixture Models

〇Riku Kumagai¹, Takashi Takekawa¹ (1.Kogakuin University)

Keywords:Clustering, Bayesian estimation

There is Bayesian inference in model-based clustering. The prior distribution in the variational EM algorithm, a Bayesian inference solution, is not updated until training is completed, but the prior distribution is used for each training. Therefore, setting the prior distribution is important to achieve proper classification with a small amount of computation, but it is difficult to set by hand. Therefore, we attempted a method to determine a new prior distribution using the posterior distribution during training. In this study, we applied the proposed method to clustering using the variational EM algorithm for NIGMM (Normal Inverse Gaussian). We used ARI (Adjusted Rand Index) as the index of the accuracy. As a result, the ARI did not change significantly, but worsened in many cases. However, there were cases in which the algorithm broke down, indicating that the proposed method can be used effectively by incorporating it with conditional branching.