Seminar Room 5 @ Institute of Statistical Mathematics
Speaker1
Jaeyong Lee (Seoul National University)
Time
13:30-14:10
Title
Dependent Species Sampling Models for Spatial Density Estimation
Abstract
In this study, we consider Bayesian nonparametric models for estimation of spatially varying density. In particular, we propose dependent species sampling models
(dSSM) which allow intuitive modeling of spatial variation.
The spatial dependency is induced in the weights. The data augmentation scheme is proposed for posterior computation.
The methods are illustrated using a simulated data and climate data.
Speaker2
Hyunjoong Kim (Yonsei University)
Time
14:15-14:55
Title
An ensemble pruning method using performance measure based on degrees of difficulty
Abstract
An ensemble pruning method for classification is discussed. Ensemble pruning is beneficial because it allows faster classification, less memory for storage, and improved accuracy of the ensemble classifier.
First, we introduce a weight adjusted voting method in classification ensemble. The weight of classifier depends on the performance measure that accounts for the degree of difficulty of training instances. Second, we derive an optimal weight to determine the priority in which classifiers are aggregated in a bootstrap aggregation ensemble. The classifier which performs better on hard-to-classify instances get higher priority, thus it has higher chance to remain in the ensemble. In the experiment on 28 real and simulation data set, we show the proposed pruned ensembles statistically perform better than existing ensemble methods. Via bias-variance decomposition, we also confirm that the proposed method reduces the classification bias while maintaining low variance.
