Abstract

Labeled data is a necessary part of modern computer science, such as machine learning and deep learning. In that context, large amount of labeled training data is required. However, collecting of labeled data as a crucial step is time consuming, error prone and often requires people involvement. On the other hand, imbalanced data is also a challenge for classification approaches. Most approaches simply predict the majority class in all cases.
In this work, we proposed several frameworks about mixture models based predictive distribution. In the case of small training data, predictive distribution is data-driven, which can take advantage of the existing training data at its maximum and don't need many labeled data. The flexibility and adaptability of Dirichlet family distribution as mixture models further improve classification ability of frameworks.
Generalized inverted Dirichlet (GID), inverted Dirichlet (ID) and generalized Dirichlet (GD) are used in this work with predictive distribution to do classification. GID-based predictive distribution has an obvious increase for activity recognition compared with the approach of global variational inference using small training data. ID-based predictive distribution with over-sampling is applied in occupancy estimation. More synthetic data are sampling for small classes. The total accuracy is improved in the end. An occupancy estimation framework is presented based on interactive learning and predictive distribution of GD. This framework can find the most informative unlabeled data and interact with users to get the true label. New labeled data are added in data store to further improve the performance of classification.