A jointly distributed semi-supervised topic model

Latent topic models are applied to analyze the low-dimensional semantic meaning of documents and images, which are widely used in object categorization. However, the unsupervised topic model cannot guarantee that the learned topics have a good relation with class labels, while manually aligning and labeling all training images are expensive and subjective in real applications. Aiming at using a small amount of partial labels to find topics much more suitable for classification, joint distribution from multi-conditional learning is adopted in this paper to generate semi-supervised topic models. Semi-supervised LDA and pLSA models are proposed when the joint distribution is known or partially known. Experimental results on natural scene categorization and head pose classification tasks show that the proposed method remains promising using only partial labels in the training process, which demonstrates the effectiveness of the proposed method.