Robust supervised topic models under label noise

Recently, some statistical topic modeling approaches have been widely applied in the field of supervised document classification. However, there are few researches on these under label noise, which exists real-world applications. For example, many large-scale datasets collected from websites or annotated by varying quality human-workers, and then a mislabeled items. In this paper, we propose two robust models for classification problems: Smoothed Labeled LDA (SL-LDA) Adaptive (AL-LDA). SL-LDA is an extension (L-LDA), classical model. The proposed model overcomes shortcoming L-LDA, i.e., overfitting noisy labels, through Dirichlet smoothing. AL-LDA iterative optimization framework based SL-LDA. At each procedure, update prior, incorporates observed concise algorithm maximizing entropy minimizing cross-entropy principles. This method avoids identifying label, common difficulty existing noise cleaning algorithms. Quantitative experimental results completely at random (NCAR) Multiple Noisy Sources (MNS) settings demonstrate our outstanding performance labels. Specially, has significant advantages relative to state-of-the-art massive noise.