| Title | : | A Two-Stage Conditional Random Field Model based Framework for Multi-Label Classification |
| Speaker | : | Abhiram Kumar Singh (IITM) |
| Details | : | Mon, 4 Sep, 2017 3:30 PM @ A. M. Turing Hall |
| Abstract: | : | Multi-label classification (MLC) involves assigning the relevant class labels to an example such as an image. One of the main issues in MLC is concerned with capturing the dependencies amongst the labels and exploiting these dependencies to obtain an improved classification performance. In the proposed two-stage framework, the first stage uses an existing approach to perform MLC for an example and gives the confidence scores for different labels as input to the second stage. A conditional random field (CRF) model is used in the second stage to capture the label dependencies and process the confidence scores using these dependencies to improve the MLC performance. The CRF is an undirected probabilistic graphical model. In the proposed framework, the CRF model is used to compute the conditional probability distribution of class labels given the confidence scores for labels. The structure of the CRF that captures the label dependencies is learnt by estimating the weights of edges in the graph such that a regularized negative pseudo-likelihood function associated with the conditional probability distribution is minimized. The results of experimental studies on benchmark datasets demonstrate the effectiveness of the proposed method in improving the MLC performance for the state-of-the-art approaches to MLC used in the first stage of the proposed framework. |
