Multi-view learning with dependent views

Multi-view algorithms, such as co-training and co-EM, utilize unlabeled data when the available attributes can be split into independent and compatible subsets. Experiments have shown that multi-view learning is sometimes beneficial for problems for which the independence assumption is not satisfied. In practice, unfortunately, it is not possible to measure the dependency between two attribute sets; hence, there is no criterion which allows to decide whether multi-view learning is applicable. We conduct experiments with various text classification problems and investigate on the effectiveness of the co-trained SVM and the co-EM SVM under various conditions, including violations of the independence assumption. We identify the error correlation coefficient of the initial classifiers as an elaborate indicator of the expected benefit of multi-view learning. Categories and Subject Descriptors I.2.6 [Artificial Intelligence]: Parameter learning General Terms Algorithms, Experimentati...

• Type: Journal
• Year: 2015
• Where: SAC
• Authors: Ulf Brefeld
• Keywords: Applied Computing, SAC 2015