Modified criterion to select useful unlabeled data for improving semi-supervised support vector machines

Multi-view based unlabeled data selection using feature transformation methods for semiboost learning

“…Illustrative examples are presented first, followed by the results for real-world data. For TBS (and DBS), as was done in [19] and [20], just 10% of the available U (i.e., β = 10) was selected based on the confident level and utilized at the subsequent training steps in Algorithm 1 (and Algorithm 2) (refer to Section 5.3 in [19] for more details on this selection)…”

“…In this section, the SemiBoost [23] algorithm and the related criteria, which are closely related to the present paper, are briefly overviewed in order to make it complete. The detailed description can also be found in the literature [19,20,23].…”

“…In this case, the confidence estimation for x i will depend on the U examples. In order to avoid this, in [19], the criterion is modified by taking a balance among the three terms in (1). This modification can be achieved through balancing the three terms through a reduction in the impact of the third term, especially when X + i ≈ X − i .…”

“…2. SB2: a single-view SemiBoost learning is performed using the modified SemiBoost algorithm [19] with L and U , in which U s is selected from U using the criterion of (2).…”

“…Then, a few examples with higher confidence levels are selected to retrain the ensemble classifier together with L. However, it is not guaranteed that adding the selected data to the training data will lead to a situation in which the classification performance can be improved [35]. Therefore, various approaches have been proposed in the literature for selecting a small amount of useful unlabeled data (U s ) from U : these include the self-training [25,30,40] and co-training [3,12,18] approaches, confidence-based approaches [19,20,23], density/distance-based approaches [8,27,28], and other approaches used in active learning (AL) algorithms [7,11,33].…”