The High Separation Probability Assumption for Semi-Supervised Learning

Semisupervised Fuzzily Weighted Adaptive Boosting for Classification

“…2) Performance comparison against alternative semisupervised classifiers: Next, the classification performances of SSFWAdaBoost-SOFIS, SSFWAdaBoost-ALMMo0 and SSFWAdaBoost-SOFBIS are compared against a number of well-known semi-supervised approaches on the 12 benchmark problems under the same experimental protocol used in the previous example. In this example, the following eight singlemodel semi-supervised classifiers are used for comparison: 1) Local and global consistence (LGC) [8]; 2) Greedy gradient max-cut (GGMC) [11]; 3) Anchor graph regularization with kernel weights (AGRK) [10]; 4) Anchor graph regularization with local anchor embedding weights (AGRL) [10]; 5) Efficient anchor graph regularization (EAGR) [51]; 6) Laplacian support vector machine (LSVM) [9]; 7) Transductive minimax probability machine (TMPM) [52], and; 8) Self-training hierarchical prototype-based classifier (STHP) [13]. In running the experiments, the externally controlled parameters, α and µ of LGC and GGMC are set to be 0.99 and 0.01, respectively, as suggested by [8], [11]; and both LGC and GGMC use the kNN graph with k = 5; AGRK, AGRL and EAGR select a total of 0.1(L + K) anchors from data by k-means (L + K is the total number of data samples); the number of the closest anchors s is set as s = 3 [51]; the iteration number of local anchor embedding is set to be 10 for AGRL [10]; as the performance of LSVM is sensitive to its parameter setting, in this study, three different parameter settings are considered for LSVM, namely, i) σ = 10, µ I = 1, µ A = 10 −6 , k = 15 (as suggested by [9]); ii) σ = 10, µ I = 0.5, µ A = 10 −6 , k = 15; iii) σ = 1, µ I = 1, µ A = 10 −6 , k = 15; and the best classification result on the unlabelled training samples of each dataset is reported; the values of two externally controlled parameters, λ and ρ of TMPM are varied from [10 −4 , 10 −3 , 10 −2 , ..., 10 4 ] as suggested by [52], and 10% of the labelled training samples are randomly selected to help TMPM determine the best parameter setting for each problem as the validation samples; STHP uses the following parameter setting γ o = 1.1, H = 6 and N = 1000 [13].…”

Balanced Contrastive Learning for Long-Tailed Visual Recognition

“…Imbalanced Semi-supervised Learning (SSL). Semisupervised learning is a subfield of machine learning that addresses scenarios where labeled training samples are limited, but an extensive amount of unlabeled data is available [26], [27], [28], [29], [65]. This scenario is directly relevant to a multitude of practical problems where it is relatively expensive to produce labeled data.…”