Deep Semi-supervised Learning for Time Series Classification

SelfMatch: Robust semisupervised time‐series classification with self‐distillation

“…Existing SOTA : For a specific data set, it represented the best‐performance algorithm, including Wei and Keogh, 10 DTW‐D, 11 SUCCESS, 12 DSSL, 24 and MDL‐SSL 15 …”

“… ConvNet : The feature extractor for the self‐supervised multitask algorithm 18 . The detailed structure of ConvNet can be found in Wang et al 7 FCN : The backbone architecture for the deep semisupervised algorithm 24 ResNet : Consisting of three residual blocks and an average pooling, shown in Figure 1. LSTMaN : Consisting of two LSTM‐based attention layers in Section 3.3. ResNet–LSTMaN : The combination of ResNet and LSTMaN (see Figure 1).…”

“…where α and β are the mean value and standard deviation of f Gau , respectively. Following Goschenhofer et al 24 and Wen et al, 58 we have ∈ α {0, 0.05, 0.10} and ∈ β {0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.20} in this paper.…”

“…48 Most researchers focus on the teacher-student interaction behavior. Offline Distillation is widely applied in TSC, 24 image classification, 48 facial expression recognition, 49 text steganography, 50 and so on.…”

“…23 Inspired by the high-and low-level linking approach in BYOT 23 and the integration of multiple SSL methods in FixMatch, 21 we design a robust semisupervised TSC model with SD, SelfMatch, to capture the representations hidden in labeled and unlabeled data. SelfMatch is different from the deep semisupervised learning algorithm (DSSL) 24 and noisy student training (NST). 25 DSSL transfers the knowledge from image classification to TSC, while NST uses a model's knowledge of the ImageNet data set to guide its knowledge of semisupervised image classification.…”