Table 1 Summary of notations used in the proposed method.
Symbol | Description |
|---|---|
\(x_i^s, y_i^s\) | Source sample and its one-hot label |
\(x^t\) | Target domain sample input |
\(r_{\theta }(x)\) | Predicted class probabilities from student network |
\(r_{\theta }(x)_j\) | Response (confidence) for class j |
\(\mathscr {U}\) | Relational matrix encoding inter-class similarity |
\(f_i^s, g_i^s\) | Feature representations of source sample \(x_i^s\) |
\(f_j^t, g_j^t\) | Feature representations of target sample \(x_j^t\) |
\(d_{\epsilon }\) | Domain discriminator with parameters \(\epsilon\) |
\(f \oplus g\) | Concatenated features for domain discrimination |
\(\lambda _{da}, \lambda _{adv}\) | Weights for domain alignment and adversarial loss |
\(\varphi , \theta\) | Parameters of teacher and student networks |
X | Concatenated feature matrix from source and target domains |
\(X_S\), \(X_T\) | Source and target domain feature matrices |
\(x_i\) | A single feature vector from X |
\(\mu\) | Mean vector of X |
\(P_{\text {pca}}\) | PCA projection matrix |
\(X_{\text {reduced}}\) | Dimensionality-reduced features after PCA |
\(\tilde{x}_i\) | L2-normalized feature vector |
\(W_{\text {all}}\) | Affinity matrix combining label and feature similarity |
\(\alpha\) | Weighting coefficient for label similarity |
\(\sigma\) | Gaussian kernel bandwidth parameter |
L | Graph Laplacian matrix |
D | Diagonal degree matrix of \(W_{\text {all}}\) |
P | Projection matrix learned via LPP |
\(\tilde{X}^s_{\text {proj}}, \tilde{X}^t_{\text {proj}}\) | Projected and normalized source and target features |
\(\mu _S, \mu _T\) | Source class mean and target cluster mean in projected space |
\(\tau _T(i)\) | Confidence score for target sample i |
\(\theta _{c(i)}\) | Class-specific pseudo-label confidence threshold |
p | Dynamic selection ratio decreasing over iterations |
\(\text {prob}(i)\) | Max predicted class probability for sample i |
\(\text {pseudoLabels}_T(i)\) | Refined pseudo-label for sample i |
\(\textbf{P}_v\) | Predicted probability matrix from view v |
\(\alpha _v\) | Weight for view v in fusion |
\(D_v\) | Distance-based score for view v |
\(\gamma _v\) | Bandwidth parameter for view weight computation |
\(\textbf{P}_{\text {fused}}\) | Fused probability matrix from all views |
\(\hat{y}_T(i)\) | Final predicted label for target sample i |
