Algorithm 1 THC Algorithm

Require:

Labeled source dataset $X_S$;

Unlabeled target dataset $X_T$;

Epoch t;

Threshold m;

Update rate ${\alpha }_s$,${\alpha }_t$.

Ensure:

Best model M.

1:Extract all source features $\mathit{S}$ and target features $\mathit{T}$; 2:Calculate source-class centroids according to Equation (6), store source-class centroids $\left\{{\mathit{f}}_i\right\}$ and target features $\mathit{T}$ in memory; 3:for$i=0$ to t do 4:    Obtain target features $\mathit{T}$ from memory, calculate Jaccard distance between samples according to Equations (1) and (2); 5:    Use threshold-based hierarchical clustering to cluster samples according to Equation (3), generate pseudo labels based on clustering results; 6:    Calculate target-cluster centroids according to Equation (7), store target-cluster centroids $\left\{{\mathit{c}}_i\right\}$ and single-sample clusters $\left\{{\mathit{v}}_i\right\}$ in memory; 7:    Fine-tune model with contrastive loss and evaluate model performance; 8:    if $mA{P}_i>mA{P}_{best}$ then 9:         $mA{P}_{best}=mA{P}_i$,          update best model M; 10:    end if 11:    Update target features $\mathit{T}$, calculate and update source-class centroids $\left\{{\mathit{f}}_i\right\}$, target-cluster centroids $\left\{{\mathit{c}}_i\right\}$ and single-sample clusters $\left\{{\mathit{v}}_i\right\}$ according to Equations (6)–(9); 12:end for