Algorithm 1: Algorithm for the parameter optimization of PDTN.

Input: the input features of source and target data: ${\left\{x_i^s\right\}}_{i=1}^{n_s}$, ${\left\{x_j^t\right\}}_{j=1}^{n_t}$;    training labels of source data: ${\left\{l_i\right\}}_{i=1}^{n_s}$; fc layers: $[f{c}_1,f{c}_2,f{c}_3]$;    learning rate: $l_r$ and trade-off parameters $\lambda$, $\gamma$, and $\mu$. Initialize:${\theta }_f$, ${\theta }_c$ randomly. Output: the optimized parameters: ${\widehat{\theta }}_f$, ${\widehat{\theta }}_c$. while the total loss $L_{total}$$<ϵ$ or iter $n<$ maxIter do (1) Generate a mini-batch features of source and target data: ${\left\{x_i^s\right\}}_{i=1}^{n_b},{\left\{x_j^t\right\}}_{j=1}^{n_b};$ (2) Extract the high-level features of source and target data: ${\{f_k^s,f_k^t\}}_{k=1}^{n_l}=G_f([x^s,x^t];{\theta }_f);$ (3) Calculate the negative-valence and positive-valence feature centers $v_n^b$ and $v_p^b$  in each mini-batch by the Equations (4) and (5); (4) Calculate the feature center of qth class $c_q^b$ in each mini-batch by the Equation (7); (5) if iter $n=1$:   Initialize global centers $v_n$, $v_p$, and $c_q$ (or $c_p$) in whole source data using steps (4) and (5);  else: $$\begin{array}{c}\hfill ▿v_n=\frac{1}{1+n_b^{\prime }}\sum _{\begin{array}{c}1\le i\le n_b^{\prime },\\ l_i\in \mathcal{N}\end{array}}(v_n^b-f_k^{s,i}),v_n\leftarrow v_n-\eta ▿v_n,\\ \hfill ▿v_p=\frac{1}{1+n_b^{″}}\sum _{\begin{array}{c}1\le j\le n_b^{″},\\ l_j\in \mathcal{P}\end{array}}(v_p^b-f_k^{s,i}),v_p\leftarrow v_p-\eta ▿v_p,\\ \hfill c_q=\frac{1}{1+n_s^q}\sum _{1\le i\le n_b^q}(c_q^b-f_k^{s,i}),c_q\leftarrow c_q-\eta ▿c_q;\end{array}$$ (6) Calculate $L_v$, $L_c$, $L_a$, $L_{ce}$, and $L_{total}$ using Equations (2), (6), and (10)–(12), respectively; (7) Update the parameter ${\theta }_f$ and ${\theta }_c$: $$\begin{array}{c}\hfill {\theta }_c\leftarrow {\theta }_c-\mu \frac{L_{ce}}{{\theta }_c},{\theta }_f\leftarrow {\theta }_f-\mu \frac{L_{total}}{{\theta }_f};\end{array}$$ (8) $n=n+1$. end while