Algorithm 2. The proposed transfer weight optimization and target positioning

Input: (1). Source domain $D_S$; (2) Target domain $D_T$; (3) Number of instances $n_T,n_S$;

Output: (1). Domain mapping of $X_S$, $X_T$; (2). Weighting transfer $W_{mi}$; (3). Target labels $y_T$;

1. Initialize W

2. for $l=1,2,\dots ,L$ do

3.   for $m=1,2,\dots ,n_T$ do

4.      for $i=1,2,\dots ,n_{of}$ do

5.         Compute co-occurrence $C\left(m,i\right)$ using Equation (2)

6.      end for

7.     Obtain the co-occurrence vector of the mth target sample

8.    end for

9. Obtain the common and specific feature matrices of different domains ($J_S{}^{(sp)}$, $J_T{}^{(sp)}$, $J_C$)

10. end for

11. for $l=1,2,\dots ,L$ do

12.   for $i=1,2,\dots ,n_{of}$ do

13.      for $m=1,2,\dots ,n_T$ do

14.         Compute $W_{mi}$ by using Equations (3)–(7)

15.      end for

16.   end for

17. end for

18. Train a classifier from $X_{S\_h}$ and $y_S$ by adjusting weights of source samples $W_{mi}$

19. Estimate $y_T$ on $X_{T\_h}$ by applying the trained classifier $f\left(\left(X_{s\_h},y_S\right),W_{mi}\right)$

20. return $X_S$, $X_T$, $y_T$, $W_{mi}$