Algorithm 1 Learning Procedure of the Proposed Framework

Input: Dataset (ꭙ1, ꭙ2, ꭙ3), (ƴ1, ƴ2, ƴ3)

Output: Trained model Ɯ, ∫

1. Initialize projector Ɯ and predictor ∫;

2. repeat

3. Sample ϰ1→Ƭ, ꭚ1→Ƭ from ꭙ1, ƴ1, respectively;

4. ΘƜ←ΘƜ -∂/∂ ΘƜ (ψ);

5. Θ∫←Θ∫ -∂/∂ Θ∫ (ψ);

6. Until ΘƜ, Θ∫ converge

7. return Ɯ, ∫