Objectives

An objective simply consists in a loss function and a decoder. Node-level objectives compute a loss for every node in a time-window graph, whereas edge-level ones compute loss for all edges. This makes node-level objectives usually faster but less powerful than edge-level objectives to capture pair-wise information.

predict_edge_supervised is the only supervised objective: it learns from attack edges that you provide. See Supervised Fine-Tuning.

Arguments

• predict_edge_supervised
  • decoder: str (1)
  • mode: str (2)
  • top_n_attacks: int (3)
  • attack_patterns: list (4)
  • max_edges_per_pattern: int (5)
  • attack_edges_path: str (6)
  • pos_weight: float (7)
• predict_edge_type
  • loss: str (8)
  • decoder: str (9)
  • balanced_loss: bool
  • use_triplet_types: bool
  • AMS
    • version: int
    • margin: float
    • scale: int
• predict_node_type
  • decoder: str (10)
  • balanced_loss: bool
• predict_masked_struct
  • loss: str (11)
  • decoder: str (12)
  • balanced_loss: bool
• detect_edge_few_shot
  • decoder: str (13)
• predict_edge_contrastive
  • decoder: str (14)
• reconstruct_node_features
  • loss: str (15)
  • decoder: str (16)
• reconstruct_node_embeddings
  • loss: str (17)
  • decoder: str (18)
• reconstruct_edge_embeddings
  • loss: str (19)
  • decoder: str (20)
• reconstruct_masked_features
  • loss: str (21)
  • mask_rate: float
  • decoder: str (22)
• one_class
  • decoder: str (23)
  • beta: float (24)
  • eps: float (25)
  • warmup: int (26)

1. Decoder used before computing loss.
   
   Available options (one selection):
   edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
2. 'scores': pick top-N from an edge_scores pkl; 'patterns': match hand-crafted TTP patterns; 'synthetic': use pre-computed embeddings for explicit attack edge definitions.
   
   Available options (one selection):
   scores
patterns
synthetic
3. (scores mode) Number of top-loss edges to use as attack examples.
   
4. (patterns mode) List of TTP pattern dicts.
   
5. (patterns mode) Max edges collected per pattern.
   
6. (synthetic mode) Path to YAML file with explicit (src_type, src_label, edge_type, dst_type, dst_label) attack edge definitions.
   
7. BCE pos_weight for the attack class (on top of 1:1 oversampling).
   
8. Loss used to predict the edge type.
   
   Available options (one selection):
   cross_entropy
BCE
9. Decoder used before computing loss.
   
   Available options (one selection):
   edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
10. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
11. 
    Available options (one selection):
    cross_entropy
BCE
12. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
13. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
14. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
15. 
    Available options (one selection):
    SCE
MSE
MSE_sum
MAE
none
16. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
17. 
    Available options (one selection):
    SCE
MSE
MSE_sum
MAE
none
18. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
19. 
    Available options (one selection):
    SCE
MSE
MSE_sum
MAE
none
20. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
21. 
    Available options (one selection):
    SCE
MSE
MSE_sum
MAE
none
22. Decoder used before computing loss.
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
23. Decoder applied to embeddings before the hypersphere; use 'none' (identity).
    
    Available options (one selection):
    edge_mlp
node_mlp
magic_gat
nodlink
inner_product
none
24. Soft-boundary fraction; radius is the (1-beta) distance quantile.
    
25. Center slack keeping |c| away from zero.
    
26. Kept for parity; a no-op (c/r update every train step).