qbiocode.apps.quvine.reproducibility.split_generator module#

Task Split Generator for QuVINE

Generates and saves task-specific data splits to ensure all methods use identical splits.

Summary#

Classes:

SplitGenerator

Generates and saves task-specific splits for reproducible benchmarking.

Reference#

class SplitGenerator(output_dir, seed_manager, registry)[source]#

Bases: object

Generates and saves task-specific splits for reproducible benchmarking.

All splits are pre-generated once and saved to disk. Methods then load these pre-generated splits instead of creating their own.

Supports three tasks: 1. Node Classification: train/val/test node indices + labels 2. Link Prediction: train/val/test edge splits + negative samples 3. Node Ranking: target nodes for ranking evaluation

__init__(output_dir, seed_manager, registry)[source]#

Initialize split generator.

Parameters:

  • output_dir (Path) – Root directory for saving splits

  • seed_manager (SeedManager) – Seed manager for reproducible split generation

  • registry (DatasetRegistry) – Dataset registry to update with split paths

generate_all_splits(dataset_name, repetition_id, G, tasks=['node_classification', 'link_prediction', 'node_ranking'], disease_nodes=None, task_config=None, split_base_dir=None)[source]#

Generate all task splits for a dataset.

Parameters:

  • dataset_name (str) – Dataset name

  • repetition_id (int) – Repetition ID

  • G (nx.Graph) – Graph to generate splits for

  • tasks (List[str]) – Tasks to generate splits for

Returns:

Mapping from task name to split file path

Return type:

Dict[str, Path]

generate_node_classification_split(dataset_name, repetition_id, G, train_ratio=0.6, val_ratio=0.2, test_ratio=0.2, min_class_size=5, disease_nodes=None, positive_class_definition='disease_seed_or_target', negative_class_definition='non_disease_nodes', split_strategy='stratified', split_base_dir=None)[source]#

Generate node classification split with synthetic labels.

Labels are generated using community detection (Louvain).

Parameters:

  • dataset_name (str) – Dataset name

  • repetition_id (int) – Repetition ID

  • G (nx.Graph) – Graph

  • train_ratio (float) – Training set ratio

  • val_ratio (float) – Validation set ratio

  • test_ratio (float) – Test set ratio

  • min_class_size (int) – Minimum nodes per class

Returns:

Path to saved split file

Return type:

Path

Generate link prediction split with held-out edges and negative samples.

Parameters:

  • dataset_name (str) – Dataset name

  • repetition_id (int) – Repetition ID

  • G (nx.Graph) – Graph

  • train_ratio (float) – Training edges ratio

  • val_ratio (float) – Validation edges ratio

  • test_ratio (float) – Test edges ratio

  • negative_sampling_ratio (float) – Ratio of negative to positive samples

Returns:

Path to saved split file

Return type:

Path

generate_node_ranking_split(dataset_name, repetition_id, G, n_targets=None, target_selection='high_degree', n_seeds=None, seed_selection='random', disease_nodes=None, excluded_nodes=None, selection_rule=None, split_base_dir=None)[source]#

Generate node ranking split with seed and target nodes.

For PPI networks, this should use disease-specific seeds/targets. For synthetic networks, generates synthetic seeds/targets.

Parameters:

  • dataset_name (str) – Dataset name

  • repetition_id (int) – Repetition ID

  • G (nx.Graph) – Graph

  • n_targets (int, optional) – Number of target nodes (default: 10% of nodes)

  • target_selection (str) – Target selection strategy (‘high_degree’, ‘random’, ‘central’)

  • n_seeds (int, optional) – Number of seed nodes (default: 5% of nodes)

  • seed_selection (str) – Seed selection strategy (‘random’, ‘high_degree’)

Returns:

Path to saved split file

Return type:

Path

load_split(split_path)[source]#

Load a pre-generated split from disk.

Parameters:

split_path (Path) – Path to split file

Returns:

Split data

Return type:

dict