TwitchDataset(root, name, transform=None, pre_transform=None, pre_filter=None)

Bases: InMemoryDataset

Twitch social network dataset loader for graph-based analysis.

Parameters:
  • root (str) –

    Root directory where the dataset should be saved

  • name (str) –

    Name of the language dataset (e.g., 'DE', 'EN', 'FR')

  • transform (callable, default: None ) –

    Function/transform that takes in a Data object and returns a transformed version. Default: None

  • pre_transform (callable, default: None ) –

    Function/transform to be applied to the data object before saving. Default: None

  • pre_filter (callable, default: None ) –

    Function that takes in a Data object and returns a boolean value, indicating whether the data object should be included. Default: None
Notes

Dataset Structure:

  • Nodes represent Twitch users
  • Edges represent user connections
  • Node features from user activities
  • Labels indicate user attributes
  • Includes train/val/test splits (80/10/10)
processed_file_names property

Names of processed data files.

Returns:
  • list[str] –

    List of processed file names
Notes

Processed files:

  • data.pt: Contains processed PyTorch Geometric data object
raw_file_names property

Names of required raw files.

Returns:
  • list[str] –

    List of required raw file names
Notes

Required files:

  • edges.csv: Network connectivity data
  • features.json: User feature data
  • target.csv: User labels
download()

Download raw data files.

Notes

Empty implementation - data should be manually placed in raw directory

load_dataset(lang)

Load raw Twitch dataset files.

Parameters:
  • lang (str) –

    Language code for the dataset (e.g., 'DE', 'EN', 'FR')
Returns:
  • tuple[csr_matrix, ndarray, ndarray] –

    Contains:

    • A: Adjacency matrix [num_nodes, num_nodes]
    • label: Binary labels [num_nodes]
    • features: Feature matrix [num_nodes, 3170]
Notes

Processing Steps:

  • Load target data:

    • Handle non-unique rows
    • Extract binary labels
    • Map node IDs

  • Load edge data:

    • Create adjacency matrix
    • Handle sparse connections

  • Load feature data:

    • Parse JSON features
    • Create sparse feature matrix

  • Reorder nodes:

    • Ensure consistent indexing
    • Map node IDs to indices
process()

Process raw data into PyTorch Geometric Data format.

Notes

Processing Steps:

  • Load dataset files:

    • Adjacency matrix
    • Node labels
    • Feature matrix

  • Convert to PyTorch format:

    • Edge indices from sparse adjacency
    • Float features from sparse matrix
    • Integer labels from binary values

  • Create Data object with:

    • Edge indices
    • Node features
    • Node labels
    • Train/val/test masks

  • Apply pre-transform if specified

  • Save processed data

Data Split:

  • Training: 80%
  • Validation: 10%
  • Testing: 10%

Features:

  • Sparse matrix handling
  • JSON feature processing
  • Random split generation
  • Optional pre-transform support