TSDBench

# ============================================================
# LargeST_protocol/tasks/general_task.yaml
# ============================================================
# Purpose:
# - This file contains only processing rules shared by all tasks and naming conventions.

# Recommended reading order for each task.yaml:
# 1) Start with sample in each task.yaml: what constitutes one sample.
# 2) Read inputs: what the model can access at inference time.
# 3) Read prediction: what the model should predict.
# 4) Finally, read leakage: which columns must be excluded from model inputs.
# ============================================================

general_defaults:
  # ------------------------------------------------------------
  # Default input processing rules.
  # Note: these are processing defaults, not input field defaults.
  # These rules specify how to process time series, tabular, or graph inputs when a task uses them,
  # but do not automatically add any input fields to a task.
  # ------------------------------------------------------------
  input_processing:
    tabular_static:
      # Unique identifier column for sensors / road segments in LargeST.
      # This column is used only to join labels, splits, and metadata.
      id_column: ID

    time_series:
      # LargeST historical traffic flow data typically uses timestamp as the time axis and sensor IDs as columns.
      timestamp_column: timestamp
      id_column: ID
      timezone: UTC

      sampling:
        # LargeST raw historical traffic flow data typically has a 15-minute resolution.
        # If your h5 file has been preprocessed to a different resolution, override this in the specific task.
        native_rate: 15min

        # By default, the task resolution matches the native resolution.
        # For example, override this to 1h in task.yaml for a ranking task with hourly decisions.
        task_rate: 15min

        # Use this aggregation method when task_rate is coarser than native_rate, for example 15min -> 1h.
        # mean is suitable for average traffic flow; sum is suitable for cumulative flow counts, subject to the data semantics.
        downsample_method: mean

        # Use this interpolation method when task_rate is finer than native_rate, for example 15min -> 5min. This task configuration is generally not recommended.
        # ffill forward-fills using the most recent observation.
        upsample_method: ffill

    graph:
      # Graph inputs are static by default, for example LargeST's adj.npy.
      # Future dynamic graphs, such as event-induced or construction-impact graphs, can be configured in individual tasks.
      directed: false
      weighted: true


  # ------------------------------------------------------------
  # Standard split names.
  # Use these standard values in splits.csv to avoid mixing train/val/valid/dev.
  # ------------------------------------------------------------
  split_names:
    train: train
    validation: val
    test: test

  # ------------------------------------------------------------
  # Inheritance rules.
  # A task can set inherit_general: true to inherit this file's processing defaults.
  # Note: inheritance only affects default rules such as input_processing / columns / split_names,
  # and does not automatically inherit any specific input fields.
  # ------------------------------------------------------------
  inheritance:
    default_inherit_general: true
    merge_strategy: recursive_override

	# ============================================================
	# LargeST_protocol/tasks/general_task.yaml
	# ============================================================
	# Purpose:
	# - This file contains only processing rules shared by all tasks and naming conventions.

	# Recommended reading order for each task.yaml:
	# 1) Start with sample in each task.yaml: what constitutes one sample.
	# 2) Read inputs: what the model can access at inference time.
	# 3) Read prediction: what the model should predict.
	# 4) Finally, read leakage: which columns must be excluded from model inputs.
	# ============================================================

	general_defaults:
	# ------------------------------------------------------------
	# Default input processing rules.
	# Note: these are processing defaults, not input field defaults.
	# These rules specify how to process time series, tabular, or graph inputs when a task uses them,
	# but do not automatically add any input fields to a task.
	# ------------------------------------------------------------
	input_processing:
	tabular_static:
	# Unique identifier column for sensors / road segments in LargeST.
	# This column is used only to join labels, splits, and metadata.
	id_column: ID

	time_series:
	# LargeST historical traffic flow data typically uses timestamp as the time axis and sensor IDs as columns.
	timestamp_column: timestamp
	id_column: ID
	timezone: UTC

	sampling:
	# LargeST raw historical traffic flow data typically has a 15-minute resolution.
	# If your h5 file has been preprocessed to a different resolution, override this in the specific task.
	native_rate: 15min

	# By default, the task resolution matches the native resolution.
	# For example, override this to 1h in task.yaml for a ranking task with hourly decisions.
	task_rate: 15min

	# Use this aggregation method when task_rate is coarser than native_rate, for example 15min -> 1h.
	# mean is suitable for average traffic flow; sum is suitable for cumulative flow counts, subject to the data semantics.
	downsample_method: mean

	# Use this interpolation method when task_rate is finer than native_rate, for example 15min -> 5min. This task configuration is generally not recommended.
	# ffill forward-fills using the most recent observation.
	upsample_method: ffill

	graph:
	# Graph inputs are static by default, for example LargeST's adj.npy.
	# Future dynamic graphs, such as event-induced or construction-impact graphs, can be configured in individual tasks.
	directed: false
	weighted: true


	# ------------------------------------------------------------
	# Standard split names.
	# Use these standard values in splits.csv to avoid mixing train/val/valid/dev.
	# ------------------------------------------------------------
	split_names:
	train: train
	validation: val
	test: test

	# ------------------------------------------------------------
	# Inheritance rules.
	# A task can set inherit_general: true to inherit this file's processing defaults.
	# Note: inheritance only affects default rules such as input_processing / columns / split_names,
	# and does not automatically inherit any specific input fields.
	# ------------------------------------------------------------
	inheritance:
	default_inherit_general: true
	merge_strategy: recursive_override