---

name: data-graph-interfaces description: "Operate DoWhy graph, pandas do-accessor, sampler, dataset, transformer, plotting, and time-series interfaces." disable-model-invocation: true metadata: disco-role: operating license: MIT

Data and Graph Interfaces

Use this sub-skill for the shared interface layer that sits between raw data, graphs, samplers, datasets, plotting, and time-lag helpers.

Route here

• Build or inspect causal graphs from NetworkX, DOT, GML, or DAGitty text.
• Align graph nodes, data columns, and lagged columns before downstream tasks.
• Use the pandas .causal.do accessor or a DoSampler.
• Choose variable types, sampler methods, or stateful reuse.
• Create synthetic datasets or reduce features before graph-based workflows.
• Prepare lagged or temporal graph abstractions.

Route away

• If the user wants classic CausalModel effect estimation, identification, do, or refutation, route to ../effect-estimation/SKILL.md.
• If the user wants dowhy.gcm sampling, interventions, attribution, or validation, route to ../graphical-causal-models/SKILL.md.
• If the user wants to learn a graph from data, treat CausalModel.learn_graph() and dowhy.graph_learners as deprecated. Hand discovery off to an external causal-discovery library, then return here only for import and alignment.

Read first

• references/data-graph-interfaces.md
• references/pandas-do-sampler.md
• references/time-series.md
• references/troubleshooting.md

Task selector

1. Need to normalize graph input? Use build_graph_from_str, CausalGraph, or build_graph.
2. Need interventional samples from a pandas frame? Use df.causal.do and the do-sampler references.
3. Need a built-in synthetic dataset or a simple feature reducer? Use the dataset and transformer references.
4. Need lagged graph or lagged column handling? Use the time-series helpers.
5. Need graph plotting? Use the plotting helpers with the optional backend note in troubleshooting.

Operating protocol

• Confirm graph nodes and DataFrame columns map cleanly before handing off to a downstream causal task.
• Prefer NetworkX DiGraph as the in-memory representation; accept DOT, GML, or DAGitty only as normalized inputs.
• For pandas do-sampling, remember that x may be a string, list, or dict. String/list keeps the original treatment values; dict sets intervention values.
• Treat stateful=True as a cache for repeated calls on the same accessor, and call reset() when you want a clean sampler state.
• If the user now needs effect estimation, refutation, or GCM queries, stop here and hand off to the sibling skill.

Minimal smoke check

Run the bundled smoke script to exercise graph parsing, pandas do-sampling, datasets, transformers, and temporal helpers:

python scripts/smoke_graph_and_do.py --help

Use the script as a safe starting point for quick validation or for adapting these interface patterns into another project.