Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Types of Digital Twins

When you select a dataset, RootCause reads it and proposes the most specific twin type the data supports (see Selecting your Dataset). There are four, and they differ on just two questions:

  • Is there a time dimension? Does each row belong to a sequence ordered in time, or is each row a standalone snapshot?

  • Are there multiple environments? Is the data one homogeneous set, or many distinct groups — products, cities, stores, customer segments — that you want modeled together?

Crossing those two questions gives the four types.

A 2x2 diagram of the four digital twin types, split by whether the data has a time dimension and whether it has multiple environments
The four twin types, by time dimension and number of environments.

The four types

1. Static

Observations with no time information. Each row is an independent case — one row per customer, say. The twin learns what drives what from a single snapshot, ignoring any time element. Best when your rows are independent cases rather than a series of dates.

2. Temporal

Observations with time: a single series ordered by a time column. The twin learns immediate effects and standard one-step temporal structure. Lagged effects — causes whose impact shows up several time steps later — need the experimental Delayed effects training option, which is off by default.

3. Multi-Environment Static

Many groups, no time. Each environment — a product, a region, a segment — is observed as a snapshot, and the twin discovers one shared cause-and-effect network across all of them while allowing a limited set of differences between them. This setting is also called pooled cross-section or multi-context data. Statistically it is not panel data, because there is no time axis, though the platform labels a trained one a Panel Static Model.

4. Multi-Environment Temporal

Many groups tracked over time — the richest type. Take an example: a table of sales orders tagged with a product and a city, recorded over time (just one example; your data will differ). It combines the time dimension of a Temporal twin with the grouping of a multi-environment twin. This is what statisticians call panel data; the platform labels a trained one a Panel Time Series Model.

Why "environments," not "groups"?

A multi-environment twin does more than sort your data into buckets. It assumes every environment shares the same underlying cause-and-effect structure, and that only a few specific mechanisms differ between them. That shared causal core is what the twin sets out to discover; the differences are treated as information, not noise.

This is why the platform says environments and not groups. In statistics a "group" is just a partition — a bucket of rows that may differ from the next bucket in any way at all. An "environment" is a stronger idea, borrowed from causal inference: a setting in which the causal mechanisms are mostly stable and only sparsely perturbed.

And the differences between environments are an asset, not a nuisance. Each environment behaves like a natural experiment: by seeing how cause and effect shift from one environment to the next, the twin can pin down the direction of relationships it could never resolve from a single pooled dataset. Modeling many environments together is therefore more accurate than modeling each one alone — provided they genuinely share a common causal core.

The research behind this idea is listed under Further reading.

How the platform chooses

The moment you select a dataset, RootCause inspects it, marks the most specific type the data supports as Recommended and preselects it, disabling the types the data cannot support. You can override the choice, but the auto-detected type is almost always the right starting point.

The builder's Choose your data step with the Dataset picker and the Type cards Static, Temporal, Multi-environment and Multi-environment temporal: Static is marked Recommended and selected, while the time-based types are disabled because the data has no time column
The Type selector after picking a dataset: Static is marked Recommended, and the types the data cannot support are disabled.

Further reading

The "environments" framing — and the finding that variation across environments helps causal discovery — comes from a body of causal-inference research:

  • Peters, J., Bühlmann, P. & Meinshausen, N. (2016). Causal inference by using invariant prediction: identification and confidence intervals. Journal of the Royal Statistical Society: Series B, 78(5), 947–1012. — introduces "environments" (Invariant Causal Prediction).

  • Arjovsky, M., Bottou, L., Gulrajani, I. & Lopez-Paz, D. (2019). Invariant Risk Minimization. arXiv:1907.02893. — popularised the term in machine learning.

  • Mooij, J. M., Magliacane, S. & Claassen, T. (2020). Joint Causal Inference from Multiple Contexts. Journal of Machine Learning Research, 21(99), 1–108. — discovery across multiple environments/contexts.

  • Perry, R., von Kügelgen, J. & Schölkopf, B. (2022). Causal Discovery in Heterogeneous Environments Under the Sparse Mechanism Shift Hypothesis. Advances in Neural Information Processing Systems, 35. — why sparse differences across environments make the full causal structure identifiable.

  • Günther, W., Ninad, U. & Runge, J. (2023). Causal Discovery for Time Series from Multiple Datasets with Latent Contexts. Uncertainty in Artificial Intelligence (UAI). — the multi-environment temporal (panel) case.

  • Sun, Y., Wang, Y., Jin, Y., Chan, D. & Koehler, J. (2017). Geo-level Bayesian Hierarchical Media Mix Modeling. Google Inc. — the hierarchical "groups / partial pooling" framing, for contrast.

Other Build a Digital Twin pages

See Digital Twin overview for the bigger picture.

Last updated