---

name: time-series description: "Use AIX360 TSICE, TSLime, and TSSaliency for local numeric explanations of univariate or multivariate time-series models, including temporal perturbations, relevant history, exogenous inputs, and integrated-gradient saliency." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Time-Series Local Explanations

Use this sub-skill for AIX360 0.3.0 local explanations over an ordered history window. It covers:

• TSLime: perturb the history, fit a local linear surrogate, and return time/feature weights.
• TSSaliency: estimate temporal and variate contributions along a path from a base signal using a model-agnostic Monte Carlo gradient.
• TSICE: perturb a selected time window and relate derived time-series features to changes in a multi-step forecast.
• tsFrame, to_np_array, model wrappers, perturbation engines, and numeric interpretation. Plotting is optional and is not part of the explanation contract.

Route here when

Choose this skill when the request mentions TSICE, TSLime, TSSaliency, temporal saliency, integrated gradients, relevant history, perturbation windows, exogenous time-series variables, or time-series attribution. Read the API reference, data-format contract, and the relevant section of workflows before running an explainer.

Route elsewhere

• Generic tabular, text, or image LIME/SHAP: use local-black-box.
• Dataset download or packaging: use datasets-and-metrics. For offline explanation tests, construct an in-memory fixture instead.
• CEM, certification, or recourse: use counterfactual-and-certification.

Operating contract

1. Normalize first. Represent a window as a pandas DataFrame with one row per time step, one numeric column per variate, and a DatetimeIndex (a tsFrame). A NumPy input must be two-dimensional (time, features) before conversion. Preserve column order and use the same order for feature_names and model reshaping.
2. Fix the axes. Record input_length=T, feature count F, and, for TSICE, forecast_lookahead=H, n_variables=F_out, and the exogenous count. TSLime accepts at least T rows and uses the last T; TSSaliency requires exactly T; TSICE accepts at least T and uses the last T.
3. Adapt the callable. Prefer a callable that handles batch arrays and returns one numeric target per sample. TSLime and TSSaliency are single- output explainers; aggregate a multi-output model before calling them. TSICE instead expects each call to return (H, n_variables) (or a length-H vector for one output). See data formats.
4. Select perturbations deliberately. Start with one small block-bootstrap engine. Add frequency, moving-average, shift, or impute engines only when their data assumptions fit the domain. Set a seed and use a small perturbation count for a smoke test; perturbation count multiplies model calls.
5. Run numeric explanation before plotting. Inspect keys and shapes, compare the instance and base predictions, and check that weights/saliency are aligned to the intended window. A plot is a presentation layer, not a verification gate. Plotting failures must not invalidate numeric output.
6. Keep experiments offline by default. Do not invoke SunspotDataset, FordDataset, or ClimateDataset merely to demonstrate the explainers: their constructors can download data. Use a deterministic tiny NumPy fixture and a local callable; see workflows.

Choose the explainer

Need	Use	Main numeric result
Local sensitivity over recent time/feature cells	TSLimeExplainer	history_weights with shape (relevant_history, F)
Temporal/variate attribution relative to a base signal	TSSaliencyExplainer	saliency with shape (T, F)
Forecast response to structured changes in a time window	TSICEExplainer	perturbation responses, feature statistics, and forecast deltas

For exact constructors, return shapes, perturbation configuration, and wrapper behavior, use api-reference.md. For failures, use troubleshooting.md rather than silently reshaping an input or output.

Interpretation guardrails

• A positive TSLime weight means the local surrogate associates increasing that cell with increasing the selected scalar model response; it is local and perturbation-distribution dependent, not a causal effect.
• A positive TSSaliency value is a positive contribution under the chosen base signal and approximate path integral. Magnitude is comparable across cells only after considering feature scales and the same base/configuration.
• TSICE signed_impact is the average forecast change for one perturbation and total_impact is a non-negative average RMS change. The output is a sampled response cloud indexed by derived statistics, not a guaranteed per-feature coefficient.
• For multi-step or multi-variable forecasts, always state whether a summary averages horizons/variables. A scalar summary can hide cancellation.

Do not use these explainers as certification, recourse, or counterfactual proof. Do not claim that a heatmap is an attribution result unless its numeric source and axes have been checked.