etna.analysis.plot_feature_relevance

plot_feature_relevance(ts: TSDataset, relevance_table: RelevanceTable, normalized: bool = False, relevance_aggregation_mode: str | Literal['per-segment'] = AggregationMode.mean, relevance_params: Dict[str, Any] | None = None, top_k: int | None = None, alpha: float = 0.05, segments: List[str] | None = None, columns_num: int = 2, figsize: Tuple[int, int] = (10, 5))

Plot relevance of the features.

The most important features are at the top, the least important are at the bottom.

For StatisticsRelevanceTable also plot vertical line: transformed significance level.

• Values that lie to the right of this line have p-value < alpha.

• And the values that lie to the left have p-value > alpha.

Parameters:

• ts (TSDataset) – TSDataset with timeseries data

• relevance_table (RelevanceTable) –

  method to evaluate the feature relevance;

  • if StatisticsRelevanceTable table is used then relevances are normalized p-values

  • if ModelRelevanceTable table is used then relevances are importances from some model

• normalized (bool) – whether obtained relevances should be normalized to sum up to 1

• relevance_aggregation_mode (str | Literal['per-segment']) – aggregation strategy for obtained feature relevance table; all the strategies can be examined at AggregationMode

• relevance_params (Dict[str, Any] | None) – additional keyword arguments for the __call__ method of RelevanceTable

• top_k (int | None) – number of best features to plot, if None plot all the features

• alpha (float) – significance level, default alpha = 0.05, only for StatisticsRelevanceTable

• segments (List[str] | None) – segments to use

• columns_num (int) – if relevance_aggregation_mode="per-segment" number of columns in subplots, otherwise the value is ignored

• figsize (Tuple[int, int]) – size of the figure per subplot with one segment in inches