Skip to content

09-create-feature-group.md

Latest commit

 

History

History
119 lines (89 loc) · 5.72 KB

09-create-feature-group.md

File metadata and controls

119 lines (89 loc) · 5.72 KB
Raw

Create a Feature Group

Use this guide to find the right pattern for your feature group.

Decision Tree

Q1: Does it load external data (file, DB, external API)?
    YES → Pattern 1: Root Feature (if data passed at runtime, also see Q18)
    NO  → Continue

Q2: How many input features?
    0 → Pattern 1: Root Feature (or DataCreator, see 01-root-features)
    1+ → Pattern 2: Derived (see Q3)

Q3: Should it be reusable via naming pattern (input__operation)?
    YES → Add Pattern 3: FeatureChainParserMixin (if 2+ inputs, use Pattern 4)
    NO  → Continue

Q4: How many output columns?
    1 → Standard
    N → Pattern 5: Multi-output (feature~0, feature~1)

Q5: Does it need fitted/trained state between runs?
    YES → Add Pattern 6: Artifact

Q6: Does it need time ordering or group-by?
    YES → Add Pattern 7: Index

Q7: Does it join data from multiple feature groups?
    YES → Add Pattern 8: Links/Joins

Q8: Framework-specific API needed?
    YES → Pattern 9: Framework-specific

Q9: Does the feature name come from configuration vs naming pattern?
    Configuration → Pattern 3 with PROPERTY_MAPPING
                    (declare the value space via allowed_values / property_spec; see Pattern 11)

Q10: Does it need custom matching logic beyond class name?
    YES → See 14-feature-matching

Q11: Does it need to filter data (time ranges, categories)?
    YES → See 15-filter-concepts

Q12: Does it need input/output validation?
    YES → See 16-validators

Q13: Does it need to match against a data connection?
    YES → See 17-data-connection-matching

Q14: Does it need explicit data type declaration?
    YES → See 18-datatypes

Q15: Could multiple FeatureGroups handle the same feature name?
    YES → See 19-domain

Q16: Do you need to track changes or ensure reproducibility?
    YES → See 20-versioning

Q17: Need to understand calculate_feature() and runtime context?
    YES → See 12-calculate-feature

Q18: Does it receive data programmatically at runtime (e.g. via API call)?
    YES → See ApiInputData docs: https://mloda-ai.github.io/mloda/in_depth/api-input-data/

Q19: Need to create feature groups dynamically or simplify complex input sources?
    YES → See 21-experimental-shortcuts

Q20: Need to define features via JSON config (e.g., for AI agents or config-driven pipelines)?
    YES → See 22-feature-config

Q21: Need to consume results incrementally as each feature group completes?
    YES → See 23-streaming

Q22: Need to build the execution plan once at startup and reuse it for repeated calls with fresh data?
    YES → See 24-realtime

Q23: Need both plan reuse and incremental per-group delivery?
    YES → Use session.stream_run() — see 23-streaming and 24-realtime

Q24: Need conditional aggregation (include only rows matching a predicate)?
    YES → See 25-masking

Pattern Guides

Pattern When to Use
01-root-features Loading data from files, APIs, databases, or generating synthetic data
02-derived-features Simple transformation with static output name
03-chained-features Reusable transforms via naming pattern (price__scaled)
04-multi-input-features Combining multiple inputs (a&b__distance)
05-multi-output-features Multiple output columns (emb~0, emb~1)
06-artifact-features Fitted models, cached computations
07-index-features Time series, group-by, window functions
08-links-joins Joining data from multiple feature groups
09-framework-specific Pandas-only, Polars-only features
10-testing-guide 3-level testing approach

Concepts

Guide What It Covers
11-options Group vs context options for feature configuration
12-calculate-feature Implementing the core computation method
13-feature-naming How to define feature names
14-feature-matching How mloda finds the right FeatureGroup
15-filter-concepts Filtering data during computation
16-validators Input/output validation
17-data-connection-matching Matching against data connections
18-datatypes Data type declaration and validation
19-domain Disambiguate feature-to-FeatureGroup matching
20-versioning Version tracking and reproducibility
21-experimental-shortcuts Dynamic creation and input source helpers
22-feature-config JSON-based feature definition for AI agents and config pipelines
23-streaming Consume results incrementally with stream_all
24-realtime Reuse execution plans with prepare + run
25-masking Conditional aggregation via mask context option