carto-composite-scoring

Use this skill whenever the user wants to create a composite score, index, or multi-variable ranking in a CARTO Workflow.

Prerequisites: Load carto-create-workflow for the development process. It covers both access paths — the MCP server's workflow tools (create_workflow, validate_workflow, run_workflow) when attached, the carto workflows CLI otherwise. Routing signals: carto-basics/references/access-paths.md.

Instructions

Step 1: Determine the scoring approach

Ask the user the following decision tree:

1. "Do you have a target/outcome variable?" (e.g. revenue, sales, crime rate)

• Yes → Supervised method using native.spatialcompositesupervised

1. "No target, but do you have expert knowledge of variable importance?" (i.e. custom weights)

• 2a. Transparency matters more than node count → native.normalize per variable + native.selectexpression with the weighted-sum literally in SQL (a_norm * 0.5 + b_norm * 0.2 + c_norm * 0.3). Inputs and weights are inspectable inline in Workflows; each normalized column is materialized; no Analytics Toolbox dependency. Use when stakeholders need to see/audit the formula, or when you're not running on AT.
• 2b. Want the AT pipeline (scaling, aggregation, bucketing) handled for you → native.spatialcompositeunsupervised with scoring_method: CUSTOM_WEIGHTS. One node, but the formula is hidden inside the component.

1. "No target, no weights?"

• → Unsupervised method with ENTROPY or FIRST_PC using native.spatialcompositeunsupervised

Success: You have identified which component and scoring method to use before designing the pipeline.

Step 2: Build the pipeline

Supervised pipeline (native.spatialcompositesupervised)

1. Load the spatial features dataset (pre-indexed at H3 or Quadbin)
2. Load the target/outcome dataset
3. Join both datasets on the spatial index column
4. Select only the relevant feature columns (drop spatial index column and geometry from feature selection — pass only actual feature variables)
5. Run native.spatialcompositesupervised with parameters:

• model_type: LINEAR_REG
• bucketize: EQUAL_INTERVALS_ZERO_CENTERED
• n_buckets: 5
• outlier_removal: true
• r_squared_threshold: 0.4

1. Output: composite score based on regression residuals (identifies areas that over/under-perform relative to the model)

Success: The workflow joins features with the target variable, selects only numeric feature columns, and produces a residual-based score per spatial cell.

Unsupervised pipeline (native.spatialcompositeunsupervised)

1. Load the spatial features dataset
2. Select only the relevant feature columns
3. Encode any categorical/ordinal variables to numeric using native.casewhen (e.g. "Lowdensityurban" → 4, "Highdensityurban" → 2)
4. Optionally reverse variables where higher = worse by multiplying by -1 in the SELECT query passed to the component
5. Run native.spatialcompositeunsupervised with parameters:

• scoring_method: CUSTOM_WEIGHTS / ENTROPY / FIRST_PC
• weights: required if CUSTOM_WEIGHTS. Wire format is a JSON-encoded string of triples `[[<column>, <weight>, <reverse>], ...]` — column is the feature column name, weight is a numeric weight (normalized internally to sum to 1), reverse is a boolean (true flips the variable so higher = worse becomes higher = better without needing a CASE WHEN). Example: '[["population_density", 0.5, false], ["accident_rate", 0.5, true]]'. Not an object map.
• scaling: RANKING
• aggregation: LINEAR
• output: RETURN_RANGE with range [0, 1]

1. Output: composite index score per location

Success: All input variables are numeric, variable directions are aligned (higher = better for the score), and the output is a normalized score per spatial cell.

Gotchas

• Provider casing & SQL dialect. This skill uses lowercase column names (BigQuery / Databricks / Postgres / Redshift convention). On Snowflake, unquoted identifiers surface UPPERCASE — reference POPULATION_DENSITY, ACCIDENT_RATE, etc. in weights, expressions, and downstream SQL. See carto-create-workflow/references/providers/<provider>.md for casing rules and SQL dialect equivalents.
• All input variables must be numeric, with aligned direction. Encode ordinal strings ("low"/"medium"/"high") via CASE WHEN first, and where "higher is worse" multiply by -1 (or set the reverse flag) before scoring — forgetting this inverts the score meaning.
• Supervised R-squared threshold (default 0.4) is permissive. If model fit is poor, the residual-based score is mostly noise. Inspect model diagnostics.
• Custom weights are normalized internally to sum to 1. The absolute values do not matter, only the ratios.
• Supervised scores are residuals, not raw values. The score identifies areas that DEVIATE from the model, not areas with the highest raw values.
• Drop the spatial index column and geometry from the feature selection — only pass actual feature variables to the scoring component.
• `weights` type label is misreported. The component schema (via read_workflow_components MCP or carto workflows components get native.spatialcompositeunsupervised --json) lists weights as type: "ColumnNumber", but the actual wire shape is a ColumnAndNumber JSON-encoded string of triples [[<column>, <weight>, <reverse>], ...] (see Step 2 above). Trust this skill and the reference templates over the reported type label.
• `indexcol.allowedColumns` is a UI hint, not a validator. The schema lists ["geoid", "h3", "quadbin"] for native.spatialcompositeunsupervised.indexcol, but it's a Workflows dropdown hint — any unique identifier column works at runtime (e.g. store_id, cell_id).

Reference Templates

Both examples use Milan spatial features data at Quadbin resolution 18.

Common Variations