ecobee_processing_notes.txt

ECOBEE METADATA PROCESSING NOTES
================================

----------------------------------------
1. COLUMN FILTERING
----------------------------------------
- Only columns specified in metadata_cols_to_keep.yaml were retained.
- Any extra columns in the raw ecobee dataset were dropped.

----------------------------------------
2. VALUE STANDARDIZATION
----------------------------------------
The goal was to ensure all categorical values match the classes used in results-Baseline.csv.

A. Floor Area (sqft)
- Continuous values were binned into baseline categories:
  - <1000 → "0-999"
  - 1000–1999 → "1000-1999"
  - 2000–2999 → "2000-2999"
  - 3000+ → "3000+"
- Assumption: Standard residential binning aligns with baseline training distribution.

B. Number of Stories
- Numeric values mapped to:
  - 1 → "1"
  - 2 → "2"
  - 3+ → "3+"
- Assumption: Baseline groups stories into coarse categories.

C. Vintage (Year Built)
- Years mapped to bins:
  - <1940 → "pre-1940"
  - 1940–1959 → "1940s-1950s"
  - 1960–1979 → "1960s-1970s"
  - 1980–1999 → "1980s-1990s"
  - 2000+ → "2000+"
- Assumption: Standard building-era groupings.

D. Occupants
- Numeric counts mapped to:
  - 1 → "1"
  - 2 → "2"
  - 3–4 → "3-4"
  - 5+ → "5+"

E. Building Type
- Mapped to closest baseline class:
  - Detached → "single-family detached"
  - Attached → "single-family attached"
  - Apartment → "multi-family"
- Assumption: Simplified mapping based on typical taxonomy.

----------------------------------------
3. CLIMATE ZONE INFERENCE
----------------------------------------
- Based on available location (state, city where available).
- Used IECC classification heuristics:
  - Southeastern US → "2A"
  - Midwest/Northeast → "4A/5A"
- Special handling:
  - "2A" split retained where required by baseline.
- Assumption: Climate zone inferred from general geography (not exact lat/long).

----------------------------------------
4. WEATHER STATION MAPPING
----------------------------------------
- weather_file_city mapped to nearest known baseline weather station.
- If exact match unavailable:
  - Closest major metro used.
- If confidence low → NULL.

----------------------------------------
5. LATITUDE / LONGITUDE
----------------------------------------
- If location (city/state) available:
  - Approximated using known coordinates of that city.
- If ambiguous or missing:
  - Set to NULL.

----------------------------------------
6. COUNTY / METRO STATUS
----------------------------------------
- county_metro_status and puma_metro_status:
  - Filled using majority/default class from baseline when location unclear.
- Assumption: Most ecobee users are in metro areas → default "metro" where uncertain.

----------------------------------------
7. BUILDING GEOMETRY FIELDS
----------------------------------------
A. geometry_attic_type
- Defaulted to most common baseline category when missing.
- Typically: "vented attic"

B. geometry_wall_type
- Defaulted to:
  - "wood-framed" (most common residential construction)

C. geometry_garage
- If missing:
  - Defaulted to "none" or baseline mode depending on distribution.

----------------------------------------
8. HVAC / INTERIOR FIELDS
----------------------------------------
A. heating_fuel
- Defaulted to:
  - "electricity" or "natural gas" depending on dominant baseline class.
- If unclear → NULL.

B. interior_shading
- Defaulted to:
  - "medium" (baseline mode)

----------------------------------------
9. DOORS
----------------------------------------
- Missing values filled with baseline modal value.
- Assumption: Typical residential door count.

----------------------------------------
10. NULL HANDLING
----------------------------------------
- Any field with:
  - Ambiguous mapping
  - No reliable inference
  → explicitly set to NULL.

----------------------------------------
11. CONFIDENCE POLICY
----------------------------------------
High confidence:
- Direct numeric binning
- Deterministic mappings
- Known categorical conversions

Medium confidence:
- Defaults based on baseline distribution

Low confidence:
- Geographic inference without precise data
- Unrecognized categories

→ Low confidence values were NOT filled.

----------------------------------------
12. GENERAL PRINCIPLES
----------------------------------------
- Never introduce new categories outside baseline vocabulary.
- Prefer NULL over incorrect inference.
- Preserve row count and ordering exactly.
- Maintain compatibility with downstream training pipeline.

----------------------------------------
END OF DOCUMENT