Ames Housing

Kaggle House Prices competition entry using R and tidymodels. Predicts sale prices of residential homes in Ames, Iowa using a random forest model.

• Algorithm: Random forest (ranger engine via tidymodels)
• Score: 0.14541 (rank 2124)
• API: Plumber service deployed on Google Cloud Run
• Live docs: https://ames-api-576662713224.europe-west12.run.app/__docs__/

Dependencies are managed by renv. Restore the environment:

renv::restore()

Or install packages from scratch:

source("install-packages.R")

Rscript ames_training.R

This will:

1. Load and preprocess training data
2. Fit a random forest model using 10-fold CV with 5 repeats
3. Pin the model to board/ (local) and ames-r-model (GCS)
4. Generate a submission CSV to output/rf_submission.csv

• SalePrice is log10-transformed during training; predictions are reversed with 10^.pred
• Categorical columns with missing values (e.g. Alley, BsmtQual) get "None" as a fill level
• Numeric columns MSSubClass, MoSold, YrSold are treated as categorical
• Remaining numeric NAs are KNN-imputed

Rscript run_api_local.R

The API serves at http://localhost:8080. It loads the model from the GCS bucket ames-r-model using the vctrs/vetiver stack. If a GCS service account JSON key is present at the repo root, it's used for auth; otherwise Application Default Credentials are used.

The service runs on Google Cloud Run. Deploy with:

docker build -t ames-api:gcs -f Dockerfile .
docker tag ames-api:gcs europe-west12-docker.pkg.dev/ames-housing-472418/docker-repo/ames-api:gcs
docker push europe-west12-docker.pkg.dev/ames-housing-472418/docker-repo/ames-api:gcs

gcloud run deploy ames-api \
  --image=europe-west12-docker.pkg.dev/ames-housing-472418/docker-repo/ames-api:gcs \
  --region=europe-west12 \
  --platform=managed \
  --service-account=gcs-read-write@ames-housing-472418.iam.gserviceaccount.com \
  --allow-unauthenticated \
  --project=ames-housing-472418

The service account needs roles/storage.objectAdmin on the bucket ames-r-model.

• POST /predict — send feature data, get price prediction
• GET /__docs__ — interactive API documentation (via rapidoc)

The plumber API authenticates to Google Cloud Storage to read the pinned model:

1. If ames-housing-472418-0f31c1a0322f.json exists locally → uses that key directly
2. Otherwise → falls back to Application Default Credentials (Cloud Run environment)

The JSON key file is gitignored.

• test_API.R is stale and references a Titanic model. Do not use it as-is.
• The model object is ranger via tidymodels; inference requires parsnip, recipes, workflows, and ranger.
• R version: 4.5.1 (via rocker/r-ver:4.5.1)
• For development workflow and commands, see AGENTS.md.