Tutorial (MovieLens)

This tutorial follows the benchmark path on MovieLens-1M. It is a narrative version of the configs/example.yaml run.

1. Choose the dataset

MovieLens-1M is small enough for quick iteration and has genre features for every movie.

datasets:
  - ml-1m

2. Pick a donor

Start with ALS. It trains on warm interactions only after the pseudo-cold split.

donors:
  - name: als
    params:
      factors: 64
      regularization: 0.05
      iterations: 20

3. Compare baselines and transfer

Grouped MP is the strong popularity baseline. linmap tests direct content-to-score transfer. knn_score_avg is the naive neighbor baseline.

methods:
  - name: grouped_most_popular_pers
  - name: linmap
  - name: knn_score_avg
    params:
      k: 20

4. Keep the split honest

splitter:
  name: pseudo_cold
  params:
    cold_frac: 0.2
    val_frac: 0.1
    n_pop_buckets: 5
    min_item_interactions: 5

The splitter removes pseudo-cold items from donor training and neighbor construction. See the evaluation protocol for the invariant.

5. Run

uv run warmbench --config configs/example.yaml --dry-run
uv run warmbench --config configs/example.yaml

Expected direction: linmap should be competitive with or stronger than the personalized Grouped MP baseline on AUC, while naive KNN is a useful check for inherited popularity.

6. Read the metrics

Ranking metrics (recall@k, precision@k, map@k, ndcg@k, mrr@k) show top-k quality. AUC is reported as an auxiliary per-user ranking signal. For final claims, compare across seeds rather than a single run.