Compresso Recsys

compresso-recsys is the recommender-system companion package for Compresso. It provides dataset loaders, checkpoint builders, checkpoint read/write helpers, and retrieval metrics for sparse representation learning experiments.

The package distribution name is compresso-recsys; the Python import is:

import compresso_recsys as cr

Install

Install from PyPI:

pip install compresso-recsys

Install optional dataset export support:

pip install "compresso-recsys[datasets]"

Install the optional scikit-learn dependency for KNN models:

pip install "compresso-recsys[knn]"

For local development:

pip install -e ../compresso
pip install -e ".[dev,datasets,test]"

Quickstart

Build a MovieLens 1M checkpoint from Python:

import compresso_recsys as cr

checkpoint_path = cr.build_recsys_checkpoint(
    dataset="ml1m",
    checkpoint_path="artifacts/ml1m/exp001.zip",
    annotation_source="genres",
)

with cr.read_checkpoint(checkpoint_path) as root:
    split = cr.load_recsys_split(root)

print(split["x_train"].shape)

Build the same kind of checkpoint from the command line:

compresso-recsys-build-checkpoint \
  --dataset ml1m \
  --checkpoint_path artifacts/ml1m/exp001.zip \
  --annotation_source genres

Amazon Reviews 2023 checkpoints can use item metadata for cold-item retrieval experiments:

compresso-recsys-build-checkpoint \
  --dataset amazon2023 \
  --amazon_category Toys_and_Games \
  --checkpoint_path artifacts/amazon_toys/temporal_exp001.zip \
  --split_mode temporal \
  --temporal_period_hours 8136 \
  --metadata_text_fields title,features,description,categories \
  --min_entity_text_words 30 \
  --annotation_source none

Temporal checkpoints use expanding histories and cumulative mixed warm/cold catalogs. Their train, validation, and test matrices are aligned by the corresponding train_item_ids, val_item_ids, and test_item_ids arrays.

What Is Included

Steam supports game metadata for item cold-start and review dates for sequential and temporal checkpoints:

checkpoint_path = cr.build_recsys_checkpoint(
    dataset="steam",
    split_mode="leave_last_out",  # Or "item_split" / "temporal".
    checkpoint_path="artifacts/steam/llo.zip",
)

The first run downloads about 1.2 GB of reviews plus game metadata; later runs reuse the local archive and parsed cache. See the dataset guide for metadata, feature availability, cold-start examples, and defaults.

• Dataset utilities for GoodBooks, MovieLens 1M, MovieLens 20M, Amazon Reviews 2023, Steam, Netflix Prize, MSD Taste Profile, Gowalla, DBbook, and Last.fm-2K.
• Optional pretrained multimodal features for ML-1M, DBbook, and Last.fm-2K, stored with item IDs, availability masks, and encoder provenance. See the dataset guide. Other datasets can use the same feature format for user-computed embeddings, including Amazon text/images.
• ZIP checkpoint format for source/target splits, embeddings, sparse embeddings, metrics, and Compresso cluster-graph stages.
• Calibrated Recall and nDCG defaults, with optional standard Recall, Precision, Hit Rate, MRR, and MAP at configurable cutoffs.
• Batched EASE, ADMM and gradient-trained TEASER cold-start models, a content-similarity cold-start baseline, dense ELSA, and lottery-ticket compressed ELSA with streaming evaluation.
• Random and popularity baselines, normalized user-user and item-item cosine KNN, and multinomial denoising and variational autoencoders (MultDAE and MultVAE).
• Sequential recommenders over chronological histories: SimpleRNN, a recurrent next-item baseline; SimpleGPT, a causal transformer with a tied head; SASRec, a modernized variant of the published self-attentive architecture trained against sampled negatives; and SimpleBidirectionalTransformer, a sequence-to-set encoder that can train on the target matrices produced by temporal splits. They are built from replaceable parts — an ItemTokenizer owning the vocabulary and a SequenceBatcher owning the context window — so bringing your own vocabulary or context length does not mean forking a trainer.
• One evaluation path for both model shapes. A sequential and a matrix model are compared through the same evaluate_recommender and compare_models calls, because a recommender is anything implementing predict_on_batch.
• One production recommendation path through model.recommend(histories, k=...). Histories and optional allow/block filters use stable item IDs; ranked item IDs and scores come back in a batch-shaped Recommendations result. Rows with fewer than k valid candidates truncate without weakening the filters.
• Unified fitted-model persistence through model.save(path) and ModelClass.load(path), using versioned ZIP archives that carry model state, configuration, and the vocabulary or candidate catalog needed for serving.
• A checkpoint-building console command: compresso-recsys-build-checkpoint.

Documentation

Release documentation is available at:

https://zombak79.github.io/compresso-recsys/

The full CLI parameter table, checkpoint split schema, and supported Amazon Reviews 2023 categories are maintained in the Checkpoint CLI Reference. Academic references and copy-ready BibTeX for EASE, TEASER, ELSA, large-scale ELSA, compressed ELSA, neighborhood models, Mult-DAE, and Mult-VAE are available in the citation guide. The implementing a recommender tutorial builds a complete Top Popular model with validation, stable-ID recommendation, persistence, evaluation, and executable contract checks.

Build the docs locally:

pip install -e ".[docs]"
sphinx-build -b html docs/source docs/build/html

License

Apache License 2.0. See LICENSE.