Citing Compresso Recsys

If you use Compresso Recsys in academic work, cite the methods, datasets, and additional metadata sources used in your experiment. Dataset citations and preprocessing-method citations serve different purposes; include both when applicable. See Datasets for the actual sources and loader defaults.

Compresso 

If your work uses the Compresso sparse-representation framework, cite the project:

@misc{compresso,
  title  = {Compresso: A PyTorch Framework for Sparse Representation Learning},
  author = {Van{\v{c}}ura, Vojt{\v{e}}ch and Giacomo Medda and Spi{\v{s}}{\'a}k, Martin and Ladislav Pe{\v{s}}ka},
  year   = {2026},
  url    = {https://github.com/zombak79/compresso}
}

EASE 

For compresso_recsys.models.EASE, cite the original EASE paper:

@inproceedings{steck2019embarrassingly,
  title={Embarrassingly shallow autoencoders for sparse data},
  author={Steck, Harald},
  booktitle={The World Wide Web Conference},
  pages={3251--3257},
  year={2019}
}

Neighborhood Models 

For compresso_recsys.models.UserKNNRecommender, cite the GroupLens user-based collaborative-filtering paper:

@inproceedings{resnick1994grouplens,
  title={GroupLens: An Open Architecture for Collaborative Filtering of Netnews},
  author={Resnick, Paul and Iacovou, Neophytos and Suchak, Mitesh and
          Bergstrom, Peter and Riedl, John},
  booktitle={Proceedings of the 1994 ACM Conference on Computer Supported
             Cooperative Work},
  pages={175--186},
  year={1994},
  doi={10.1145/192844.192905}
}

For compresso_recsys.models.ItemKNNRecommender, cite the original item-based collaborative-filtering paper:

@inproceedings{sarwar2001item,
  title={Item-Based Collaborative Filtering Recommendation Algorithms},
  author={Sarwar, Badrul and Karypis, George and Konstan, Joseph and
          Riedl, John},
  booktitle={Proceedings of the 10th International Conference on World Wide Web},
  pages={285--295},
  year={2001},
  doi={10.1145/371920.372071}
}

Mult-VAE, Mult-DAE, and AutoRec 

For compresso_recsys.models.MultDAETrainer or compresso_recsys.models.MultVAETrainer, cite the paper that introduced both models:

@inproceedings{liang2018variational,
  title={Variational Autoencoders for Collaborative Filtering},
  author={Liang, Dawen and Krishnan, Rahul G. and Hoffman, Matthew D. and
          Jebara, Tony},
  booktitle={Proceedings of the 2018 World Wide Web Conference},
  pages={689--698},
  year={2018},
  doi={10.1145/3178876.3186150}
}

AutoRec established the collaborative-filtering autoencoder architecture that preceded Mult-DAE:

@inproceedings{sedhain2015autorec,
  title={AutoRec: Autoencoders Meet Collaborative Filtering},
  author={Sedhain, Suvash and Menon, Aditya Krishna and Sanner, Scott and
          Xie, Lexing},
  booktitle={Proceedings of the 24th International Conference on World Wide Web},
  pages={111--112},
  year={2015},
  doi={10.1145/2740908.2742726}
}

TEASER 

For compresso_recsys.models.TEASER or compresso_recsys.models.TEASERGDTrainer, cite the original TEASER paper:

@inproceedings{depauw2022who,
  title={Who do you think I am? Interactive User Modelling with Item Metadata},
  author={De Pauw, Joey and Ruymbeek, Koen and Goethals, Bart},
  booktitle={Proceedings of the 16th ACM Conference on Recommender Systems},
  pages={640--643},
  year={2022},
  doi={10.1145/3523227.3551470}
}

ELSA 

For standard compresso_recsys.models.ELSA training, cite the original ELSA paper:

@inproceedings{vanvcura2022scalable,
  title={Scalable linear shallow autoencoder for collaborative filtering},
  author={Van{\v{c}}ura, Vojt{\v{e}}ch and Alves, Rodrigo and Kasalick{\`y}, Petr and Kord{\'\i}k, Pavel},
  booktitle={Proceedings of the 16th ACM conference on recommender systems},
  pages={604--609},
  year={2022}
}

ELSA with sampled output candidates 

If your work uses ELSAConfig.max_output to train against sampled output candidates at scale, also consider citing the large-scale ELSA evaluation:

@article{10.1145/3748335,
  author = {Van\v{c}ura, Vojt\v{e}ch and Kasalick\'{y}, Petr and Alves, Rodrigo and Kord\'{\i}k, Pavel},
  title = {Evaluating Linear Shallow Autoencoders on Large Scale Datasets},
  year = {2025},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3748335},
  doi = {10.1145/3748335},
  journal = {ACM Trans. Recomm. Syst.},
}

Compressed ELSA 

If your work uses compresso_recsys.models.CompressedELSA, cite the sparse-representation paper. The original ELSA citation above may also be appropriate when ELSA itself is central to the work.

@inproceedings{vanvcura2026efficient,
  title={Efficient Learning of Sparse Representations from Interactions},
  author={Van{\v{c}}ura, Vojt{\v{e}}ch and Spi{\v{s}}{\'a}k, Martin and Alves, Rodrigo and Pe{\v{s}}ka, Ladislav},
  booktitle={Proceedings of the ACM Web Conference 2026},
  pages={8577--8580},
  year={2026}
}

SASRec 

For compresso_recsys.models.SASRecTrainer, cite the original SASRec paper. Its sequential objective, tied item scoring, and published MovieLens hyperparameters are SASRec-derived. The transformer block uses a modernized pre-norm attention architecture, so published results are a point of comparison rather than exact implementation parity:

@inproceedings{kang2018self,
  title={Self-Attentive Sequential Recommendation},
  author={Kang, Wang-Cheng and McAuley, Julian},
  booktitle={2018 IEEE International Conference on Data Mining (ICDM)},
  pages={197--206},
  year={2018},
  doi={10.1109/ICDM.2018.00035}
}

SimpleRNN 

For compresso_recsys.models.SimpleRNNTrainer, GRU4Rec is the relevant recurrent recommendation reference. Compresso’s model supports GRU or LSTM encoders and uses full-catalog next-item cross-entropy on user histories; it does not reproduce GRU4Rec’s session-parallel batching or original ranking losses. Cite this paper as methodological background and describe the actual configuration used:

@inproceedings{hidasi2016session,
  title={Session-based Recommendations with Recurrent Neural Networks},
  author={Hidasi, Balazs and Karatzoglou, Alexandros and Baltrunas, Linas and Tikk, Domonkos},
  booktitle={International Conference on Learning Representations},
  year={2016},
  url={https://arxiv.org/abs/1511.06939}
}

SimpleGPT and SimpleBidirectionalTransformer 

compresso_recsys.models.SimpleGPTTrainer uses a causal Transformer with a learned prefix and tied item embeddings by default. Its implementation follows the nanoGPT architecture, adapted to item sequences. compresso_recsys.models.SimpleBidirectionalTransformerTrainer reuses those building blocks with bidirectional attention and a pooled CLS state to predict an unordered target set. Cite the Transformer foundation, and attribute nanoGPT when discussing the implementation:

@inproceedings{vaswani2017attention,
  title={Attention Is All You Need},
  author={Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob
          and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia},
  booktitle={Advances in Neural Information Processing Systems},
  volume={30},
  year={2017},
  url={https://arxiv.org/abs/1706.03762}
}

@misc{karpathy_nanogpt,
  author={Karpathy, Andrej},
  title={{nanoGPT}},
  howpublished={GitHub repository},
  url={https://github.com/karpathy/nanoGPT},
  note={Record the revision used in your experiment}
}

These are Compresso baseline implementations, not separate published methods. In particular, the bidirectional model is not an implementation of BERT4Rec’s masked-item objective. Its set-valued multinomial loss is related to the Mult-VAE/Mult-DAE formulation above; cite that work when the loss is relevant.

Content, Popularity, and Random Baselines 

compresso_recsys.models.ContentRecommender uses item features and similarity scoring; cite the feature/embedding model actually used to create those features. compresso_recsys.models.PopularityBaseline and compresso_recsys.models.RandomBaseline are elementary baselines with no single originating paper attributed by this package. Report their settings and the Compresso Recsys version or commit. Catalog adapters and model base classes are infrastructure, not additional recommendation methods.

Datasets 

MovieLens 1M and 20M 

For compresso_recsys.datasets.MovieLens1M and compresso_recsys.datasets.MovieLens20M, cite the dataset history paper and state the exact variant. GroupLens usage terms remain applicable.

@article{harper2015movielens,
  title={The MovieLens Datasets: History and Context},
  author={Harper, F. Maxwell and Konstan, Joseph A.},
  journal={ACM Transactions on Interactive Intelligent Systems},
  volume={5},
  number={4},
  year={2015},
  doi={10.1145/2827872},
  url={https://doi.org/10.1145/2827872}
}

Goodbooks-10k 

For compresso_recsys.datasets.Goodbooks, credit the Goodbooks-10k release. The upstream repository supplies the dataset rather than a designated research paper; use a dataset/software citation and retain its attribution and usage terms. Goodbooks-10k is distinct from McAuley Lab’s Goodreads datasets.

@misc{zygmuntz2017goodbooks,
  author={{zygmuntz}},
  title={{Goodbooks-10k}: Ten Thousand Books, Six Million Ratings},
  year={2017},
  howpublished={Dataset, release v1.0},
  url={https://github.com/zygmuntz/goodbooks-10k}
}

Amazon Reviews 2023 

For compresso_recsys.datasets.AmazonReviews2023, use the citation recommended by the official dataset page. Report the category, source revision, and filtering; citations for older Amazon releases do not identify this collection.

@article{hou2024bridging,
  title={Bridging Language and Items for Retrieval and Recommendation},
  author={Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and
          Chen, Xiusi and McAuley, Julian},
  journal={arXiv preprint arXiv:2403.03952},
  year={2024},
  url={https://arxiv.org/abs/2403.03952}
}

Steam 

For compresso_recsys.datasets.Steam, the upstream Steam page lists SASRec (SASRec) and the two references below. Report that the adapter uses steam_reviews.json.gz and steam_games.json.gz: it does not load bundle interactions or the BERT4Rec remapped sequence release.

@inproceedings{wan2018monotonic,
  title={Item Recommendation on Monotonic Behavior Chains},
  author={Wan, Mengting and McAuley, Julian},
  booktitle={Proceedings of the 12th ACM Conference on Recommender Systems},
  year={2018},
  url={https://cseweb.ucsd.edu/~jmcauley/pdfs/recsys18b.pdf}
}

@inproceedings{pathak2017steam,
  title={Generating and Personalizing Bundle Recommendations on Steam},
  author={Pathak, Apurva and Gupta, Kshitiz and McAuley, Julian},
  booktitle={Proceedings of the 40th International ACM SIGIR Conference on
             Research and Development in Information Retrieval},
  year={2017},
  url={https://cseweb.ucsd.edu/~jmcauley/pdfs/sigir17.pdf}
}

Netflix Prize 

For compresso_recsys.datasets.NetflixPrize, cite the original dataset description. Internet Archive is the adapter’s download location, not the dataset’s creator. Original Netflix terms remain applicable.

@inproceedings{bennett2007netflix,
  title={The Netflix Prize},
  author={Bennett, James and Lanning, Stan},
  booktitle={Proceedings of KDD Cup and Workshop 2007},
  year={2007},
  url={https://www.cs.uic.edu/~liub/KDD-cup-2007/proceedings/The-Netflix-Prize-Bennett.pdf}
}

MSD / Taste Profile 

For compresso_recsys.datasets.TasteProfile, cite the Million Song Dataset paper and explicitly identify the Taste Profile subset contributed by The Echo Nest, using train_triplets.txt. The adapter loads user-song play counts, not the full MSD audio features or song metadata. See the MSD source for attribution.

@inproceedings{bertinmahieux2011million,
  title={The Million Song Dataset},
  author={Bertin-Mahieux, Thierry and Ellis, Daniel P. W. and Whitman, Brian
          and Lamere, Paul},
  booktitle={Proceedings of the 12th International Society for Music
             Information Retrieval Conference},
  year={2011},
  url={https://ismir2011.ismir.net/papers/OS6-1.pdf}
}

Gowalla 

For compresso_recsys.datasets.Gowalla, cite the paper requested by SNAP. State that the adapter uses raw check-ins; the social graph is not downloaded.

@inproceedings{cho2011friendship,
  title={Friendship and Mobility: User Movement in Location-Based Social Networks},
  author={Cho, Eunjo and Myers, Seth A. and Leskovec, Jure},
  booktitle={Proceedings of the 17th ACM SIGKDD International Conference on
             Knowledge Discovery and Data Mining},
  year={2011},
  url={https://snap.stanford.edu/data/loc-gowalla.html}
}

Additional Movie and Book Descriptions 

The MovieLens and Goodbooks adapters also load generated descriptions from beeFormer. If those descriptions contribute to your experiment’s item features, cite beeFormer in addition to the original dataset:

@inproceedings{vancura2024beeformer,
  title={beeFormer: Bridging the Gap Between Semantic and Interaction Similarity
         in Recommender Systems},
  author={Van{\v{c}}ura, Vojt{\v{e}}ch and Kord{\'\i}k, Pavel and Straka, Milan},
  booktitle={Proceedings of the 18th ACM Conference on Recommender Systems},
  pages={1102--1107},
  year={2024},
  doi={10.1145/3640457.3691707},
  url={https://github.com/recombee/beeformer}
}

Evaluation Protocol 

Held-out users are evaluated under strong generalization: each user’s history is split into a fold-in part the model sees and a held-out part it is scored against. eval_holdout_frac defaults to 0.2, matching the 80/20 split described by Liang et al. If you report numbers from user_split, cite the Mult-VAE/Mult-DAE paper above.

The default eval_draws=1 produces one source/target split per held-out user. Opting into five independent draws per user with eval_draws=5 follows the ELSA line of work; cite the ELSA papers above when reporting under that protocol.

Dataset preprocessing 

Netflix and Taste Profile support/feedback defaults follow the Mult-VAE paper above. Steam’s every-review-as-feedback convention follows BERT4Rec, while Gowalla’s 10-interaction CF support defaults follow NGCF:

@inproceedings{sun2019bert4rec,
  title={{BERT4Rec}: Sequential Recommendation with Bidirectional Encoder
         Representations from Transformer},
  author={Sun, Fei and Liu, Jun and Wu, Jian and Pei, Changhua and Lin, Xiao
          and Ou, Wenwu and Jiang, Peng},
  booktitle={Proceedings of the 28th ACM International Conference on
             Information and Knowledge Management},
  year={2019},
  doi={10.1145/3357384.3357895},
  url={https://arxiv.org/abs/1904.06690}
}

@inproceedings{wang2019ngcf,
  title={Neural Graph Collaborative Filtering},
  author={Wang, Xiang and He, Xiangnan and Wang, Meng and Feng, Fuli and Chua, Tat-Seng},
  booktitle={Proceedings of the 42nd International ACM SIGIR Conference on
             Research and Development in Information Retrieval},
  year={2019},
  url={https://arxiv.org/abs/1905.08108}
}

These citations explain the defaults, not exact reproduction of published results. Report the actual split_mode, support thresholds, seed, and metric definitions. Compresso’s LLO uses three stage targets; temporal splits use global time windows; neither is the original BERT4Rec sampled-ranking protocol. The raw Gowalla adapter does not use NGCF/LightGCN’s published train/test files.

Statistical Comparison 

If you report significance from compresso_recsys.stats, cite the methods it implements. See Comparing Models Statistically for what each one contributes and a ready-to-adapt methods paragraph.

The paired bootstrap behind the confidence interval:

@article{efron1979bootstrap,
  title={Bootstrap Methods: Another Look at the Jackknife},
  author={Efron, Bradley},
  journal={The Annals of Statistics},
  volume={7},
  number={1},
  pages={1--26},
  year={1979},
  doi={10.1214/aos/1176344552}
}

The randomization test used for the p-value, and the comparison of significance tests that motivates choosing it for retrieval evaluation:

@inproceedings{smucker2007comparison,
  title={A Comparison of Statistical Significance Tests for Information
         Retrieval Evaluation},
  author={Smucker, Mark D. and Allan, James and Carterette, Ben},
  booktitle={Proceedings of the Sixteenth ACM Conference on Information and
             Knowledge Management},
  series={CIKM '07},
  pages={623--632},
  year={2007},
  publisher={ACM},
  doi={10.1145/1321440.1321528}
}

Why a Monte Carlo p-value is computed as (1 + extreme) / (B + 1) rather than as a plain proportion:

@article{phipson2010permutation,
  title={Permutation P-values Should Never Be Zero: Calculating Exact
         P-values When Permutations Are Randomly Drawn},
  author={Phipson, Belinda and Smyth, Gordon K.},
  journal={Statistical Applications in Genetics and Molecular Biology},
  volume={9},
  number={1},
  pages={Article 39},
  year={2010},
  doi={10.2202/1544-6115.1585}
}

The multiple-testing correction applied across a comparison report:

@article{holm1979simple,
  title={A Simple Sequentially Rejective Multiple Test Procedure},
  author={Holm, Sture},
  journal={Scandinavian Journal of Statistics},
  volume={6},
  number={2},
  pages={65--70},
  year={1979}
}

The Monte Carlo p-value form and the convention of choosing the resample count so that alpha * (B + 1) is an integer:

@book{davison1997bootstrap,
  title={Bootstrap Methods and their Application},
  author={Davison, A. C. and Hinkley, D. V.},
  publisher={Cambridge University Press},
  year={1997},
  doi={10.1017/CBO9780511802843}
}

SWAP multimodal ML-1M, DBbook, and Last.fm-2K 

For pretrained features, cite the resource paper and the versioned dataset. Continue citing MovieLens for ML-1M. For DBbook, this adapter uses the reconstructed interaction release described by SWAP, not a claim of retrieving the original ESWC challenge service.

Paper: See the Movie, Hear the Song, Read the Book. Data: Zenodo record 15403972. Processing: authors’ repository.

@inproceedings{spillo2025multimodal,
  author = {Spillo, Giuseppe and Musacchio, Elio and Musto, Cataldo and
            de Gemmis, Marco and Lops, Pasquale and Semeraro, Giovanni},
  title = {See the Movie, Hear the Song, Read the Book: Extending
           MovieLens-1M, Last.fm-2K, and DBbook with Multimodal Data},
  booktitle = {Proceedings of the Nineteenth ACM Conference on Recommender Systems},
  year = {2025},
  doi = {10.1145/3705328.3748162}
}

@misc{spillo2025multimodaldata,
  author = {Spillo, Giuseppe and Musacchio, Elio and Musto, Cataldo and
            de Gemmis, Marco and Lops, Pasquale and Semeraro, Giovanni},
  title = {See the Movie, Hear the Song, Read the Book: Extending
           MovieLens-1M, Last.fm 2K, and DBBook with multimodal Data},
  year = {2025},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.15403972}
}

For Last.fm, the original README requests attribution to Last.fm and suggests the HetRec workshop citation below. See GroupLens HetRec 2011 and the supplied README for usage terms; availability in a public archive does not waive the original dataset restrictions.

@inproceedings{cantador2011hetrec,
  author = {Cantador, Iv{\'a}n and Brusilovsky, Peter and Kuflik, Tsvi},
  title = {2nd Workshop on Information Heterogeneity and Fusion in
           Recommender Systems (HetRec 2011)},
  booktitle = {Proceedings of the 5th ACM Conference on Recommender Systems},
  year = {2011},
  publisher = {ACM}
}