A Worrying Analysis of Neural Recommendations, Now Executable
This repository wraps dozens of conference-published deep learning recommenders in a unified framework to test whether they actually beat well-tuned classical baselines like KNN and matrix factorization.

What it does
This repository bundles implementations of neural recommendation algorithms from top-tier conferences into a single evaluation harness. It pits them against a comprehensive suite of older methods—ItemKNN, SLIM_BPR, EASE_R, BPRMF, and others—each subjected to Bayesian hyperparameter optimization. The framework automates data loading, splitting, metric computation, and result tabulation to enable fair, reproducible comparisons.
The interesting bit
Rather than reimplementing papers from scratch, the authors wrapped the original authors’ code—stored in _github and _original folders—with minor fixes and unified Wrapper classes. That design choice intentionally preserves the original implementations while eliminating interface bias, making the framework a reproducibility engine built around a skeptical question: do these neural models actually advance the state of the art?
Key highlights
- Wraps original deep learning recommender code from RecSys, CIKM, and IJCAI with unified
Wrapperinterfaces - Baseline arsenal includes
UserKNN,ItemKNN,PureSVD,FunkSVD,BPRMF,SLIM_ElasticNet,EASE_R, and hybrid variants - Standardized evaluation via
EvaluatorHoldoutandEvaluatorNegativeSamplewith automated metric computation DataReaderobjects handle dataset acquisition and preprocessing, caching splits inData_manager_split_datasets- Bayesian hyperparameter search via
SearchBayesianSkoptusing fixed search spaces across all experiments - Ships with full experimental results, hyperparameters, and LaTeX table generation for four published studies
Caveats
- Some datasets, such as ConvNCF, cannot be downloaded automatically and require manual placement
- The codebase targets Python 3.6 and requires compiling Cython extensions, which may friction modern setups
- A post-hoc correction to HR and NDCG calculations under random holdout affects only a few results; the authors note the overall conclusions stand
Verdict Essential reading—and running—for anyone publishing neural recommender research who hasn’t seriously benchmarked against tuned older methods. Skip it if you are only looking for a lightweight, drop-in recommendation library.
Frequently asked
- What is MaurizioFD/RecSys2019_DeepLearning_Evaluation?
- This repository wraps dozens of conference-published deep learning recommenders in a unified framework to test whether they actually beat well-tuned classical baselines like KNN and matrix factorization.
- Is RecSys2019_DeepLearning_Evaluation open source?
- Yes — MaurizioFD/RecSys2019_DeepLearning_Evaluation is open source, released under the AGPL-3.0 license.
- What language is RecSys2019_DeepLearning_Evaluation written in?
- MaurizioFD/RecSys2019_DeepLearning_Evaluation is primarily written in Python.
- How popular is RecSys2019_DeepLearning_Evaluation?
- MaurizioFD/RecSys2019_DeepLearning_Evaluation has 991 stars on GitHub.
- Where can I find RecSys2019_DeepLearning_Evaluation?
- MaurizioFD/RecSys2019_DeepLearning_Evaluation is on GitHub at https://github.com/MaurizioFD/RecSys2019_DeepLearning_Evaluation.