Changelog

Changelog#

All notable changes to CoreRec will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.5.4] - 2026-06-05#

Added#

  • Model zoo expanded 14 -> 41 production models, all conforming to the unified contract (fit/predict/recommend/save/load + collapse guard), built on shared GPU-capable bases and covered by 112 CI-gated contract tests:

    • Deep CTR / feature-interaction (12): FM, AFM, NFM, DeepFM, DCN, AutoInt, xDeepFM, FiBiNet, PNN, WideDeep, GMF, MLP (corerec.engines.deep_ctr).

    • Sequential / session (6): GRU4Rec, Caser, BST, DIN, DIEN, NARM (corerec.engines.sequential).

    • Classic CF (4): ItemKNN, UserKNN, EASE, SLIM (corerec.engines.classic_cf).

    • Auto-encoder CF (2): MultVAE, MultiDAE (corerec.engines.vae_cf).

    • Graph + native MF (3): NGCF (corerec.engines.graph_cf), ALS/WMF and Item2Vec (corerec.engines.matrix_factorization).

    • With the larger zoo CoreRec’s best model leads three of the four standard benchmarks (ML-100K via SLIM/UserKNN, Gowalla and Yelp2018 via LightGCN).

  • corerec.serving.OnlineRecommender — ANN-backed online serving engine (sub-millisecond retrieval at million-item scale via FAISS HNSW/IVF), with incremental add_items / fold_in_user (freshness without retraining) and a graceful popularity fallback for cold-start users. See the Online Serving guide.

  • Native GPU sparse trainers: an efficient sparse LightGCN and a strengthened BPR (OnlineRecommender.from_interactions(df, model="lightgcn"|"bpr")). CoreRec’s LightGCN ranks #1 on the Gowalla and Yelp2018 large-scale graph benchmarks.

  • corerec.evaluation.evaluate(model, test, train, k=...) — one-call standard top-K evaluation (NDCG/MAP/MRR/Precision/Recall/HitRate@K).

  • BaseRecommender.card() — reproducibility/audit metadata (params + versions).

  • task argument (implicit/rating) with negative sampling for DCN, DeepFM, GNNRec.

Fixed#

  • Deep-model output collapse: DCN/DeepFM/GNNRec trained on observed-only data (all-positive labels) collapsed to a constant. Negative sampling + the task contract restore them to parity with reference implementations (e.g. DCN NDCG@10 0.005 → 0.296 on ML-100K). A post-fit guard warns on score collapse.

  • Vectorized full-ranking inference for deep models (recommend latency ~236 ms → ~1.5 ms/user); removed NCF’s per-row iterrows data prep.

  • VectorIndex now supports inner-product HNSW, so dot-product models (MF/ALS) are served correctly rather than distorted by cosine normalization.

  • Lazy torch import in corerec.utils.seed — pure collaborative-filtering usage (e.g. SAR) no longer pulls in PyTorch (~754 MB → ~124 MB resident).

  • task="rating" models now round-trip correctly through save/load.

[0.5.3] - 2026-03-31#

Added#

  • Safe bundle persistence (corerec_safe_v1) for all 14 production models

  • Production platform tests (retrieval, ranking, reranking, serving, pipelines)

  • Model catalog documentation pages

Fixed#

  • Safe bundle ID map round-trip for integer keys

  • Predict score parity on save/load for all production models

  • Documentation alignment (QuickStart, tutorials, top_k, artifact paths)

Changed#

  • Unified recommend(..., top_k=) API across production models

[Unreleased - historical]#

Added#

  • LICENSE file (MIT License)

  • Proper packaging configuration (pyproject.toml, setup.py in root)

  • Comprehensive requirements files (requirements.txt, requirements-dev.txt, requirements-test.txt)

  • CI/CD workflows using GitHub Actions (test.yml, lint.yml, docs.yml)

  • CONTRIBUTING.md with detailed contribution guidelines

  • Pre-commit hooks configuration (.pre-commit-config.yaml)

  • Improved .gitignore to exclude build artifacts

  • CRITICAL_ISSUES_ANALYSIS.md - comprehensive codebase analysis

  • QUICKSTART_FIXES.md - guide for fixing critical issues

  • CHANGELOG.md (this file)

  • CLI tool with tab completion (corerec command)

Changed#

  • Moved setup.py from production/ to root directory

  • Updated .gitignore to follow Python best practices

  • Improved package structure for better import consistency

Fixed#

  • Build artifacts (*.o, *.pyc, pycache) cleanup from repository

  • Package installation issues (setup.py now in correct location)

Security#

  • Added Bandit security scanning to CI/CD

  • Added Safety dependency vulnerability scanning

[0.5.1] - 2024-10-XX#

Added#

  • Multiple recommendation engines (DCN, DeepFM, GNNRec, MIND, NASRec, SASRec)

  • Unionized Filter Engine (collaborative filtering algorithms)

  • Content Filter Engine (content-based filtering algorithms)

  • Comprehensive API with BaseRecommender base class

  • Extensive examples directory

Known Issues#

  • Two base classes exist (BaseRecommender and BaseCorerec) - consolidation needed

  • Circular import risks in core_rec.py

  • Incomplete type hints

  • C++ extensions not properly configured in setup

  • Missing comprehensive test coverage reports

[0.5.0] - 2024-09-XX#

Added#

  • Initial release of CoreRec

  • Basic recommendation models

  • Core API structure

For detailed changes, see the GitHub Releases page.

Contents