William-Huang274/Yelp-Offline-Recommendation-Pipeline
Offline Yelp recommender for Louisiana restaurant discovery, with candidate fusion, XGBoost rerank, QLoRA/DPO sidecar experiments, and release validation.
What's novel
Offline Yelp recommender for Louisiana restaurant discovery, with candidate fusion, XGBoost rerank, QLoRA/DPO sidecar experiments, and release validation.
Code Analysis
16 files read · 6 roundsA rigorous offline restaurant recommendation pipeline that fuses multi-source candidates (popularity, ALS, graph, BGE semantic) into a frozen pool, then reranks the top-50 with a custom DCN-V2 mixture-of-experts model trained on structured, NLP, and sequence features, evaluated via a one-shot full-d
Strengths
Exceptional evaluation rigor (full-denominator, no target injection, 15-model OOF ablation, paired bootstrap with exact p-values, frozen test dispatch guard) combined with genuinely non-trivial model implementations (MoE cross layers, residual dual encoder with in-batch contrastive masking, pointer-copy Transformer with learned repeat/explore gate). The DuckDB-based feature assembly and memory-mapped training loops show production-grade engineering discipline.
Weaknesses
The test suite is limited to contract/surface validation (manifest structure, metric consistency, dispatch safety) with no unit tests of the core algorithms or edge-case coverage of the model code. The 64 numbered scripts in a flat directory, while well-named, create a monolithic procedural pipeline where cross-script dependencies are implicit (e.g., 167 imports 163 via importlib) rather than through a shared library.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
9
Contributors
1
Files
185
Active weeks
2
Repository
Language
Python
Stars
1
Forks
0
License
MIT