Netflix catalog: EDA → sparse retrieval → dense / hybrid / rerank

Author: Nikolaos (Nikos) Mavrapidis (NikosMav)

Two showcase tracks in one repo:

1. EDA + Boolean / TF-IDF — original 2023 case study, cleaned and runnable (netflix_data_analysis.ipynb)
2. Catalog retrieval product — Boolean, TF-IDF, BM25, dense MiniLM, hybrid RRF, and CPU cross-encoder rerank with honest offline metrics (RETRIEVAL.md, python -m retrieval)

This is catalog text search. It is not a production recommender, not collaborative filtering, not TESSI, and not RAG-over-the-web.

5-minute story

Step	What	Where
Explore the catalog	Cleaning + charts	netflix_data_analysis.ipynb
Sparse retrieval	Boolean / TF-IDF / BM25	notebook + python -m retrieval
Dense / hybrid / rerank	Embed → fuse → CE rerank → eval	python -m retrieval, RETRIEVAL.md

Headline (same 28 hand-labeled queries — regenerate with python -m retrieval eval):

Baseline on main:

method	recall@5	recall@10	ndcg@5	ndcg@10	mrr
boolean	0.3159	0.4012	0.3185	0.3527	0.4440
tf-idf	0.4502	0.5446	0.4555	0.4912	0.5013
dense(title+desc)	0.5849	0.6209	0.5710	0.5656	0.6304
dense(title-only)	0.4241	0.4499	0.4286	0.4269	0.5081
hybrid(tfidf+dense)	0.5059	0.6922	0.4941	0.5626	0.5853

Extended ablations (BM25, metadata text, CE rerank):

method	recall@5	recall@10	ndcg@5	ndcg@10	mrr
bm25	0.5248	0.5645	0.5167	0.5253	0.5637
dense(title+desc+meta)	0.5735	0.6856	0.5825	0.6193	0.6786
hybrid(bm25+dense,meta)	0.6552	0.7421	0.6351	0.6605	0.7065
dense+rerank	0.6167	0.7062	0.6179	0.6469	0.6930
hybrid+rerank	0.6882	0.7627	0.6862	0.7053	0.7601

BM25 replaces Boolean as the serious lexical baseline. Metadata helps dense recall@10 / MRR more than sparse early precision. Cross-encoder rerank over top-50 gives the largest early-rank lift. Full table + interpretation: RETRIEVAL.md.

Clone and run

git clone https://github.com/NikosMav/DataAnalysis-Netflix.git
cd DataAnalysis-Netflix

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# --- Track 1: EDA + sparse TF-IDF notebook ---
jupyter notebook netflix_data_analysis.ipynb

# --- Track 2: retrieval product (no paid API, CPU default) ---
python -m retrieval query "war between vietnam and usa" --method bm25 --top-k 10
python -m retrieval query "feel-good cooking competition show" --method dense-rerank
python -m retrieval query "dark crime thriller set in Scandinavia" --method hybrid-rerank
python -m retrieval eval --failures   # regenerates results/*.json

# Optional walkthrough notebook
jupyter notebook netflix_dense_retrieval.ipynb

Runtime: sparse notebook needs a few GB RAM for pairwise matrices. Dense first run downloads MiniLM (~80MB) and embeds ~7.8k rows; rerank downloads ms-marco MiniLM CE (~80MB). Then caches under .cache/.

What’s in the box

Path	Purpose
netflix_data_analysis.ipynb	EDA + Boolean/TF-IDF case study (kept intact)
retrieval/	Package + CLI (query, eval, index)
data/labeled_queries.json	28 author-labeled queries with relevant_show_ids (unchanged)
results/eval_metrics.json	Last regenerated metric table
RETRIEVAL.md	Full case study: method, baseline + extended results, limits
netflix_dense_retrieval.ipynb	Thin package walkthrough

Data

Under data/: Netflix titles dump, slim IMDb join for the EDA chart, and retrieval labels. Provenance in data/README.md.

Limitations (read these)

• No invented metrics. Tables above come from python -m retrieval eval on the shipped labels.
• 28 queries, author-labeled. Enough for an honest demo, not a public IR leaderboard.
• Metadata ablation is uneven. Genre/cast/director/country help some methods and hurt others — see RETRIEVAL.md.
• Catalog search ≠ recommender ≠ TESSI ≠ web RAG.

License

MIT — see LICENSE.md. Netflix catalog © Netflix (public dump via Kaggle). IMDb data subject to IMDb non-commercial terms.