Recommendation Engine

Collaborative filtering over a user–post likes table: factorise the interaction matrix, find similar users, and recommend what they liked.

Served as a small Flask endpoint.

---

How it works

Recommender pivots the likes table into a user–item matrix, then reduces it with truncated SVD to 10 latent features:

self.svd = TruncatedSVD(n_components=10)
user_item_matrix_svd = self.svd.fit_transform(self.user_item_matrix)
self.user_similarity = cosine_similarity(user_item_matrix_svd)

Similarity is computed in the reduced space rather than on raw likes. Two users who like different individual posts within the same underlying topic still come out similar, which is the point of the factorisation — raw cosine similarity on a sparse matrix would call them unrelated.

Recommendations are then a similarity-weighted sum over what everyone else liked, normalised by total similarity, with the top N returned.

Data

Expects a CSV at data/likes.csv (not committed) with one row per interaction:

user_id,post_id,like
1,101,1
1,104,1
2,101,1

pivot_table fills unobserved pairs with 0.

Running it

pip install -r requirements.txt
python app.py

curl "http://localhost:5000/recommend?user_id=1&top_n=5"

{ "recommended_posts": [104, 92, 17, 55, 3] }

An unknown user_id returns 400 — this is pure collaborative filtering, so a user with no interaction history cannot be served. Handling that cold start would mean falling back to popularity or content-based ranking.

Layout

app.py               Flask route
src/recommender.py   matrix construction, SVD, similarity, ranking
src/main.py          CLI entry point

Limitations

The model is fit once at construction from the whole CSV, so new likes need a restart to take effect, and the user–item matrix is held entirely in memory. Fine at small scale; a real deployment would want incremental updates and a stored factorisation.