Learning-Resource Recommender 📚

A lightweight hybrid recommender that blends collaborative filtering with content-based signals to surface personalised learning resources — with keyword explanations for every suggestion.

---

How It Works

Three signals, weighted and blended:

Signal	Model	What it captures
Collaborative	ALS (implicit feedback)	Patterns from similar users
Collaborative	LogisticMF (implicit)	Second opinion on user-item affinity
Content	TF-IDF cosine similarity	Keyword overlap with user history

Each recommendation includes an explanation: the top TF-IDF terms connecting the item to the user's history.

---

Results (sample dataset)

Metric	Score
MAP@10	1.000
NDCG@10	1.000
Recall@10	1.000

Evaluated on a 5-user × 8-item sample with leave-one-out holdout.

---

Screenshot

---

Quickstart

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Browse recommendations in the UI
streamlit run app.py

# Run tests
PYTHONPATH=. pytest tests/ -v

---

Project Structure

src/
├── data.py         — load CSVs, build sparse interaction matrix + TF-IDF content matrix
├── ids.py          — user/item ↔ index mappings
├── als_train.py    — ALS model (implicit library)
├── lmf_train.py    — LogisticMF model (implicit library)
├── hybrid.py       — scoring functions + keyword explainer
└── metrics.py      — P@K, R@K, AP@K, MAP@K, NDCG@K
tests/
├── test_data.py    — data loading, matrix shapes, mappings
├── test_hybrid.py  — scoring functions, explainer
└── test_metrics.py — ranking metric correctness
data/sample/
├── items.csv       — item_id, title, genre flags
└── ratings.csv     — user_id, item_id, rating, timestamp

---

Data Format

items.csv

item_id,title,Comedy,Drama,Action,...
1,Intro to Linear Algebra,0,1,0,...

ratings.csv

user_id,item_id,rating,ts
1,1,1,1000

Swap in any implicit-feedback dataset that follows this schema (e.g. MovieLens 100K).