Hybrid Recommendation Engine

The problem

Collaborative filtering has nothing to say about a user or item it has never seen.

Approach

Collaborative filtering, content-based filtering and SVD matrix factorisation blended at a fixed 0.4/0.3/0.3 to address the cold start problem, and evaluated on catalog coverage, category diversity and average rating rather than raw accuracy.

How it works

• Collaborative filtering, content-based filtering and SVD matrix factorisation each produce candidate recommendations.
• A hybrid layer blends them at a fixed 0.4 collaborative, 0.3 content-based, 0.3 SVD, which is what covers the cold start neither pure approach solves alone.
• Quality is judged on catalog coverage, category diversity and the average rating of what gets recommended, rather than on raw accuracy.
• A Streamlit dashboard lets you inspect recommendations per user.

Key decisions

Hybrid specifically for cold start

Collaborative filtering knows nothing about a new item and content-based filtering only ever suggests more of the same. Combining them covers each other's blind spot rather than averaging two mediocre answers.

Judge the list, not the prediction

Accuracy on a held-out rating says little about a short ranked list someone actually sees. Coverage, diversity and the average rating of served recommendations correspond to the product surface instead.

What the measurements showed

• 3,000 users, 500 products across 7 categories, 129,782 ratings, 91.35% sparsity.
• Hybrid score is a fixed blend: 0.4 collaborative, 0.3 content-based, 0.3 SVD.
• Recommendations average above 4.0 of 5.0 on rating, with coverage spread across categories rather than concentrated in popular items.

Running it

Full setup, configuration and API reference are in the repository README.

Stack

• Python
• scikit-learn
• pandas
• NumPy
• SVD

Written up in

• Building a Hybrid Recommendation SystemDataDrivenInvestor