Machine Learning Book Recommender (Portfolio)

An end‑to‑end project exploring data cleaning, recommendation modeling, and deployment.

Project Note: This app uses the Book‑Crossing dataset as a starting point. The raw data contained significant noise and inconsistencies. I implemented extensive cleaning and enrichment (e.g., standardized metadata, subject mapping), but a small number of errors can still persist. The app exists to showcase the ML pipeline and engineering, not perfect catalog data.

What This Demonstrates

End‑to‑end ML build: data ingestion → cleaning → modeling → deployment.
Recommendation UX: cold‑start via subjects; warm personalization via ratings.
Evaluation in the wild: popularity ranking and similar‑book browsing to gather signal.
MLOps mindset: scheduled retraining on a separate server with automated deployment.
Performance & scale: low‑latency candidate retrieval and reranking.

How to Use the App

Create Your Profile: Sign up and select your favorite genres. These are used to generate your first set of recommendations.
Explore Initial Recommendations: Until you've rated at least 10 books, suggestions are mostly based on your favorite genres. You can update them anytime from your profile.
Browse Popular Books: Visit the Search page without entering a title to see books ranked by popularity (using a Bayesian formula). This is a great way to find books you’ve already read and rate them quickly.
Rate at Least 10 Books: Once you’ve rated 10 or more books, the system unlocks fully personalized recommendations based on your own preferences.
Keep Exploring and Rating: The more books you rate, the better your recommendations become. You can also find similar books on any book's detail page.

Behind the Scenes

Data pipeline: Book‑Crossing base with extensive cleaning/enrichment (standardized metadata, subject mapping).
Retrieval: implicit‑feedback ALS embeddings for fast user–item candidate generation.
Content similarity: attention‑pooled subject embeddings scored via cosine similarity.
ANN search: FAISS indexes over ALS and subject spaces for efficient top‑K lookups.
Popularity: Bayesian adjustment balances rating averages with rating counts.
Reranking: LightGBM merges metadata and decomposed embeddings to refine results.
Automation: scheduled training on a separate server; artifacts exported and hot‑reloaded without downtime.