📚 Books Recommender System

An item-based collaborative filtering recommender built on the Book-Crossing ratings data. It finds books that the same readers rated similarly, using k-Nearest Neighbors. A Streamlit app returns five similar books with cover images. The notebook documents the full data preparation and model training.

Live demo: https://books-recommender-system-vineet416.streamlit.app/

Project Files

File	Description
app.py	Streamlit app: pick a book, get 5 similar books with covers
research.ipynb	Loading, cleaning, filtering, pivot table, model training, artifact export
artifacts/model.pkl	Fitted NearestNeighbors model (brute force, Euclidean)
artifacts/book_pivot.pkl	Book × user rating matrix (742 × 888)
artifacts/book_names.pkl	The 742 titles shown in the dropdown
artifacts/final_rating.pkl	Cleaned ratings joined with book metadata (59,850 rows)
requirements.txt	Python dependencies (unpinned)
.devcontainer/devcontainer.json	Python 3.11 dev container that starts the app on port 8501
LICENSE	MIT license

Dataset

Source: Book-Crossing dataset, read from Datasets/ (not in this repo): BX-Books.csv (271,360 rows), BX-Users.csv (278,858), BX-Book-Ratings.csv (1,149,780). Modeling table final_rating: 59,850 rows × 9 columns, no missing values.

Column	Description
user_id	Anonymous reader ID (888 unique)
ISBN	Edition ID (2,176 unique)
rating	0–10; 0 means no explicit rating
title / author	Book title (742 unique) and author (584 unique)
year / publisher	Publication year (text) and publisher (357 unique)
image_url	Large cover image URL
number_of_ratings	Title's rating count before de-duplication (50–363)

Approach

1. Understand the data: load the three CSVs, check shapes, rename columns, drop small and medium image columns.
2. Filter users: keep users with more than 200 ratings (899 of 105,283 raters) → 526,356 ratings.
3. Join and filter books: merge ratings with books on ISBN, keep titles with at least 50 ratings, drop duplicate (user, title) pairs.
4. Build the matrix: pivot to title × user, fill missing with 0, convert to a sparse CSR matrix.
5. Train: NearestNeighbors(algorithm='brute'), k = 6 (the book itself + 5 neighbors).
6. Serve: pickle four artifacts; app.py looks up neighbors and fetches cover URLs.

Key Insights

Recomputed from the artifacts (the raw CSVs are not available).

Finding	Value
Data kept after filters	59,850 of 1,149,780 ratings (5.2%)
Explicit ratings (1–10)	14,961 (25.0%); the other 75.0% are zeros. Matrix density is 2.27%
Explicit rating average / median	7.96 / 8; 66.7% of ratings are 8–10, only 2.9% are 1–4
Most explicitly rated	The Lovely Bones (104, avg 8.32), Bridget Jones's Diary (84, 7.49), The Da Vinci Code (82, 8.59)
Highest average (≥ 20 ratings)	The Two Towers 9.74, The Return of the King 9.57, The Secret Garden 9.46
Most rows, few real ratings	Wild Animus: 363 rows, only 38 explicit, avg 4.18
Example output	Harry Potter and the Chamber of Secrets → Books 4, 3, 1, then Exclusive, The Cradle Will Fall
"Hub" books	Exclusive is in 579 of 742 top-5 lists, No Safe Place in 515 (3 explicit ratings each); 434 books (58.5%) are never recommended

Tech Stack

• Language and data: Python 3.11, pandas, NumPy, SciPy (sparse CSR matrix)
• Model: scikit-learn NearestNeighbors (item-based collaborative filtering)
• App and tooling: Streamlit, pickle, Jupyter Notebook, VS Code Dev Container

How to Use

1. Install dependencies: pip install -r requirements.txt
2. Start the app: streamlit run app.py (opens at http://localhost:8501)
3. Choose a book, then click Show Recommendation.
4. To retrain, place the three CSVs in Datasets/ and run all cells in research.ipynb.