Book Recommendation Engine

A machine learning project that recommends books similar to a given title using k-NN (k-nearest neighbours) with cosine similarity.

The engine is trained on the Book-Crossings dataset, which contains thousands of user ratings on books, and suggests titles based on books that have been liked by similar users.

---

Description

The project uses a collaborative filtering recommender built from scratch using Python. It:

• Processes and filters the Book-Crossings dataset to keep active users and popular books.
• Builds a book-user matrix where each row is a book, and each column is a user.
• Calculates the cosine similarity between books to find those with similar preferences.
• Uses k-NN to recommend books similar to a given title.
• Includes evaluation using Recall@K to measure how often the model correctly recommends books a user may like.

---

How it Works

1. Loading the Data: The Book-Crossings dataset is downloaded and loaded.
2. Filtering: Removes all users with less than min_user_ratings ratings, and all books with less than min_book_ratings ratings.
3. Building the Matrix: Creates a binary book x user matrix, where 1 represents a user rating a book positively, and 0 represents otherwise. It may also apply TF-IDF like weighting to try to reduce popularity bias.
4. Training: Fits a k-NN model using cosine similarity on the matrix.
5. Evaluation: Splits data by holding out 1 positive rating per user, and measuring how often the held out book appears in the top K neighbours (Recall@K).
6. Recommendation: Given a title, finds the most similar books based on the patterns of the other readers.

---

Dataset

As mentioned, we use the Book-Crossings dataset. It contains:

• ~270,000 books.
• ~78,000 users.
• ~1,000,000 ratings.

Within, it contains three CSV files:

• BX-Books.csv, containing ISBN, title, and author.
• BX-Book-Ratings.csv, containing user ID, ISBN, and rating.
• BX-Users.csv, containing user details. This file is not used.

---

Usage

1. Clone the repo.

git clone https://github.com/cadhann/book_recommendation_engine.git
cd book_recommendation_engine

2. Set up environment.

python -m venv .venv
# On Powershell
.venv\Scripts\Activate.ps1
# In Command Prompt
.venv\Activate.bat
# In Bash
source .venv/bin/activate
pip install -r requirements.txt

3. Train the recommender.

python -m scripts.train --min-user-ratings 200 --min-book-ratings 100 --k 5

This will:

• Download & preprocess the data.
• Build the item x user matrix.
• Train the k-NN model.
• Evaluate it using Recall@K.
• Save the model & matrix to artifacts/

Output example:

Downloading data (if required)...
Loading CSVs...
Filtering active users and popular books...
Building item-user matrix (titles x users)...
Training k-NN model...
Evaluating...
Recall@5: 0.462
Saving artifacts...
Done.

4. Generate recommendations.

When trained, similar books can be queried.

python -m scripts.recommend --title "Animal Farm" --top-k 5

Output example:

Loading artifacts...
Base title: Animal Farm
1) 1984 (distance: 0.70)
2) The Corrections: A Novel (distance: 0.75)
3) To Kill a Mockingbird (distance: 0.83)
4) Lord of the Flies (distance: 0.83)
5) Shopgirl : A Novella (distance: 0.83)

The lower the distance, the more similar the book. Here, we can see that 1984 (by the same author as Animal Farm, George Orwell) is matched, along with To Kill a Mockingbird and Lord of the Flies, two novels with similar themes to Animal Farm.