Semantic Book Recommender System

An end-to-end AI-powered book recommendation system that combines semantic search, zero-shot classification, and emotion-based filtering to provide personalized book suggestions.

---

Project Overview

This project builds a content-based recommendation system using modern NLP techniques. Instead of relying on keywords, it understands the meaning (semantics) of book descriptions.

Users can:

• Describe a book they want
• Filter by category (Fiction / Nonfiction)
• Filter by emotional tone (Happy, Sad, Suspenseful, etc.)
• Receive visually rich recommendations with thumbnails

📂 Project Pipeline

1️⃣ Data Collection

• Dataset: 7k Books with Metadata

• Fields used:

  • Title
  • Authors
  • Description
  • Rating
  • Categories
  • Thumbnail

---

2️⃣ Data Cleaning & Preprocessing

• Removed missing values:

  • description
  • num_pages
  • average_rating
  • published_year

• Filtered low-quality data:

  • Removed books with descriptions < 25 words

• Created new features:

  • title_and_subtitle
  • tagged_description (ISBN + description)

• Handled missing thumbnails:

books["large_thumbnail"] = np.where(
    books["thumbnail"].isna(),
    "cover-not-found.jpg",
    books["thumbnail"]
)

---

3️⃣ Exploratory Data Analysis (EDA)

• Missing value visualization (heatmaps)
• Correlation analysis (Spearman)
• Description length analysis
• Category distribution

---

4️⃣ Zero-Shot Classification

Used Hugging Face model:

facebook/bart-large-mnli

Purpose:

• Convert messy categories → simple labels:

  • Fiction
  • Nonfiction

Process:

1. Evaluate model using 300 Fiction + 300 Nonfiction samples
2. Compute accuracy
3. Predict missing categories
4. Fill missing values in dataset

---

5️⃣ Emotion Analysis

Each book is tagged with emotional scores:

• Joy
• Sadness
• Anger
• Fear
• Surprise

Used for tone-based filtering in recommendations.

---

6️⃣ Semantic Embeddings

Used local model:

sentence-transformers/all-MiniLM-L6-v2

Why?

• Fast
• Lightweight
• No API cost
• Optimized for semantic similarity

---

7️⃣ Vector Database (Chroma)

• Stored embeddings for all books
• Enabled fast similarity search
• Used persistent storage to avoid recomputation

---

8️⃣ Recommendation Logic

Step-by-step:

1. User inputs query
2. Convert query → embedding
3. Retrieve top-k similar books
4. Filter by category (optional)
5. Rank by emotional tone (optional)
6. Return top results

---

9️⃣ User Interface (Gradio)

Built an interactive dashboard:

• Text input for query

• Dropdown for category

• Dropdown for emotional tone

• Gallery output with:

  • Book cover
  • Title
  • Authors
  • Short description

▶️ Run the App

python main.py

---

📸 Example Query

"A story about love and forgiveness"

👉 Output:

• Semantically similar books
• Filtered by category
• Sorted by emotional tone

---