---
title: Bangla Book Recommender
emoji: 📚
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 5.45.0
app_file: app.py
pinned: false
license: cc-by-nc-4.0
short_description: Cold-start recommendations from 127K Bangla books
tags:
  - recommendation
  - bangla
  - lightgcn
  - two-tower
---

# 📚 Bangla Book Recommender

A cold-start book recommendation interface for the **RokomariBG** dataset —
the first large-scale multi-entity heterogeneous graph dataset for Bangla book
recommendation, built from publicly scraped reviews and metadata on
[Rokomari.com](https://rokomari.com), Bangladesh's largest online bookstore.

The Space lets a visitor pick a few books they have enjoyed and receive
recommendations from either of two benchmarked models from the paper:

| Model | Notes |
|---|---|
| **Neural Two-Tower** | Best benchmarked model. Item tower fuses ID, content (title, summary, author, publisher), and metadata (price, rating, pages); user tower combines a user-ID embedding with pooled history. Strongest at cold-start. Covers all 127K books. |
| **LightGCN** | Pure graph collaborative filtering with 4 GCN layers and uniform layer averaging. Trained on the 16K-book subset with sufficient interaction history. |

## How recommendation works (cold-start mode)

A new visitor has no `user_id` in the trained models. So instead of looking up
a stored user embedding, the Space derives one on the fly:

1. The visitor selects **N** books they have enjoyed (Bangla or English).
2. Each model has a precomputed **book embedding** for every book in its
   training catalogue.
3. The visitor's **taste vector** is computed as the L2-normalised mean of the
   selected books' embeddings.
4. **Cosine similarity** is computed between the taste vector and every book
   embedding. The seed books are masked out, and the top-K nearest books are
   returned.
5. Liked recommendations can be **promoted into the seed set**, after which
   re-running yields progressively more personalised results.

The pipeline is pure NumPy at inference time — no model is loaded at runtime,
only the precomputed embeddings — so the Space is fast and cheap to host on
the free CPU tier.

## File layout

All artefacts live in the Space root:

```
.
├── app.py                       # Gradio app
├── requirements.txt
├── README.md                    # this file
├── books_metadata.parquet       # title, author, category, rating, summary, url
├── two_tower_book_emb.npy       # (N, 256) float32, L2-normalisable
├── two_tower_book_ids.npy       # (N,)     book_id strings, row-aligned
├── lightgcn_book_emb.npy        # (16266, 64) float32, L2-normalisable
└── lightgcn_book_ids.npy        # (16266,)    book_id strings, row-aligned
```

## Citation

```bibtex
@misc{ahmed2026personalizedbanglabookrecommendation,
  title  = {Towards Personalized Bangla Book Recommendation:
            A Large-Scale Multi-Entity Book Graph Dataset},
  author = {Rahin Arefin Ahmed and Md. Anik Chowdhury and
            Sakil Ahmed Sheikh Reza and Devnil Bhattacharjee and
            Muhammad Abdullah Adnan and Nafis Sadeq},
  year   = {2026},
  eprint = {2602.12129},
  archivePrefix = {arXiv},
  primaryClass  = {cs.IR},
  url    = {https://arxiv.org/abs/2602.12129}
}
```

## Links

- 📄 Paper: <https://arxiv.org/abs/2602.12129>
- 💾 Dataset: <https://huggingface.co/datasets/DevnilMaster1/Bangla-Book-Recommendation-Dataset>
- 🐙 Code: <https://github.com/DevnilMaster/Bangla-Book-Recommendation-Dataset>

License: **CC BY-NC 4.0** (matches the dataset licence)