Datasets:

deem-data
/

erase-bench

This document lists the datasets under RecBole/dataset/ used in this project: download links (no RecBole auto-download) and how to build each from raw using scripts in this repository.

Summary

Dataset (folder)	Task	Main file	Raw source
30music	SBR	`30music.inter`	ReMAP Lab (link below)
nowp	SBR	`nowp.inter`	Zenodo
amazon_reviews_books	SBR	`amazon_reviews_books.inter`	Amazon Reviews (Books)
amazon_reviews_grocery_and_gourmet_food	CF	`amazon_reviews_grocery_and_gourmet_food.inter`	Amazon Reviews (Grocery)
movielens	CF	`movielens.inter`	MovieLens (Kaggle/GroupLens)
rsc15	SBR	`rsc15.inter`	RecSys Challenge 2015 (Kaggle)
tafeng	NBR	`tafeng_merged.json`	Ta-Feng (Kaggle)
dunnhumby	NBR	`dunnhumby_merged.json`	dunnhumby / Kaggle
instacart	NBR	`instacart_merged.json`	Instacart / Kaggle

1. 30music

Folder: dataset/30music/
Task: Sequential recommendation (SBR)
Main file: 30music.inter
Header: user_id:token session_id:token item_id:token timestamp:float

Download

Source: ReMAP Lab, Politecnico di Milano (30Music listening and playlists dataset, RecSys 2015).
Link: 30Music dataset (SharePoint)
Citation: Turrin, R., Quadrana, M., Condorelli, A., Pagano, R., & Cremonesi, P. "30Music listening and playlists dataset", RecSys 2015.

Building from raw (this repo)

This repo does not contain a conversion script from 30music raw files to 30music.inter. You need to produce a tab-separated .inter with the header above (one row per user/session/item/timestamp) and save it as dataset/30music/30music.inter.
After the main file exists, forget sets can be created with:
- dataset/30music/create_forget_sets.py (requires 30music.inter and tracks/tags files as in the script).

2. NowP (NowPlaying)

Folder: dataset/nowp/
Task: SBR
Main file: nowp.inter
Header: user_id:token session_id:token item_id:token timestamp:float

Download

Source: NowP (Zenodo) — music listening dataset.
Place the CSV that contains session-level data in dataset/nowp/. The script expects sessions_2018.csv with columns including user_id, session_id, timestamp, and an item identifier (e.g. musicbrainz_id).

Building from raw (this repo)

Download from Zenodo and put sessions_2018.csv in dataset/nowp/.
From dataset/nowp/ run:
```
python preprocess_nowp.py
```
- Step 1: Reads sessions_2018.csv, writes nowp_temp.inter with columns user_id, session_id, item_id (musicbrainz_id), timestamp.
- Step 2: Sorts and deduplicates (external sort) into nowp.inter.
- Step 3: Removes the temp file.
Output: dataset/nowp/nowp.inter in RecBole atomic format.

Amazon Reviews (shared source for Books and Grocery)

Both amazon_reviews_books and amazon_reviews_grocery_and_gourmet_food use the same data family; only the category (and thus the files/folder) differs.

Download (one source, pick your category)

Source: Amazon product review. Provides per-category review data as *.jsonl.gz (and optionally meta_*.jsonl.gz for metadata).
Format: Each review JSONL line has user_id, asin (product id), rating, timestamp. Scripts ignore meta_* files and use only the review *.jsonl.gz for the chosen category.
Put the category’s review files in the matching dataset folder: dataset/amazon_reviews_books/ or dataset/amazon_reviews_grocery_and_gourmet_food/.

3. Books

Folder: dataset/amazon_reviews_books/
Task: SBR (with sessions and rating)
Main file: amazon_reviews_books.inter
Header: user_id:token session_id:token item_id:token rating:float timestamp:float

Download: Use the Books category from the Amazon Reviews source above. Place the Books review *.jsonl.gz (and optionally meta_Books.jsonl.gz for sensitive-item scripts) in dataset/amazon_reviews_books/.

Building from raw (this repo)

Step 1 – Raw .inter (no sessions):
From dataset/amazon_reviews_books/:
```
python convert_to_inter.py [--output_file amazon_reviews_books_raw.inter] [--files file1.jsonl.gz file2.jsonl.gz ...]
```
Reads all non-meta .jsonl.gz files, extracts user_id, asin, rating, timestamp, writes amazon_reviews_books_raw.inter (tab-separated, no type suffixes).
Step 2 – Deduplicate and filter:
Same directory:
```
python deduplicate.py --input amazon_reviews_books_raw.inter --output amazon_reviews_books_clean.inter
```
Writes RecBole-style header and deduplicated, sorted data; keeps only users/items with at least 5 interactions (default).
Step 3 – Add sessions:
Same directory:
```
python add_sessions.py --input_file amazon_reviews_books_clean.inter --output_file amazon_reviews_books.inter --time_window_hours 1.0
```
Groups by user and time window (default 1 hour), assigns session_id, writes amazon_reviews_books.inter with the full header above.
Optional: sensitive items and forget sets use identify_sensitive_items.py and generate_forget_sets.py (see scripts’ help and meta_Books.jsonl.gz for metadata).

4. Food

Folder: dataset/amazon_reviews_grocery_and_gourmet_food/
Task: CF (no sessions)
Main file: amazon_reviews_grocery_and_gourmet_food.inter
Header: user_id:token item_id:token rating:float timestamp:float

Download: Use the Grocery and Gourmet Food category from the Amazon Reviews source above. Place that category’s review *.jsonl.gz in dataset/amazon_reviews_grocery_and_gourmet_food/.

Building from raw (this repo)

Step 1 – Raw .inter:
From dataset/amazon_reviews_grocery_and_gourmet_food/:
```
python preprocess_data.py --output_file amazon_reviews_grocery_and_gourmet_food_raw.inter [--files ...]
```
Reads all non-meta .jsonl.gz in the directory (or the files you pass), extracts user_id, asin (as item_id), rating, timestamp; writes a tab-separated file with a simple header. Default: all *.jsonl.gz except meta_*.
Step 2 – Deduplicate:
Same directory:
```
python deduplicate.py --input amazon_reviews_grocery_and_gourmet_food_raw.inter --output amazon_reviews_grocery_and_gourmet_food.inter
```
Writes the RecBole header and deduplicated/filtered data (min 5 interactions per user/item by default). Result is the main CF .inter (no session_id).

5. MovieLens

Folder: dataset/movielens/
Task: CF
Main file: movielens.inter
Header: user_id:token item_id:token rating:float timestamp:float

Download

Source: MovieLens 20M (Kaggle) — use rating.csv. Place it in dataset/movielens/.

Building from raw (this repo)

Step 1 – Raw .inter:
From dataset/movielens/:
```
python preprocess_movielens.py --input-file rating.csv --output-file movielens_raw.inter
```
Reads userId, movieId, rating, timestamp; converts timestamp to Unix; writes tab-separated movielens_raw.inter (with a simple header line).
Step 2 – Deduplicate and filter:
Same directory:
```
python deduplicate.py --input movielens_raw.inter --output movielens.inter
```
Writes RecBole header and deduplicated/filtered data (min 5 interactions per user/item by default). Output: movielens.inter.
Optional: sensitive movies and forget sets use identify_sensitive_movies.py and generate_forget_sets.py in the same folder.

6. RSC15 (RecSys Challenge 2015)

Folder: dataset/rsc15/
Task: SBR (session-based; session = user_id in some configs)
Main file: rsc15.inter
Header: session_id:token item_id:token timestamp:float (or in some scripts user_id:token is used for session id).

Download

Source: RSC15 (Kaggle) — RecSys Challenge 2015 / YOOCHOOSE-style session data.

Building from raw (this repo)

Convert the challenge CSV into two RecBole .inter files with header session_id:token item_id:token timestamp:float (or user_id:token item_id:token timestamp:float if your pipeline uses user_id for session):
- rsc15.train.inter
- rsc15.test.inter
  One row per (session, item, timestamp). How you split train/test is up to you (e.g. by time or as in the challenge).
Merge train + test:
From dataset/rsc15/:
```
python merge_train_and_test.py
```
Reads rsc15.train.inter and rsc15.test.inter, concatenates, sorts by first column and timestamp, writes rsc15.inter. Note: The script currently writes header user_id:token item_id:token timestamp:float. If your SBR config expects session_id:token, either rename the column in the script or ensure your config treats that column as session id.
For train/val/test splits from raw CSV, see dataset/rsc15/train_val_test_unlearn_split_v2.py (reads CSV with session_id, item_id, timestamp and produces splits).

7. TaFeng (NBR)

Folder: dataset/tafeng/
Task: Next-basket recommendation (NBR)
Main file: tafeng_merged.json (not .inter)

Download

Source: TaFeng (Kaggle) — use tafeng_all_months_merged.csv.zip.

Building from raw (this repo)

This repo does not include a script that converts Ta-Feng CSV → tafeng_merged.json. You must build the JSON yourself (or adapt the dunnhumby script logic).
Required JSON format: One object: user_id (string) → list of baskets. Each basket = list of item IDs (ints or strings), in chronological order. Example:
```
{ "user_1": [[1, 2, 3], [4, 5]], "user_2": [[10, 20]] }
```
Save as dataset/tafeng/tafeng_merged.json. NBR models expect at least 4 baskets per user (configurable). After that, fraud baskets (for poisoning experiments) can be generated with dataset/create_fraud_baskets_nbr.py.

8. Dunnhumby (NBR)

Folder: dataset/dunnhumby/
Task: NBR
Main file: dunnhumby_merged.json

Download

Source: dunnhumby source files or dunnhumby The Complete Journey (Kaggle) — use transaction_data.csv. Place it in dataset/dunnhumby/.

Building from raw (this repo)

From dataset/dunnhumby/:
```
python create_dunnhumby_merged_v2.py
```
- Reads transaction_data.csv.
- Mapping: household_key → user_id, BASKET_ID → basket, PRODUCT_ID → item_id, TRANS_TIME → ordering.
- Groups by household, then by basket; sorts baskets by time; outputs list of product lists per user.
- Keeps only users with at least 4 baskets.
Script writes dunnhumby_merged_v2.json. To use it as the main NBR file, either rename/copy to dunnhumby_merged.json or point config to dunnhumby_merged_v2.json if supported.
Fraud baskets: dataset/create_fraud_baskets_nbr.py --dataset dunnhumby (reads dunnhumby_merged.json or the path you use).

9. Instacart (NBR)

Folder: dataset/instacart/
Task: NBR
Main file: instacart_merged.json

Download

Source: Instacart Market Basket Analysis (Kaggle). If the official data is unavailable: instacart-orders (GitHub).

Building from raw (this repo)

This repo does not include a script that converts Instacart CSV/tables → instacart_merged.json. Build the JSON as for tafeng/dunnhumby.
Required JSON format: Same as tafeng: user_id (string) → list of baskets, each basket = list of item IDs, chronological. Save as dataset/instacart/instacart_merged.json. Minimum baskets per user (e.g. 4) is enforced by the NBR data loader. Then you can run dataset/create_fraud_baskets_nbr.py --dataset instacart.

RecBole atomic format (quick reference)

Separator: Tab between columns.
Header: First line = column names with type suffix, e.g. user_id:token, item_id:token, rating:float, timestamp:float, session_id:token.
File naming: Main interaction file is {dataset_name}.inter in dataset/{dataset_name}/.
NBR: Uses {dataset_name}_merged.json (user → list of baskets), not .inter, when using NextBasketDataset.

For more on RecBole data format, see the official docs (e.g. atomic files and dataset list).