---
pretty_name: MAGE multimodal recommendation datasets
language:
- en
tags:
- recommendation
- multimodal
- graph-learning
- arxiv:2610.37409
configs:
- config_name: baby
  default: true
  data_files:
  - split: train
    path: data/baby/train.parquet
  - split: validation
    path: data/baby/validation.parquet
  - split: test
    path: data/baby/test.parquet
- config_name: sports
  data_files:
  - split: train
    path: data/sports/train.parquet
  - split: validation
    path: data/sports/validation.parquet
  - split: test
    path: data/sports/test.parquet
- config_name: clothing
  data_files:
  - split: train
    path: data/clothing/train.parquet
  - split: validation
    path: data/clothing/validation.parquet
  - split: test
    path: data/clothing/test.parquet
- config_name: elec
  data_files:
  - split: train
    path: data/elec/train.parquet
  - split: validation
    path: data/elec/validation.parquet
  - split: test
    path: data/elec/test.parquet
---

# MAGE datasets

Prepared Baby, Sports, Clothing, and Electronics datasets for **Can Complementary Signals Bridge Similarity Islands? Manifold-Augmented Graph Embedding for Multimodal Recommendation**.

- [Paper](https://arxiv.org/abs/2610.37409)
- [Code and training instructions](https://github.com/Alex-bupt/mage)
- [Input format](https://github.com/Alex-bupt/mage/blob/main/docs/DATA_FORMAT.md)

The release contains fixed interaction splits, item and user ID mappings, visual and text feature matrices, an expanded text-feature catalog, and offline relation graphs. It provides numeric features rather than raw images or review text. These prepared Amazon-category benchmarks build on the data used by [GUME](https://github.com/Alex-bupt/mage/blob/main/docs/upstream/GUME_README.md) and MMRec.

## Contents

`archive/MAGE_datasets_2026-10-08.zip` is an unchanged mirror of the prepared dataset archive previously distributed through Google Drive. Extract it with a ZIP64-compatible utility to obtain:

```text
MAGE_datasets/
  baby/
  sports/
  clothing/
  elec/
  README.txt
  SHA256SUMS.txt
```

Electronics uses `elec` as its directory and dataset name in the code.

| Dataset | Recommendation items | Expanded text rows | Stage-1 `--seed-max` |
| --- | ---: | ---: | ---: |
| Baby | 7,050 | 71,317 | 7,049 |
| Sports | 18,357 | 77,832 | 18,356 |
| Clothing | 23,033 | 84,695 | 23,032 |
| Electronics | 63,001 | 498,196 | 63,000 |

Each dataset directory includes:

| File | Contents |
| --- | --- |
| `<dataset>.inter` | Original interaction table and split labels |
| `<dataset>_hard_eval.inter` | Same training/validation rows, with a restricted hard test subset |
| `i_id_mapping.csv`, `u_id_mapping.csv` | Original catalog/user identifiers and zero-based model IDs |
| `image_feat.npy` | Aligned visual features, `(items, 4096)`, float64 |
| `text_feat.npy` | Augmented text features, `(items, 512)`, float32 |
| `text_feat_raw.npy` | Unmodified recommendation-item text features |
| `text_feat_full.npy`, `i_id_mapping_full.csv` | Expanded text catalog and aligned item mapping |
| `related_by_id.json` | Relations over the expanded item IDs |
| `p2b_map_filtered.json` | Complementary edges between recommendation items |
| `offline_manifest.json` | Preparation settings, shapes and feature hashes |

The `.inter` files and the mapping `.csv` files are **tab-separated**. Interaction columns are `userID`, `itemID`, `rating`, `timestamp`, and `x_label`. The split labels are `0` = training, `1` = validation, and `2` = test. Keep the supplied IDs, row order and splits unchanged. The user mapping retains the original benchmark's pseudonymous user identifiers.

## Explore interaction splits

The Parquet files under `data/` expose the original interaction table as four Hugging Face configurations: `baby`, `sports`, `clothing`, and `elec`. Each retains the five source columns and separates rows according to the existing `x_label`; no split is resampled. These tables also power the dataset viewer.

```python
from datasets import load_dataset

dataset = load_dataset("Alex6657/MAGE", "baby")
print(dataset["train"][0])
```

| Configuration | Users | Training rows | Validation rows | Test rows |
| --- | ---: | ---: | ---: | ---: |
| `baby` | 19,445 | 118,551 | 20,559 | 21,682 |
| `sports` | 35,598 | 218,409 | 37,899 | 40,029 |
| `clothing` | 39,387 | 197,338 | 40,150 | 41,189 |
| `elec` | 192,403 | 1,304,382 | 192,403 | 192,403 |

The Parquet configurations contain interactions only. Download the archive for feature matrices, mappings, graphs and the separate hard-test files needed for the full MAGE workflow.

## Download the training data

```python
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="Alex6657/MAGE",
    repo_type="dataset",
    filename="archive/MAGE_datasets_2026-10-08.zip",
)
print(archive)
```

The archive is **1,882,496,634 bytes**. SHA-256:

```text
8412c252d01efbcadc48101b3d3634191318e554d11219a93119b6253fb0651a
```

The extracted archive contains `SHA256SUMS.txt` for its files. From `MAGE_datasets/`, verify it with `sha256sum -c SHA256SUMS.txt` where available.

## Prepare text features for training

The archived `text_feat.npy` files use Stage-1 gap retention **0.60**. The current Baby, Sports and Clothing configurations in the code repository use **0.20**, **0.40** and **0.40**, respectively. For those configurations, regenerate `text_feat.npy` from `text_feat_full.npy` using the repository's Stage-1 commands before training. Preserve the original files and use a fresh working directory without graph caches.

`--seed-max` is the inclusive last recommendation item ID. Expanded text rows supply neighbor features and must not be passed directly as the Stage-2 training feature matrix. The first recommendation-item rows of the expanded mapping must match the seed mapping. All text embeddings have 512 columns; visual features have 4096 columns.

Electronics has 480 zero visual-feature rows, representing missing visual signal.

## Use and attribution

These files are intended for research on multimodal recommendation. Preserve the original benchmark attribution and applicable third-party terms. This mirror does not grant additional rights to the underlying datasets. The code repository retains its [upstream references and notices](https://github.com/Alex-bupt/mage/blob/main/docs/upstream/GUME_README.md).

Please cite the [MAGE paper](https://arxiv.org/abs/2610.37409) when using the prepared features and graphs, together with the relevant original data and method sources.