Can Complementary Signals Bridge Similarity Islands? Manifold-Augmented Graph Embedding for Multimodal Recommendation
Paper • 2610.37409 • Published
Search is not available for this dataset
userID int64 0 19.4k | itemID int64 0 7.05k | rating float64 1 5 | timestamp int64 982M 1.41B | x_label int64 0 0 |
|---|---|---|---|---|
0 | 1,587 | 5 | 1,373,932,800 | 0 |
0 | 1,879 | 5 | 1,373,932,800 | 0 |
0 | 0 | 5 | 1,373,932,800 | 0 |
1 | 2,828 | 5 | 1,372,464,000 | 0 |
1 | 5,403 | 5 | 1,367,452,800 | 0 |
1 | 1,434 | 5 | 1,372,377,600 | 0 |
1 | 0 | 5 | 1,372,464,000 | 0 |
2 | 2,846 | 2 | 1,395,187,200 | 0 |
2 | 6,578 | 3 | 1,387,497,600 | 0 |
2 | 4,011 | 5 | 1,387,497,600 | 0 |
3 | 1,886 | 5 | 1,376,697,600 | 0 |
3 | 0 | 5 | 1,376,697,600 | 0 |
3 | 638 | 4 | 1,376,697,600 | 0 |
3 | 2,194 | 4 | 1,376,697,600 | 0 |
4 | 5,523 | 5 | 1,396,310,400 | 0 |
4 | 3,408 | 5 | 1,396,310,400 | 0 |
4 | 5,070 | 4 | 1,396,310,400 | 0 |
5 | 4,914 | 4 | 1,390,521,600 | 0 |
5 | 6,291 | 5 | 1,399,680,000 | 0 |
5 | 1,372 | 5 | 1,393,977,600 | 0 |
6 | 771 | 5 | 1,374,019,200 | 0 |
6 | 5,058 | 5 | 1,374,019,200 | 0 |
6 | 0 | 5 | 1,374,019,200 | 0 |
6 | 1,232 | 5 | 1,374,019,200 | 0 |
6 | 3,063 | 4 | 1,374,019,200 | 0 |
6 | 4,144 | 5 | 1,374,019,200 | 0 |
6 | 1,308 | 1 | 1,374,019,200 | 0 |
6 | 6,166 | 4 | 1,374,019,200 | 0 |
6 | 6,244 | 5 | 1,374,019,200 | 0 |
7 | 4,657 | 5 | 1,372,550,400 | 0 |
7 | 4,832 | 4 | 1,374,796,800 | 0 |
7 | 3,183 | 4 | 1,370,995,200 | 0 |
8 | 1,143 | 4 | 1,369,958,400 | 0 |
8 | 3,408 | 4 | 1,401,667,200 | 0 |
8 | 1,307 | 4 | 1,398,124,800 | 0 |
8 | 0 | 3 | 1,398,124,800 | 0 |
8 | 5,494 | 3 | 1,398,124,800 | 0 |
8 | 5,166 | 5 | 1,393,027,200 | 0 |
8 | 3,485 | 5 | 1,398,124,800 | 0 |
8 | 2,146 | 4 | 1,393,027,200 | 0 |
8 | 6,716 | 2 | 1,401,667,200 | 0 |
8 | 5,606 | 4 | 1,398,124,800 | 0 |
9 | 6,266 | 5 | 1,373,328,000 | 0 |
9 | 5,519 | 5 | 1,367,884,800 | 0 |
9 | 6,498 | 5 | 1,386,720,000 | 0 |
9 | 5,987 | 5 | 1,383,004,800 | 0 |
9 | 1,210 | 5 | 1,373,328,000 | 0 |
9 | 6,730 | 2 | 1,383,091,200 | 0 |
9 | 4,793 | 5 | 1,376,956,800 | 0 |
9 | 4,367 | 5 | 1,383,523,200 | 0 |
9 | 4,666 | 3 | 1,384,992,000 | 0 |
9 | 4,788 | 4 | 1,368,662,400 | 0 |
9 | 6,473 | 5 | 1,383,004,800 | 0 |
9 | 4,626 | 5 | 1,373,328,000 | 0 |
10 | 2,089 | 5 | 1,323,388,800 | 0 |
10 | 2,368 | 5 | 1,350,345,600 | 0 |
10 | 2,045 | 5 | 1,320,364,800 | 0 |
10 | 4,445 | 5 | 1,323,388,800 | 0 |
10 | 4,310 | 5 | 1,320,364,800 | 0 |
11 | 2,300 | 3 | 1,361,836,800 | 0 |
11 | 5,125 | 5 | 1,361,836,800 | 0 |
11 | 0 | 5 | 1,361,836,800 | 0 |
11 | 6,165 | 5 | 1,361,836,800 | 0 |
11 | 3,661 | 3 | 1,361,836,800 | 0 |
12 | 6,145 | 5 | 1,374,364,800 | 0 |
12 | 2,312 | 5 | 1,377,907,200 | 0 |
12 | 0 | 4 | 1,374,364,800 | 0 |
12 | 1,755 | 5 | 1,377,907,200 | 0 |
12 | 6,579 | 5 | 1,374,364,800 | 0 |
12 | 4,339 | 5 | 1,377,907,200 | 0 |
12 | 4,520 | 5 | 1,374,710,400 | 0 |
13 | 3,131 | 5 | 1,402,790,400 | 0 |
13 | 3,946 | 5 | 1,402,790,400 | 0 |
13 | 0 | 5 | 1,402,790,400 | 0 |
14 | 1,169 | 4 | 1,397,520,000 | 0 |
14 | 473 | 3 | 1,369,008,000 | 0 |
14 | 14 | 3 | 1,385,424,000 | 0 |
14 | 6,072 | 3 | 1,396,742,400 | 0 |
14 | 5,428 | 4 | 1,396,742,400 | 0 |
14 | 1,158 | 3 | 1,385,424,000 | 0 |
14 | 4,240 | 4 | 1,387,670,400 | 0 |
15 | 5,955 | 2 | 1,368,835,200 | 0 |
15 | 3,726 | 5 | 1,373,587,200 | 0 |
15 | 6,210 | 4 | 1,386,720,000 | 0 |
15 | 6,859 | 5 | 1,392,768,000 | 0 |
15 | 4,695 | 4 | 1,374,105,600 | 0 |
15 | 6,732 | 5 | 1,394,582,400 | 0 |
15 | 6,937 | 3 | 1,405,900,800 | 0 |
15 | 6,747 | 3 | 1,384,473,600 | 0 |
15 | 5,069 | 4 | 1,373,587,200 | 0 |
15 | 6,728 | 5 | 1,376,870,400 | 0 |
15 | 6,575 | 4 | 1,389,571,200 | 0 |
15 | 5,918 | 3 | 1,367,193,600 | 0 |
15 | 4,284 | 5 | 1,382,659,200 | 0 |
15 | 0 | 3 | 1,381,968,000 | 0 |
15 | 4,790 | 2 | 1,370,217,600 | 0 |
16 | 6,706 | 3 | 1,397,088,000 | 0 |
16 | 1,500 | 3 | 1,397,088,000 | 0 |
16 | 1,179 | 2 | 1,397,088,000 | 0 |
16 | 5,479 | 5 | 1,399,593,600 | 0 |
End of preview. Expand in Data Studio
MAGE datasets
Prepared Baby, Sports, Clothing, and Electronics datasets for Can Complementary Signals Bridge Similarity Islands? Manifold-Augmented Graph Embedding for Multimodal Recommendation.
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 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:
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.
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
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:
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.
Please cite the MAGE paper when using the prepared features and graphs, together with the relevant original data and method sources.
- Downloads last month
- -