Henry236/pog-testset-v2

POG FITB / CP test set V2

Two evaluation sets for outfit compatibility on POG / Alibaba iFashion, rebuilt so that distractors are drawn only from items a model could plausibly have learned. Used as the shared benchmark for a COMP9727 (UNSW) group project comparing four recommenders.

This is a derived benchmark split, not the POG dataset. It holds nothing but POG item ids. No images, no titles, no user data. To get the items themselves, go to the original source linked below.

Files

File	Rows	Size
pog_fitb_test_v2_min3.txt	82,463 questions	41.8 MB
pog_cp_test_v2_min3.txt	198,928 rows (99,464 real + 99,464 corrupted)	59.5 MB
rebuild_test_set_v2.ipynb	—	31 KB
SHA256SUMS.txt	—	—

Despite the .txt extension both data files are JSON arrays, indent=4. Load them with json.load.

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download("Henry236/pog-testset-v2", "pog_fitb_test_v2_min3.txt",
                       repo_type="dataset")
fitb = json.load(open(path, encoding="utf-8"))

Format

FITB (fill in the blank). Hide one item from a real outfit, pick it from 4 candidates. label is the index of the correct candidate.

{
  "example_id": "fill_in_the_blank_nondisjoint_test_000000",
  "items": ["6ed185bab890de3b2e07aa68911deead", "b32370b52c4f436caea0416816fe9427"],
  "blank_index": 0,
  "candidates": ["3e1c99e3bca206ab...", "495a9f3927162870...",
                 "b9b575c6a189dd71...", "fe9062f58cc7503e..."],
  "label": 0
}

CP (compatibility prediction). Score a real outfit above a corrupted one, reported as AUC. label is 1 for real, 0 for corrupted.

{
  "example_id": "compatibility_nondisjoint_test_000000",
  "items": ["487869864d000441...", "671ba95500defea5...", "f0016c4f17fdc9dd..."],
  "label": 1
}

The nondisjoint in example_id is inherited from the upstream POG test files, which follow the Polyvore-Outfits naming convention.

Why "min3", and what it fixes

The distractors in the original test files are sampled from the whole 4,747,039 item catalogue. Only 583,464 of those items (12.3%) ever appear in an outfit, so about 88% of distractors are items no model has trained on.

That is a shortcut. All four FITB candidates come from one category, and a model falls back to a category-average embedding for items it does not know, so the three unknown candidates score identically and the known answer stands out. A model can score well by recognising which item it has seen before, without learning anything about compatibility.

V2 draws distractors only from items appearing 3 or more times in the training outfits. Nothing else changes. The question, the blank index, the correct answer and the example_id are all carried over, so V2 lines up question-by-question with the upstream file.

Two ways the fix can backfire, both handled:

Trap	Handling
The correct answer is itself out of vocabulary, making it the only unrecognised candidate	Those questions are dropped
A corrupted outfit holds more unknown items than the real one	Swaps only ever replace a known item with a known item of the same category, and the unknown-item profile of real and corrupted sets is asserted equal

The threshold of 3 lines up with the k-core threshold used by the graph model and sits inside the Item2Vec vocabulary (min_count = 2), so candidates are known to every model being compared. At a threshold of 1 the distractors are only 36% likely to be graph nodes against 88% for the answer, a 51 point gap. At 3 that gap closes.

Sanity check on the result: a control that scores candidates purely on "have I seen this item before" gets 0.2502 on V2, against a chance level of 0.2500. Recognition is worth nothing here. That is the property the rebuild was for.

Provenance

rebuild_test_set_v2.ipynb is the exact notebook that produced these files, included so the split can be audited and regenerated. It is deterministic: RANDOM_SEED = 42, MIN_APPEARANCES = 3.

Inputs, all derived from the original POG release:

Input	What it is
pog_train_outfit_data.txt	911,821 training outfits, the 90% split of POG outfit_data.txt
pog_item_data.csv	item id → category id, parsed from POG item_data.txt
pog_fitb_test_outfit_data.txt, pog_cp_test_outfit_data.txt	the upstream test files this rebuilds

The notebook asserts its output before writing: zero out-of-vocabulary candidates, zero duplicate candidate sets, unique example_ids, questions and blank indices unchanged from upstream, an even spread of answer positions, and for CP an equal count of real and corrupted rows with matching outfit-length and unknown-item profiles.

Code that consumes these files: https://github.com/QuangMinhPhan23/outfit-recommender

Source

The underlying data is POG / Alibaba iFashion: 1,013,136 outfits, 4.7M items, 19.2M user click records, released with the paper below.

• Paper: https://arxiv.org/abs/1905.01866
• Data and original test files: https://github.com/wenyuer/POG

The upstream release states no license. These derived files are redistributed for research and coursework, with attribution, on the understanding that they carry no rights beyond whatever the original release grants. If you are the POG authors and want this taken down, open a discussion on this repo and it will be removed.

Citation

Cite the original dataset paper:

@inproceedings{chen2019pog,
  title     = {POG: Personalized Outfit Generation for Fashion Recommendation
               at Alibaba iFashion},
  author    = {Chen, Wen and Huang, Pipei and Xu, Jiaming and Guo, Xin and
               Guo, Cheng and Sun, Fei and Li, Chao and Pfadler, Andreas and
               Zhao, Huan and Zhao, Binqiang},
  booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on
               Knowledge Discovery and Data Mining (KDD)},
  year      = {2019},
  pages     = {2662--2670},
  doi       = {10.1145/3292500.3330652}
}

The FITB and CP protocols, and the nondisjoint split naming, come from the Polyvore-Outfits line of work:

@inproceedings{vasileva2018typeaware,
  title     = {Learning Type-Aware Embeddings for Fashion Compatibility},
  author    = {Vasileva, Mariya I. and Plummer, Bryan A. and Dusad, Krishna and
               Rajpal, Shreya and Kumar, Ranjitha and Forsyth, David},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2018}
}

@inproceedings{han2017bilstm,
  title     = {Learning Fashion Compatibility with Bidirectional LSTMs},
  author    = {Han, Xintong and Wu, Zuxuan and Jiang, Yu-Gang and Davis, Larry S.},
  booktitle = {ACM Multimedia},
  year      = {2017}
}

If you use the min-3 rebuild specifically, please also link back to this repo, so that the distractor change is not mistaken for the upstream split.