license: cc-by-nc-sa-4.0
Dataset Description
SHIFT15M was introduced as a benchmark for evaluating set-to-set matching models when the training and test distributions differ. Many machine learning methods assume that training and test data are independently and identically distributed, but this assumption is often violated in real-world applications. In fashion, trends change over time, causing shifts in item appearance, prices, user preferences, and outfit composition.
The dataset supports experiments under several types of distribution shifts, including covariate shift and target shift. It also provides multiple benchmark tasks, such as regression, classification, and set-to-set matching.
Supported Tasks
SHIFT15M supports the following tasks:
| Task | Task Type | Shift Type | Input | Output |
|---|---|---|---|---|
| NumLikesRegression | Regression | Target shift | (N, 25) |
(N, 1) |
| SumPricesRegression | Regression | Covariate shift / Target shift | (N, 1) |
(N, 1) |
| ItemPriceRegression | Regression | Target shift | (N, 4096) |
(N, 1) |
| ItemCategoryClassification | Classification | Target shift | (N, 4096) |
(N, 7) |
| Set2SetMatching | Set-to-set matching | Covariate shift | (N, 4096) × (M, 4096) |
(1) |
These tasks are provided through the official software package for handling the dataset.
Citation
@INPROCEEDINGS {10208629,
author = { Kimura, Masanari and Nakamura, Takuma and Saito, Yuki },
booktitle = { 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) },
title = {{ SHIFT15M: Fashion-specific dataset for set-to-set matching with several distribution shifts }},
year = {2023},
pages = {3508-3513},
}