File size: 4,270 Bytes
eeed8eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
---
pretty_name: Feature Selection Benchmark Datasets (HRLFS)
task_categories:
- tabular-classification
- tabular-regression
tags:
- feature-selection
- tabular
- reinforcement-learning
- benchmark
size_categories:
- 100K<n<1M
---

# Feature Selection Benchmark Datasets (HRLFS)

This repository hosts the **21 benchmark datasets** used in our ACM TKDD paper:

> **Comprehend, Divide, and Conquer: Feature Subspace Exploration via Multi-Agent Hierarchical Reinforcement Learning**
> Weiliang Zhang, Xiaohan Huang, Yi Du, Ziyue Qiao, Qingqing Long, Zhen Meng, Yuanchun Zhou, Meng Xiao
> *ACM Transactions on Knowledge Discovery from Data (TKDD), 2026*

- 📄 Paper code: [https://github.com/coco11563/HARLFS](https://github.com/coco11563/HARLFS)

## Dataset Summary

The 21 datasets are publicly available and collected from the Feature Selection Benchmark, NCBI Gene Expression Omnibus (GEO), UCI, Kaggle, OpenML, libSVM, etc. They cover three task types — binary classification (C), multi-class classification (MC), and regression (R) — and span diverse fields including biology, finance, image, and synthetic data. Sample sizes range from 253 to 83,733 and feature dimensions from 21 to 20,670.

| Dataset | Task | #Samples | #Features | File |
|---|---|---|---|---|
| SpectF | C | 267 | 44 | `spectf.hdf` |
| SVMGuide3 | C | 1,243 | 21 | `svmguide3.hdf` |
| German Credit | C | 1,001 | 24 | `german_credit.hdf` |
| Credit Default | C | 30,000 | 25 | `credit_default.hdf` |
| SpamBase | C | 4,601 | 57 | `spam_base.hdf` |
| Megawatt1 | C | 253 | 38 | `megawatt1.hdf` |
| Ionosphere | C | 351 | 34 | `ionosphere.hdf` |
| Mice-Protein | MC | 1,080 | 77 | `mice_protein.hdf` |
| Coil-20 | MC | 1,440 | 400 | `coil-20.hdf` |
| MNIST | MC | 10,000 | 784 | `mnist.hdf` |
| Otto | MC | 61,878 | 93 | `otto.hdf` |
| Jannis | MC | 83,733 | 54 | `jannis.hdf` |
| Cao | MC | 4,186 | 13,488 | `cao.hdf` |
| Han | MC | 2,746 | 20,670 | `han.hdf` |
| Openml_586 | R | 1,000 | 25 | `openml_586.hdf` |
| Openml_589 | R | 1,000 | 25 | `openml_589.hdf` |
| Openml_607 | R | 1,000 | 50 | `openml_607.hdf` |
| Openml_616 | R | 500 | 50 | `openml_616.hdf` |
| Openml_618 | R | 1,000 | 50 | `openml_618.hdf` |
| Openml_620 | R | 1,000 | 25 | `openml_620.hdf` |
| Openml_637 | R | 500 | 50 | `openml_637.hdf` |

## Data Format

Each dataset is stored as a single HDF5 file with two keys:

- `raw_train` — training split (80% of samples)
- `raw_test` — test split (20% of samples)

Each key stores a `pandas.DataFrame` where **the last column is the label** and all preceding columns are features.

## Usage

Download the files with `huggingface_hub` and load them with `pandas`:

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Shaow/Feature_Selection_Dataset",
    filename="spam_base.hdf",
    repo_type="dataset",
)

train = pd.read_hdf(path, key="raw_train")
test = pd.read_hdf(path, key="raw_test")

X_train, y_train = train.iloc[:, :-1].to_numpy(), train.iloc[:, -1].to_numpy()
X_test, y_test = test.iloc[:, :-1].to_numpy(), test.iloc[:, -1].to_numpy()
```

To reproduce the experiments in the paper, place the `.hdf` files under the `./data` directory of the [HARLFS repository](https://github.com/coco11563/HARLFS) and run, e.g.:

```bash
python HRLFS.py --dataset spam_base
```

Note: reading these files requires `pandas` and `tables` (PyTables): `pip install pandas tables`.

## Licensing

All datasets are redistributed from publicly available sources (Feature Selection Benchmark, NCBI GEO, UCI, Kaggle, OpenML, libSVM). Please refer to the original sources for their respective license terms; this collection is provided for research purposes only.

## Citation

If you use these datasets, please cite our paper:

```bibtex
@article{zhang2026comprehend,
  title={Comprehend, Divide, and Conquer: Feature Subspace Exploration via Multi-Agent Hierarchical Reinforcement Learning},
  author={Zhang, Weiliang and Huang, Xiaohan and Du, Yi and Qiao, Ziyue and Long, Qingqing and Meng, Zhen and Zhou, Yuanchun and Xiao, Meng},
  journal={ACM Transactions on Knowledge Discovery from Data},
  year={2026}
}
```

## Contact

For questions, please contact the corresponding author: **Meng Xiao** (shaow@cnic.cn).