README.md

---
license: cc-by-nc-4.0
tags:
- TAAC2026
- recommendation
size_categories:
- 1K<n<10K
---

# TAAC2026 Second Round Demo Dataset (1000 Samples)

A sample dataset containing 1000 user-item interaction records for the second round of [TAAC2026 competition](https://algo.qq.com). This dataset uses a **flat column layout** — all features are stored as individual top-level columns instead of nested structs/arrays.

# Dataset Overview

| Property | Value |
|---|---|
| **File** | `demo_1000.parquet` |
| **Rows** | 1,000 |
| **Columns** | 142 |
| **File Size** | ~40 MB |

## Columns

The 142 columns fall into **7 categories**:

| Category | Count | Arrow Type | Description |
|---|---|---|---|
| **ID & Label** | 5 | `int64` / `int32` | Core identifiers, label, and timestamp |
| **User Int Features** | 54 | `int64` / `list<int64>` | Discrete user features, including both single-value scalar features (such as age, gender, etc.) and multi-value array features (like marital status, etc.), describing user basic attributes and preferences. |
| **User Dense Features** | 17 | `list<float>` | Continuous-valued user features, including embeddings and other aligned signals for some corresponding integer features. |
| **Item Int Features** | 17 | `int64` / `list<int64>` | Discrete item features, including item categories, types, and other basic information, as well as multi-label information for items. |
| **Item Dense Features** | 4 | `list<float>` | Continuous-valued item features, including embeddings. |
| **Domain Sequence Features** | 45 | `list<int64>` | Behavioral sequence features from 4 domains |


## Detailed Column Schema

### ID & Label Columns (5 columns)
All these 5 columns have no null value.

| | | | | | |
|--------|---------|---------|------------|------------|---------- |
| **Column** | user_id | item_id | label_type | label_time | timestamp
| **Data Type** | `int64` | `int64` | `int32` | `int64` | `int64`

### User Int Features (54 columns)
- `user_int_feats_{1,3,4,48-59,82,86,92-110}`: Scalar `int64`, total 36 columns.
- `user_int_feats_{15, 60, 62-66, 80, 89-91, 111, 115-118, 121-122}`: Array `list<int64>`, total 18 columns.

### User Dense Features (17 columns)
- `user_dense_feats_{61, 87, 120, 123, 130-132}`: Array `list<float>`, total 7 columns, representing user embedding features([SUM](https://arxiv.org/abs/2311.09544) , [LFM4Ads](https://arxiv.org/abs/2508.14948) ).
- `user_dense_feats_{62-66, 89-91, 118, 121}`: Array `list<float>`, total 10 columns, corresponding to the integer features user_int_feats_{62-66, 89-91, 118, 121}, with the same length.
  - An Example:
  `user_int_feats_62`: [1, 2, 3], `user_dense_feats_62`: [10.5, 20, 15.5]
  
    **Explanation:** Here, the two arrays are aligned element by element. For example, the 1st element in user_int_feats_62 (value 1) denotes a specific entity or category, while the 1st element in user_dense_feats_62 (value 10.5) provides some statistics for that element, such as a dwell time, a score/probability.

### Item Int Features (17 columns)
- `item_int_feats_{5-10, 12-13, 16, 81, 83-85, 114, 119, 126}`: Scalar `int64`, total 16 columns.
- `item_int_feats_{11}`: Array `list<int64>`, total 1 column.

### Item Dense Features (4 columns)
- `item_dense_feats_{124, 127-129}`: Array `list<float>`, total 4 columns, representing item embedding features.


### Domain Sequence Features (45 columns)
`list<int64>` sequences from 4 behavioral domains:
- `domain_a_seq_{38-46}`: 9 columns
- `domain_b_seq_{67-79, 88}`: 14 columns
- `domain_c_seq_{27-37, 47}`: 12 columns
- `domain_d_seq_{17-26}`: 10 columns

---

## Usage

```python
import pyarrow.parquet as pq
import pandas as pd

# Read the parquet file
df = pd.read_parquet("demo_1000.parquet")

print(df.shape)       # (1000, 142)
print(df.columns)     # ['user_id', 'item_id', 'label_type', ...]
```

With Hugging Face `datasets`:
```python
from datasets import load_dataset

ds = load_dataset("TAAC2026/second_round_sample_1000")
print(ds)
```