TAAC2026/data_sample_1000
收藏资源简介:
--- license: cc-by-nc-4.0 tags: - TAAC2026 - recommendation --- # TAAC2026 Demo Dataset (1000 Samples) > [!WARNING] ⚠️**Update[2026.04.10]:** > This demo dataset has been updated to newest version with the following changes: > - The parquet file is now a **flat column layout**, with all features as top-level columns. > - Add a sequence feature, rename feature names and update some features. > Participants should refer to the updated `demo_1000.parquet` and this `README.md` for the latest schema and data details. A sample dataset containing 1000 user-item interaction records for the [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** | 120 | | **File Size** | ~39 MB | ## Columns The 120 columns fall into **6 categories**: | Category | Count | Data Type | Description | |---|---|---|---| | **ID & Label** | 5 | `int64` / `int32` | Core identifiers, label, and timestamp | | **User Int Features** | 46 | `int64` / `list<int64>` | Integer-valued user features (scalar or array) | | **User Dense Features** | 10 | `list<float>` | Float-array user features | | **Item Int Features** | 14 | `int64` / `list<int64>` | Integer-valued item features (scalar or array) | | **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 | Data Type | |---|---| | `user_id` | `int64` | | `item_id` | `int64` | | `label_type` | `int32` | | `label_time` | `int64` | | `timestamp` | `int64` | > [!NOTE] **Note:** > When `user_int_feats_{fid}` and `user_dense_feats_{fid}` share the same `{fid}`, they are aligned and jointly describe the same entity or signal. ### User Int Features (46 columns) - `user_int_feats_{1,3,4,48-59,82,86,92-109}`: Scalar `int64`, total 35 columns. - `user_int_feats_{15, 60, 62-66, 80, 89-91}`: Array `list<int64>`, total 11 columns. ### User Dense Features (10 columns) - `user_dense_feats_{61-66, 87, 89-91}`: Array `list<float>`, total 10 columns. ### Item Int Features (14 columns) - `item_int_feats_{5-10, 12-13, 16, 81, 83-85}`: Scalar `int64`, total 13 columns. - `item_int_feats_{11}`: Array `list<int64>`, total 1 column. ### 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, 120) print(df.columns) # ['user_id', 'item_id', 'label_type', ...] ``` With Hugging Face `datasets`: ```python from datasets import load_dataset ds = load_dataset("TAAC2026/data_sample_1000") print(ds) ```
license: cc-by-nc-4.0 tags: - TAAC2026 - 推荐(recommendation) # TAAC2026 演示数据集(1000条样本) > [!警告] ⚠️**更新[2026.04.10]:** > 本演示数据集已更新至最新版本,变更如下: > - Parquet文件现已采用**扁平化列布局(flat column layout)**,所有特征均作为顶层列存储。 > - 新增序列特征,重命名特征名称并更新部分特征。 > 参赛选手请参阅更新后的`demo_1000.parquet`与本`README.md`以获取最新的数据模式(schema)与数据详情。 本示例数据集包含1000条用户-物品交互记录,用于[TAAC2026竞赛](https://algo.qq.com/)。该数据集采用**扁平化列布局(flat column layout)**——所有特征均以独立顶层列形式存储,而非嵌套结构体与数组(nested structs/arrays)。 ## 数据集概览 | 属性 | 取值 | |---|---| | **文件** | `demo_1000.parquet` | | **行数** | 1,000 | | **列数** | 120 | | **文件大小** | ~39 MB | ## 列分类 120个列分为**6大类**: | 类别 | 数量 | 数据类型 | 描述 | |---|---|---|---| | **ID与标签** | 5 | `int64` / `int32` | 核心标识符、标签与时间戳 | | **用户整型特征** | 46 | `int64` / `list<int64>` | 整型用户特征(标量或数组) | | **用户稠密特征** | 10 | `list<float>` | 浮点数组型用户特征 | | **物品整型特征** | 14 | `int64` / `list<int64>` | 整型物品特征(标量或数组) | | **领域序列特征** | 45 | `list<int64>` | 来自4个领域的行为序列特征 | --- ## 详细列数据模式(schema) ### ID与标签列(5列) 此5列均无空值。 | 列名 | 数据类型 | |---|---| | `user_id` | `int64` | | `item_id` | `int64` | | `label_type` | `int32` | | `label_time` | `int64` | | `timestamp` | `int64` | > [!提示] **注意:** > 当`user_int_feats_{fid}`与`user_dense_feats_{fid}`共享相同的`{fid}`时,二者为对齐特征,共同描述同一实体或信号。 ### 用户整型特征(46列) - `user_int_feats_{1,3,4,48-59,82,86,92-109}`:标量`int64`,共35列。 - `user_int_feats_{15, 60, 62-66, 80, 89-91}`:数组`list<int64>`,共11列。 ### 用户稠密特征(10列) - `user_dense_feats_{61-66, 87, 89-91}`:数组`list<float>`,共10列。 ### 物品整型特征(14列) - `item_int_feats_{5-10, 12-13, 16, 81, 83-85}`:标量`int64`,共13列。 - `item_int_feats_{11}`:数组`list<int64>`,共1列。 ### 领域序列特征(45列) `list<int64>`序列来自4个行为领域: - `domain_a_seq_{38-46}`:9列 - `domain_b_seq_{67-79, 88}`:14列 - `domain_c_seq_{27-37, 47}`:12列 - `domain_d_seq_{17-26}`:10列 --- ## 使用方法 python import pyarrow.parquet as pq import pandas as pd # 读取Parquet文件 df = pd.read_parquet("demo_1000.parquet") print(df.shape) # (1000, 120) print(df.columns) # ['user_id', 'item_id', 'label_type', ...] 使用Hugging Face `datasets`库时: python from datasets import load_dataset ds = load_dataset("TAAC2026/data_sample_1000") print(ds)




