T-ECD

# T-ECD: T-Tech E-commerce Cross-Domain Dataset


⭐️ **T-ECD** is a large-scale synthetic cross-domain dataset for recommender systems research, created by T-Bank's RecSys R&D team.
It captures real-world e-commerce interaction patterns across multiple domains while ensuring complete anonymity through synthetic generation.

🎯 Overview
T-ECD represents user interactions across five different e-commerce domains within a banking ecosystem:

- **Marketplace** — browsing and interacting with items in an e-commerce marketplace.
- **Retail** — interactions within a retail delivery service, including cart additions and completed orders.
- **Payments** — online and offline financial transactions between users and brands.
- **Offers** — responses to promotional content such as impressions, clicks, and partner transitions.
- **Reviews** — explicit user feedback in the form of ratings and embeddings of textual comments.

**Scale:**
- **~135B** interactions
- ~44M users
- ~30M items
- **1300+ days of temporal coverage**

Additionally, we provide **T-ECD Small** - a compact version containing 1B interactions that excludes the Payments domain.

<div style="font-size: 1.1em;">

| Metric | T-ECD Small | T-ECD Full |
|--------|-------------|------------|
| 🔄 **Interactions** | ~1B | **~135B** |
| 👥 **Users** | ~3.5M | **~44M** |
| 📦 **Items** | ~2.6M | **~30M** |
| 🏪 **Brands** | ~29K | **~1M** |
| 📅 **Temporal Coverage** | 200+ days | **1300+ days** |
| 🌐 **Domains** | 4 (excl. Payments) | **5 (all domains)** |

</div>

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/Y3hHv_cipdq2p4A9jiQoz.png" style="max-width: 80%; height: auto;">

Cross-domain consistency is achieved by aligning identifiers across all domains:
- the same `user_id` always refers to the same individual user, and
- the same `brand_id` always refers to the same brand entity.

This alignment allows researchers to seamlessly link interactions from different services, enabling studies in transfer learning, cross-domain personalization, and multi-task modeling.

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/QG0DavvcvccN1GcN_gRL6.png" style="max-width: 80%; height: auto;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/s8a8iC4RmUjsD_hzOVPvD.png" style="max-width: 80%; height: auto;">

---

### 📂 Data Schema

The dataset is stored in **Parquet** format with daily partitions (`{day}`).
The directory structure is as follows:

```
t-ecd/
├── users.pq
├── brands.pq
├── marketplace/
│ ├── events/{day}.pq
│ └── items.pq
├── retail/
│ ├── events/{day}.pq
│ └── items.pq
├── payments/
│ ├── events/{day}.pq
│ └── receipts/{day}.pq
├── offers/
│ ├── events/{day}.pq
│ └── items.pq
└── reviews/{day}.pq
```

#### Data availability
<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/c2Clc9bNxL9i7jgGBfBq2.png" style="max-width: 80%; height: auto;" alt="Temporal distribution of events over domains">
*Temporal distribution of events over domains*
In line with real-world industrial environments, domain-specific data availability varies in historical depth.
This reflects practical constraints including data retention policies and product lifecycle stages -
newer e-commerce services naturally have shorter histories compared to established banking domains like payments and transactions.

### ⚙️ Events and Catalogs

- **Events**: Each domain provides logs of user interactions with the following possible columns:

- `action_type` — interaction type (e.g., view, click, add-to-cart, order, transaction).
- `subdomain` — surface where the interaction occurred (recommendations, catalog, search, checkout, campaign); available in Marketplace and Retail.
- `item_id` — present in Marketplace, Retail, and Offers; identifies a specific product or offer.
- `brand_id` — present in all domains; denotes the seller, store, or partner associated with an item, offer, or transaction.
- `price` — represents the monetary value of the interaction.
- `count` — represents the amount of items in single interaction.
- `os` — user operating system, available in Marketplace and Retail.

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/Q7aeb_I-Yf-rcqyPDTOLa.png" style="max-width: 80%; height: auto;" >

- **Item catalogs (`items.pq`)**: Available for Marketplace, Retail, and Offers. Each entry includes:
- `item_id`
- `brand_id`
- category information (if available)
- pretrained embedding (if available)

- **User catalog (`users.pq`)**: Contains anonymized user attributes such as region and socio-demographic cluster.

- **Brand catalog (`brands.pq`)**: Contains `brand_id`, brand-level metadata, and embeddings.

#### 🧾 Special Structures

- **Receipts (`payments/receipts/{day}.pq`)**:
Some transactions include detailed receipts with purchased items, their quantities, and prices.
Items are aligned with Marketplace and Retail catalogs, enabling fine-grained cross-domain linkage at the product level.

- **Reviews (`reviews/{day}.pq`)**:
Provide explicit ratings per brand.
Raw text reviews are not included; instead, we release pretrained text embeddings to preserve privacy while enabling multimodal research.
---

### 🛠️ Data Collection

T-ECD was generated through a multi-step process:
1. **Sampling of event chains**: sequences of interactions were sampled from real logs of T-Bank ecosystem services.
2. **Anonymization**: user and brand identifiers were pseudonymized; sensitive attributes removed.
3. **Synthetic generation**: based on real distributions and event patterns, new synthetic interaction chains were produced, preserving structural properties such as sparsity, heavy tails, cross-domain overlaps, and behavioral contexts.

This process ensures that the dataset is privacy-preserving while remaining representative of industrial recommender system data.

## ⚠️ Important Note on Temporal Data Usage

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/zaPAcuD3CItTzP2PBkErs.png" style="max-width: 80%; height: auto;">

**To prevent data leakage, events from the final 12 hours should not be used for prediction tasks.**
The dataset contains temporal noise that requires maintaining a minimum 12-hour gap between the timestamp of the most recent user event and the prediction timestamp.
This constraint applies to both training and testing scenarios to avoid temporal data leakage.

## Download

#### Basic Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="t-tech/T-ECD",
repo_type="dataset",
allow_patterns="dataset/full/",
local_dir="./t_ecd_data",
token="<your_hf_token>"
)
```

#### Selective Download

For advanced usage including selection of domains and date ranges we provide custom downloader [tecd_downloader.py](https://huggingface.co/datasets/t-tech/T-ECD/blob/main/tecd_downloader.py)

Example usage:

```python
from tecd_downloader import download_dataset

download_dataset(
token="<your_hf_token>",
dataset_path="dataset/small",
local_dir="t_ecd_small_partial",
domains=["retail", "marketplace"],
day_begin=1300,
day_end=1308,
max_workers=10
)
```

---

### 🔐 License

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0) licence

---

# T-ECD：T-Tech电商跨域数据集


⭐️ **T-ECD** 是由T-Bank的推荐系统（RecSys）研发团队打造的大规模合成跨域数据集，用于推荐系统领域的研究。该数据集通过合成生成技术实现全匿名化处理，同时捕捉多领域下的真实电商交互模式。

🎯 数据集概览
T-ECD涵盖了银行生态体系内五个不同电商领域的用户交互行为：
- **电商商城（Marketplace）**：用户在电商商城内浏览并与商品进行交互的行为。
- **零售配送（Retail）**：零售配送服务内的交互行为，包括加购商品与完成订单的操作。
- **支付交易（Payments）**：用户与品牌之间的线上及线下金融交易行为。
- **营销活动（Offers）**：用户对推广内容的响应行为，包括曝光、点击以及合作伙伴跳转操作。
- **用户评论（Reviews）**：以评分和文本评论嵌入向量（embedding）形式呈现的显性用户反馈。

**数据集规模：**
- 约**1350亿**次交互
- 约4400万用户
- 约3000万商品
- 覆盖时长超**1300天**

此外，我们还推出了**T-ECD精简版（T-ECD Small）**：该精简版本不含支付交易领域，仅包含约10亿次交互。

| 指标 | T-ECD精简版 | T-ECD完整版 |
|--------|-------------|------------|
| 🔄 **交互次数** | ~10亿 | **~1350亿** |
| 👥 **用户数** | ~350万 | **~4400万** |
| 📦 **商品数** | ~260万 | **~3000万** |
| 🏪 **品牌数** | ~2.9万 | **~100万** |
| 📅 **覆盖时长** | 200+天 | **1300+天** |
| 🌐 **覆盖领域** | 4个（不含支付交易） | **5个（全领域）** |

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/Y3hHv_cipdq2p4A9jiQoz.png" style="max-width: 80%; height: auto;">

跨域一致性通过对齐各领域的标识符实现：
- 相同的`user_id`始终对应同一用户个体
- 相同的`brand_id`始终对应同一品牌实体

这种对齐方式可让研究人员无缝关联不同服务的交互数据，支持迁移学习、跨域个性化以及多任务建模等方向的研究。

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/QG0DavvcvccN1GcN_gRL6.png" style="max-width: 80%; height: auto;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/s8a8iC4RmUjsD_hzOVPvD.png" style="max-width: 80%; height: auto;">

---

### 📂 数据模式（Data Schema）
数据集以**Parquet（列式存储格式）**格式存储，并按日分区（`{day}`）进行组织。目录结构如下：

t-ecd/
├── users.pq
├── brands.pq
├── marketplace/
│ ├── events/{day}.pq
│ └── items.pq
├── retail/
│ ├── events/{day}.pq
│ └── items.pq
├── payments/
│ ├── events/{day}.pq
│ └── receipts/{day}.pq
├── offers/
│ ├── events/{day}.pq
│ └── items.pq
└── reviews/{day}.pq

#### 数据可用性
<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/c2Clc9bNxL9i7jgGBfBq2.png" style="max-width: 80%; height: auto;" alt="Temporal distribution of events over domains">
*各领域事件的时间分布*
与真实工业场景一致，各领域数据的历史覆盖深度存在差异。这反映了实际约束条件，包括数据留存政策与产品生命周期阶段：相较于支付交易等成熟的银行生态领域，新兴电商服务的历史数据时长自然更短。

### ⚙️ 事件与目录数据
- **事件数据**：各领域均提供用户交互日志，其可能包含以下字段：
- `action_type`：交互类型（例如浏览、点击、加购、下单、交易等）。
- `subdomain`：交互发生的场景（包括推荐位、商品目录、搜索、结账页、营销活动等）；仅在电商商城与零售配送领域可用。
- `item_id`：在电商商城、零售配送与营销活动领域均存在，用于标识特定商品或营销活动。
- `brand_id`：全领域均存在，用于标识与商品、营销活动或交易相关的卖家、店铺或合作伙伴。
- `price`：表示本次交互的交易金额。
- `count`：表示单次交互中的商品数量。
- `os`：用户操作系统；仅在电商商城与零售配送领域可用。

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/Q7aeb_I-Yf-rcqyPDTOLa.png" style="max-width: 80%; height: auto;" >

- **商品目录（`items.pq`）**：在电商商城、零售配送与营销活动领域均可用，每条记录包含以下内容：
- `item_id`
- `brand_id`
- 分类信息（若可用）
- 预训练嵌入向量（若可用）

- **用户目录（`users.pq`）**：包含经过匿名化处理的用户属性，例如地区与社会人口学聚类标签。

- **品牌目录（`brands.pq`）**：包含`brand_id`、品牌元数据以及嵌入向量。

#### 🧾 特殊数据结构
- **交易凭证（`payments/receipts/{day}.pq`）**：
部分交易包含详细的交易凭证，记录了购买商品、商品数量与价格。商品与电商商城及零售配送领域的目录数据对齐，可实现细粒度的跨域商品级关联。

- **用户评论（`reviews/{day}.pq`）**：
提供针对品牌的显性评分。原始文本评论未公开，我们仅发布预训练文本嵌入向量以保护用户隐私，同时支持多模态研究。

---

### 🛠️ 数据采集流程
T-ECD通过多阶段流程生成：
1. **事件链采样**：从T-Bank生态服务的真实日志中采样交互序列。
2. **匿名化处理**：对用户与品牌标识符进行假名化处理，移除敏感属性。
3. **合成数据生成**：基于真实数据分布与交互模式生成全新的合成交互序列，保留数据的结构特性，例如稀疏性、重尾分布、跨域重叠性以及行为上下文。

该流程确保数据集在保护用户隐私的同时，能够真实反映工业级推荐系统数据的特性。

## ⚠️ 时序数据使用注意事项

<img src="https://cdn-uploads.huggingface.co/production/uploads/645d4947f5760d1530d55023/zaPAcuD3CItTzP2PBkErs.png" style="max-width: 80%; height: auto;">

**为防止数据泄露，请勿使用最后12小时内的事件数据开展预测任务。**
数据集包含时序噪声，因此需要确保最新用户事件的时间戳与预测时间戳之间至少间隔12小时。
该约束适用于训练与测试全流程，以避免时序数据泄露。

## 数据集下载

#### 基础下载
python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="t-tech/T-ECD",
repo_type="dataset",
allow_patterns="dataset/full/",
local_dir="./t_ecd_data",
token="<your_hf_token>"
)

#### 选择性下载
针对需要选择领域与时间范围的高级使用场景，我们提供了自定义下载工具 [tecd_downloader.py](https://huggingface.co/datasets/t-tech/T-ECD/blob/main/tecd_downloader.py)

示例用法：
python
from tecd_downloader import download_dataset

download_dataset(
token="<your_hf_token>",
dataset_path="dataset/small",
local_dir="t_ecd_small_partial",
domains=["retail", "marketplace"],
day_begin=1300,
day_end=1308,
max_workers=10
)

---

### 🔐 开源许可
本数据集采用知识共享署名-非商业性使用-相同方式共享4.0国际许可协议（CC BY-NC-SA 4.0）进行发布。