---
license: apache-2.0
task_categories:
- tabular-classification
size_categories:
- 100M<n<1B
---
We introduce TRACE, the first benchmark dataset for post-click GMV prediction with delayed feedback.

📁 Data structure includes:
1.  **Features**:
    *   **User/Item/Contextual Attributes**: 22 such attributes are provided (e.g., `feature_0` to `feature_21`).
    *   **Click Timestamp**: `click_time`.

2.  **Labels and Sequence Information**:
    *   **Gross Merchandise Volume (GMV) Sequence**: The sequence of transaction amounts for all purchases associated with a click. This is represented by `dirpay_amt` for the current purchase and `prev_dirpay_amt` for previous purchases.
    *   **Purchase Count Sequence**: The sequence indicating the order of each purchase within the attribution window for a given click. `count` represents the current purchase's order, and `total_counts` indicates the final number of purchases.
    *   **Purchase Timestamp Sequence**: The sequence of timestamps for each purchase. `pay_time` typically represents the timestamp of the current purchase, and `prev_pay_time` contains timestamps of previous purchases.
    *   **Repurchase Indicator (`multi_tag`)**: A binary label (0 or 1) indicating whether the GMV generated by a click is from a single purchase (0) or a repurchase (1). This is derived from the number of purchases.
    *   **Final Ground-Truth GMV Label**: The total GMV accumulated by the end of the attribution window. This is calculated as the sum of `dirpay_amt`.