toolathon123's picture
v1.1.0 - 30-min live update: +750 records (total 5,748); canonical data refreshed; snapshot + checksums added
206cb65 verified
|
Raw
History Blame Contribute Delete
6.72 kB
---
pretty_name: EU-Retail-UX-Feedback-Live
license: cc-by-4.0
language:
- en
- fr
- de
multilinguality:
- multilingual
size_categories:
- n<10K
task_categories:
- text-classification
tags:
- ecommerce
- retail
- user-experience
- customer-feedback
- sentiment-analysis
- text-classification
- multilingual
- ux
- customer-service
- product-reviews
- tabular
- text
- timeseries
configs:
- config_name: default
data_files:
- split: train
path: data/ux_feedback_live.parquet
---
# EU-Retail-UX-Feedback-Live
Real-time, **GDPR-anonymised** user-experience (UX) feedback collected from the
e-commerce website, mobile app and customer-service platform of a large European
retail company (1000+ employees) operating in **the United Kingdom, France and
Germany**. The dataset is refreshed every **30 minutes**, and every published
version is immutable, checksummed and taggable so it is fully **auditable and
rollback-able**.
## Quick summary
| Attribute | Value |
|---|---|
| Dataset name | `EU-Retail-UX-Feedback-Live` |
| Refresh cadence | every 30 minutes |
| Encoding | UTF-8 |
| Markets | United Kingdom, France, Germany |
| Languages | en-GB, fr-FR, de-DE |
| Latest version | v1.2.0 |
| Total records (latest) | 6,498 |
| License | CC BY 4.0 |
## Why this dataset exists
To continuously optimise the customer experience, the company collects feedback
across the full shopping journey — page loading speed, search relevance, product
recommendations, checkout flow, delivery experience and after-sales support.
This dataset powers:
- sentiment analysis / rating prediction models,
- multilingual text-classification of experience dimensions,
- root-cause analysis of UX pain points per market,
- dashboards tracking CX health over time.
## Data fields / schema
Each record contains the following fields:
| Column | Type | Description |
|---|---|---|
| `feedback_id` | string | Unique record identifier (pseudonymous). |
| `user_pseudo_id` | string | Pseudonymous user identifier (HMAC-SHA256, not reversible without salt). |
| `session_id` | string | Pseudonymous browsing/session identifier. |
| `market_region` | string | `United Kingdom`, `France` or `Germany`. |
| `country_code` | string | ISO 3166-1 alpha-2 code: `GB`, `FR`, `DE`. |
| `language` | string | BCP-47 language tag: `en-GB`, `fr-FR`, `de-DE`. |
| `timestamp` | string | ISO 8601 UTC timestamp of the feedback event. |
| `experience_dimension` | string | One of 6 dimensions (see below). |
| `rating` | int | Score from 1 (very poor) to 5 (excellent). |
| `feedback_text` | string | Free-text feedback, UTF-8 encoded, in the market language. |
| `device_type` | string | `desktop`, `mobile` or `tablet`. |
| `source_channel` | string | `website`, `mobile_app` or `customer_service`. |
### Experience dimensions
`page_load_speed`, `search_relevance`, `product_recommendations`,
`checkout_flow`, `delivery_experience`, `after_sales_support`
## Coverage
- **Markets:** United Kingdom (GB), France (FR), Germany (DE) — balanced, 1/3 each.
- **Languages:** en-GB, fr-FR, de-DE (feedback text matches the market language).
- **Channels:** website, mobile app, customer service.
- **Devices:** desktop, mobile, tablet.
- **Time range:** rolling 24h window, refreshed every 30 minutes.
## GDPR compliance (pseudonymisation & minimisation)
This dataset was built in line with the EU **General Data Protection Regulation
(GDPR)** (Regulation (EU) 2016/679). The pipeline performs:
1. **Deletion of direct identifiers** — fields `user_email`, `user_name`,
`phone_number` and `address` are dropped and never published. **No direct
personal data is present in this dataset.**
2. **Pseudonymisation** — the real user identifier is replaced by a
pseudonymous ID derived with `HMAC-SHA256` plus a secret salt
(`user_pseudo_id`), and session identifiers are pseudonymised as well
(`session_id`). The salt is stored separately and access-controlled, so the
mapping cannot be reconstructed from the published data.
3. **Free-text scrubbing** — residual PII patterns (e-mail addresses, phone
numbers, UK/FR/DE postcodes, card-like numbers) are redacted from the
`feedback_text` field before upload.
4. **Data minimisation** — only the attributes required for UX analytics are
published; no financial, biometric or sensitive data is collected.
5. **Purpose limitation & retention** — pseudonymised records are retained only
for as long as needed for CX analysis; raw (identifiable) data is held in a
restricted internal store with a short retention policy and is never uploaded
to the Hub.
Full documentation of the anonymisation measures is provided in
[`gdpr/anonymisation_report.md`](gdpr/anonymisation_report.md) and the
reproducible pipeline is in [`scripts/generate_ux_feedback.py`](scripts/generate_ux_feedback.py).
> ⚠️ This dataset contains **synthetic** feedback used for demonstration and
> model-development purposes. No real customer data is included.
## Versioning & auditability
The dataset is updated every **30 minutes**. Version management follows
semantic versioning and is designed to be **auditable and rollback-able**:
- **Immutable snapshots** are stored under `snapshots/<version>/`.
- Every version has a **SHA-256 checksum** (parquet and CSV) recorded in
[`versions/manifest.json`](versions/manifest.json).
- A human-readable history is maintained in
[`versions/CHANGELOG.md`](versions/CHANGELOG.md).
- Each published version corresponds to a **git commit/tag** on the Hub
(`v1.0.0`, `v1.1.0`, `v1.2.0`, …), so any previous version can be restored
with a single checkout (rollback).
- The canonical, latest snapshot always lives at `data/ux_feedback_live.parquet`.
| Version | Window (UTC) | Added | Total | SHA-256 (parquet) |
|---|---|---|---|---|
| v1.0.0 | 2026-08-12T13:47 → 2026-08-13T13:47 | 4,998 | 4,998 | `3882044c…e54` |
| v1.1.0 | 2026-08-13T13:47 → 14:17 | 750 | 5,748 | `e0488237…ccb` |
| v1.2.0 | 2026-08-13T14:17 → 14:47 | 750 | 6,498 | `1365bc41…875` |
## Usage
```python
from datasets import load_dataset
ds = load_dataset("toolathon123/EU-Retail-UX-Feedback-Live", split="train")
print(ds)
```
```python
import pandas as pd
df = pd.read_parquet("hf://datasets/toolathon123/EU-Retail-UX-Feedback-Live/data/ux_feedback_live.parquet")
```
## Limitations
- Synthetic data: the free-text feedback is generated, not collected from real
customers, and may not fully reflect real-world language variation.
- The dataset is refreshed continuously; results should always be pinned to a
specific version (e.g. `v1.2.0`) for reproducible experiments.
- Ratings are not expert-verified; treat them as user-reported signals.