README.md

---
license: other
pretty_name: CAPT Public-Trace Dataset and Landmark Analysis
language:
- en
tags:
- digital-platforms
- algorithmic-visibility
- amazon-bestsellers
- new-york-times
- longitudinal
- research
configs:
- config_name: amazon_daily_completed
  data_files:
  - split: train
    path: data/amazon_daily_completed.csv
- config_name: nyt_weekly_completed
  data_files:
  - split: train
    path: data/nyt_weekly_completed.csv
---

# CAPT dataset

This repository accompanies the submitted paper *Algorithmic Visibility as a
Platform-Mediated Outcome: Public Signals and Future Product Persistence in
Digital Marketplaces*.

The release contains public Amazon Australia bestseller-surface observations,
public New York Times bestseller-list records, the completed longitudinal
panels used for the paper's sensitivity analysis, reproducible landmark-model
code, and the corresponding derived results.

## Repository contents

| Path | Description |
|---|---|
| `data/amazon_daily_completed.csv` | Amazon daily panel, 14 April–18 May 2026 |
| `data/nyt_weekly_completed.csv` | NYT weekly panel, list dates 19 April–24 May 2026 |
| `analysis/capt_analyze_jmis_landmark.py` | Landmark cohort construction and logistic models |
| `analysis/capt_plot_nyt_amazon_comparison.py` | Title/author normalization and matching functions |
| `analysis/capt_validation_schema.py` | Shared filesystem helper |
| `results/jmis_landmark_book_level.csv` | Derived ASIN-level analytic table |
| `results/jmis_landmark_nested_models.csv` | Nested-model performance |
| `results/jmis_landmark_summary.json` | Model estimates and validation metrics |


## Dataset

Predictors use information dated no later than 30 April 2026. The prediction
landmark is 4 May 2026. Outcomes are product presence on the directly captured
Amazon snapshots of 11 May (7-day retention) and 18 May (14-day retention).
Continuous predictors are standardized within the analysis sample. Model
validation uses an 80/20 stratified holdout and five-fold stratified
cross-validation with seed `20260629`.

## Reproduce the sensitivity analysis

From the repository root:

```bash
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

python analysis/capt_analyze_jmis_landmark.py \
  --amazon data/amazon_daily_completed.csv \
  --nyt data/nyt_weekly_completed.csv \
  --output-dir reproduced_results
```

The command writes the book-level table, nested-model table, and JSON model
summary to `reproduced_results/`.

## Load with `datasets`

```python
from datasets import load_dataset

amazon = load_dataset(
    "bonsai44/CAPT",
    "amazon_daily_completed",
)
nyt = load_dataset(
    "bonsai44/CAPT",
    "nyt_weekly_completed",
)
```

## Fields

The Amazon panel includes snapshot date, category node and label, page and rank,
title, author, ASIN, public product URL, visible rating average and count,
format, price, and provenance fields.

The NYT panel includes scrape date, list slug and title, list page date, rank,
weeks-on-list text, title, author, publisher, description, ISBN-13, public book
URL, and provenance fields.

## Limitations and responsible use

These are bounded public ranking-surface observations, not unit sales,
conversions, advertising exposure, recommendation logs, or a complete sample of
either platform. Amazon ASINs identify product-format records rather than unique
works. Title/author matching may contain false positives or false negatives.
The analysis is observational and does not identify causal effects.

The CSVs include product titles, author names, identifiers, descriptions, and
URLs originating from third-party public pages. Before making the repository
public, the depositor is responsible for confirming that redistribution
complies with the relevant platform terms and publisher or institutional
requirements. The MIT license in `LICENSE` applies to the accompanying software,
not to third-party database rights or third-party content.

## Citation

Replace the placeholders below with the final author and publication details
before publishing the dataset:

```bibtex
@misc{capt_public_trace_2026,
  author       = {Hang Thanh Bui, Priyadharshini Muthukannan},
  title        = {Algorithmic Visibility as a Platform-Mediated Outcome: Public Signals and Future Product Persistence in Digital Marketplaces},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/bonsai44/CAPT}},
  note         = {Data and code accompanying the submitted manuscript}
}
```