Marketplace Chameleons: Demystifying the Scam Ecosystem of Second-Hand Online Platforms

This repository contains the processed data, LLM prompts, and analysis code used in the paper *Marketplace Chameleons: Demystifying the Scam Ecosystem of Second-Hand Online Platforms*. The artifact consists of three complementary components used in the paper: - `ProcessedData/`: processed scam cases used for downstream measurement. - `LLMPrompt/`: the LLM-assisted filtering, structuring, and taxonomy pipeline. - `AnalysisCode/`: scripts for quantitative analyses and validation. ## Processed Data The `ProcessedData/` directory contains two CSV files with 3,392 cases covering 43 fine-grained scam categories: - `chinese_dataset.csv`: 1,143 cases from RedNote, Douban, and Zhihu. - `english_dataset.csv`: 2,249 cases from Reddit. Raw forum posts, usernames, account identifiers, direct links, and Telegram materials are not included. The processed records are derived research data and should not be interpreted as platform-confirmed ground truth or as a census of scams. Each CSV contains the following fields: - `case_id`: release-local case identifier. - `structed_json`: structured representation generated by the analysis pipeline. The spelling is retained for compatibility with the released files. - `category_id`: assigned fine-grained scam category. - `confidence`: model-reported classification confidence, not a calibrated probability. - `time`: original post time at the available granularity. The `structed_json` field contains the structured case representation used by the analysis pipeline. Its main attributes include victim and scammer roles, traded item, payment channel, transaction stage, trigger event, scam mechanism, behavioral sequence, channel migration, scam outcome, victim loss, recovery outcome, and platform. All 3,392 released rows have unique release-local identifiers, valid structured JSON, and confidence values in the interval [0, 1]. ## LLM Pipeline The `LLMPrompt/` directory contains the implementation of the LLM-assisted analysis pipeline for data filtering, structured extraction, classification, and taxonomy construction. It includes separate scripts for the Chinese and English datasets: - `chinese_prompt.py` - `english_prompt.py` Running this pipeline requires the original source inputs, a taxonomy file, a configured model, and user-provided API credentials. Credentials must not be committed to this repository. ## Analysis Code The `AnalysisCode/` directory provides implementations of the aggregate analyses reported in the paper, including category, platform, item, payment, loss, success-rate, and recovery summaries. The scripts can be run individually from the repository root after completing the environment setup. The directory contains seven analysis scripts: - `analysis_count.py`: scam category distribution analysis. - `analysis_platform.py`: platform-level scam analysis. - `analysis_item.py`: scam-related item analysis. - `analysis_payment.py`: payment method analysis. - `analysis_money.py`: victim financial loss analysis. - `analysis_successrate.py`: scam success rate analysis. - `analysis_recovery.py`: victim recovery outcome analysis. The directory also includes [`Validation Criteria.md`](AnalysisCode/Validation%20Criteria.md), which documents the annotators' tasks and decision criteria used during validation. ## Software Setup Python 3.9 or later is required to run the LLM pipeline and analysis scripts; it is not required to access or inspect the CSV dataset. ```bash python3 -m venv .venv source .venv/bin/activate python -m pip install --upgrade pip python -m pip install -r requirements.txt ``` ## Data Scope and Responsible Use The source code is released under the MIT License; see [`LICENSE`](LICENSE). The processed data are derived from public online-community content and do not include raw posts, usernames, account identifiers, direct links, or Telegram materials. The data are not covered by the MIT License; their reuse may be subject to applicable source-platform terms, copyright, privacy, database rights, and research-ethics requirements. Users should not use the data to identify, target, or harass individuals.