tg-forward-graph

Collection and analysis pipeline for mapping forward relationships between public Telegram channels.

Produces a directed weighted graph where nodes are channels and edges represent observed message forwards, as a basis for structural analysis of information ecosystems.

Approach

The pipeline maps structure, not identity. It measures which channels relay which, how often, and over what period. It does not attribute content to individuals, and does not attempt to infer intent or coordination from relay patterns. Applied to a cluster of Italian-language political channels; see the sampling note for the full seed list and selection criteria.

Data source

Public channel previews (t.me/s/<channel>), accessible without an account or authentication. No channel subscription, no interaction, no login required to reproduce the collection.

Requests are paginated backwards with a 2-second delay. Only publicly accessible content is read.

Requirements

python >= 3.10
requests beautifulsoup4 lxml polars

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

Usage

Three stages, each reading from the previous. Raw data is never modified in place; processed/ is always regenerable.

1. Collect — reads config/seeds.txt (one channel handle per line, no @), writes one JSONL file per channel to raw/.

python collect.py

2. Inspect — prints collection statistics and ranks channels most frequently forwarded but not yet observed, for snowball expansion.

python stats.py

3. Build — normalises, deduplicates, applies the time window, writes Parquet to processed/.

python build.py

Output

processed/messaggi.parquet — one row per message channel, msg_id, ts, testo, fwd_da, fwd_nome

processed/archi.parquet — one row per directed edge origine, destinazione, peso, primo, ultimo

peso is the forward count; primo and ultimo bound the observed activity of that edge, allowing temporal filtering downstream.

Parameters

FINESTRA_GG in build.py sets the analysis window in days (default 90). The window is anchored to the most recent message in the dataset rather than to execution time, so results are deterministic on a fixed input.

Vary this parameter to test how sensitive results are to the window choice.

Repository contents

Included: pipeline code, sampling methodology, aggregated edge list (channel pairs with counts).

Not included: raw message archives, full message text, session files, credentials. Message content is collected locally for analysis but not redistributed.

Methodology and limitations

See notes/campionamento.md (Italian) for the full sampling procedure, expansion thresholds, stopping criterion, and declared limitations.

Key limitations in brief:

• Snowball bias. The sample reflects what is connected to the initial seeds. Isolated clusters are invisible.
• Observational asymmetry. Seed channels are observed directly; other nodes appear only as forward destinations. Their outbound degree is unknown, not zero.
• Bounded depth. Five pages per channel. Effective coverage varies with posting frequency.

Interpretation

Observed forwarding measures relay, not coordination, and certainly not inauthenticity. Establishing that an observed pattern differs from chance requires a null model, which is not part of this repository.

Licence

Code: MIT. Aggregated data: CC BY 4.0.