mayacheruty/BON-Synthetic-Influencers

Bot Or Not (BON): Synthetic Influencer Authenticity Dataset

Can you trust an influencer's audience before you pay them? BON is a 10,000-row synthetic dataset for influencer due diligence: telling real audiences apart from fake or inflated ones (bought followers, bots, click farms, AI-fabricated personas).

Fully synthetic. No real people, accounts, or personal data. Numbers in the text are illustrative.

At a glance

Rows	10,000
Columns	26 (behavioral metrics + text + labels)
Task	Fraud detection, retrieval, report generation
Modality	Tabular + English text
Ground truth	Labeled by a simulator (not by a model)
License	MIT

What it is

Each row is a fictional influencer profile: account metrics (followers, engagement, audience quality), a few pieces of model-written text (bio, comments, caption), and a ground-truth label saying whether the account is authentic and, if not, what kind of fake it is.

Why it exists

Influencer fraud costs brands billions a year, and buyers cannot easily tell a real audience from a padded one. Detection needs labeled data, but real data never comes with a trustworthy "is this fake" label. A simulator fixes that: because we decide the truth in code, the detection task can be measured honestly.

How it was built

• Simulator (Python) picks one of 9 archetypes and generates the metrics from that type's fingerprint. The archetype is the ground-truth label, set by code.
• Language model (Qwen2.5-3B-Instruct) writes only the surface text to match the archetype.
• Hardened for realism: feature ranges were widened and noised, and sophisticated fraud was added (a case_type column) so some frauds hide their signature and some clean accounts look suspicious.

Notebook	Open
Synthetic data generation
Exploratory data analysis

The 9 archetypes

Archetype	Group	Fraud?
authentic_organic	real	no
micro_real	real	no
viral_legit	real	no
lightly_boosted	gray zone	no
bought_followers	fake audience	yes
click_farm_audience	fake audience	yes
bought_engagement	fake engagement	yes
engagement_pod	fake engagement	yes
ai_persona	fake identity	yes

Columns

Group	Columns
Identity + labels	profile_id, archetype, is_fraud, fraud_severity, case_type, niche
Account	follower_count, following_count, posts_count, account_age_months
Engagement	avg_likes, avg_comments, avg_views, engagement_rate, like_comment_ratio, view_follower_ratio
Authenticity signals	follower_following_ratio, growth_spike_score, audience_geo_mismatch, audience_real_ratio, comment_authenticity, profile_pic_flag
Text	bio, sample_comments, post_caption, audience_note

Leakage note: comment_authenticity and profile_pic_flag are signals the simulator assigns that a real auditor cannot read directly. Drop them from an honest classifier and recover that signal from the text.

What the data shows

The short version: fake and real influencers leave different traces, and a model can tell them apart about 90% of the time using only the signals a real auditor could actually see.

A few things we learned:

• The data is clean and fair. No missing values, no duplicates, and the text has zero AI giveaways. All nine influencer types appear in equal numbers, so nothing is skewed.
• Every kind of fake has its own tell. Bought followers get almost no likes. Click farms have audiences in the wrong countries. Bought likes arrive with no real comments. No single number catches everyone, so the model weighs several signals together, the way a human analyst would.

• We caught our own mistake. The first model scored a suspicious 100%. That perfect score was the red flag: the task was too easy and one column was quietly leaking the answer. We removed it, roughed the data up to look more like the messy real world, and re-tested. The honest result held at about 90%.
• It fails on the safe side. When the model is wrong it usually flags a real creator who happens to look shady, rather than letting a fraud slip through. For a due-diligence tool, that is exactly the mistake you want it to make.

Headline (fraud detection, honest features only): AUC 0.903, accuracy 0.824. With every feature it reaches 0.960, but 0.903 is the number we stand behind.

Honest framing

This is a synthetic proof of concept. The fingerprints are grounded in documented real-world fraud signals, but synthetic classes separate more cleanly than reality, so 0.903 is an upper bound, not a production claim. The balanced classes and severity scores are learning-design choices, not estimates of how common fraud actually is (real fraud is a skewed minority).

Files

• bon_dataset.csv - the dataset (10,000 rows)
• BON_DataGen.ipynb - synthetic data generation (simulator + LLM + hardening)
• BON_EDA.ipynb - exploratory data analysis