🎥 Live demo: an interactive Streamlit dashboard (
streamlit_app.py) is included. Deploy it free on Streamlit Community Cloud (entrypointstreamlit_app.py) and replace this line + the badge above with your*.streamlit.appURL.
CreatorPulse is a single-file data-science portfolio project. It simulates two years of weekly behavioral data for 500 creators, injects realistic pre-burnout decay signals, and then runs the full applied-DS toolkit end-to-end: exploratory analysis, statistical hypothesis testing, an A/B-test simulation, survival analysis, a gradient-boosted classifier, and a causal difference-in-differences estimate — all from one reproducible script.
Read this first. Every number in this repo comes from synthetic data generated with a fixed seed (
np.random.seed(42)). There is no external dataset, scraping, or API. All metrics below — ROC-AUC, lift, hazard ratios, "ROI" — are illustrative outputs of the simulation, not measurements of real-world creators. The value of this project is the methodology and code rigor, not empirical findings. It is framed honestly on purpose.
git clone https://github.com/koutilyaY/creatorpulse
cd creatorpulse
python3 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
python creatorpulse_analysis.pyWhat happens: the script prints a sectioned analysis log to the terminal and writes a 6-panel dashboard image, creatorpulse_dashboard.png, to the current working directory. It also calls plt.show(), so a matplotlib window will pop up — suppress it for headless/CI runs:
MPLBACKEND=Agg python creatorpulse_analysis.pyRuntime is ~37 seconds on a laptop. The slow part is the 500-iteration permutation test inside the difference-in-differences section.
The creator economy is large, and when a top creator stops posting, the downstream cost — lost content, broken brand deals, churned audiences — is real. The premise of this project is that burnout leaves a behavioral fingerprint before it becomes visible: posting cadence fragments, engagement decays, sentiment dips, and response rates fall in the weeks leading up to a creator going dark.
CreatorPulse encodes that hypothesis directly into a data simulator, then asks the natural follow-up questions a data scientist would: Is the signal statistically detectable? Can a model use it to predict who is at risk? Does an intervention measurably help? Did a platform change causally push creators harder? Each section answers one of those with an appropriate method.
This is a methodology showcase. Because the burnout signal is injected by the generator, the analysis should recover it — the point is demonstrating that the right tool is applied correctly to each question, with leakage controls and honest reporting.
No real creators, platforms, or APIs are touched. creatorpulse_analysis.py generates everything internally:
| Property | Value |
|---|---|
| Creators | 500 |
| Observation window | 104 weeks (2 years) |
| Weekly records | 52,000 |
| Seed | np.random.seed(42) (plus per-section seeds 123 / 77) |
| Burnout signal | injected in the 8 weeks before each burned-out creator's burnout_week |
| Feature cutoff | weeks 1–52 only feed the model (prevents post-burnout leakage) |
Generated per-creator attributes: archetype (Lifestyle, Gaming, Finance, Comedy, Education, Fitness), platform, subscriber count (log-normal), team size, years active, monetization status. Archetype-specific base posting frequency and burnout probability drive the simulation; solo / long-tenure / TikTok creators get an added burnout risk bump.
Weekly behavioral series: posts_per_week, engagement_rate, caption_sentiment, comment_toxicity, response_rate. Each picks up decay during the pre-burnout signal window and collapses post-burnout.
Each section maps to a function in the script and answers a distinct question.
Burnout rates broken out by archetype, platform, and team size; a direct pre-signal vs. signal-window engagement comparison to confirm the injected decay is present.
- Shapiro–Wilk normality check (justifies going non-parametric).
- Mann–Whitney U with rank-biserial effect size (healthy vs. pre-burnout engagement).
- Welch's t-test on posting frequency.
- Chi-square test of independence (platform × burnout) with Cramér's V.
- Bonferroni correction across the three primary tests.
A simulated "Wellness Check" intervention: at-risk creators split into treatment/control, with two-sample t-test, Cohen's d, retention lift, and a statistical power estimate.
- Kaplan–Meier curves, solo vs. team creators, via
KaplanMeierFitter. - Log-rank test for curve separation.
- Cox Proportional Hazards (
CoxPHFitter, ridge-penalized) for hazard ratios across team size, tenure, monetization, and log-subscribers.
Leakage-controlled features built only from weeks 1–52: 12-week OLS trend slopes (engagement / posts / sentiment), coefficient of variation in posting, week-over-week engagement-drop counts, response-rate decline, plus creator metadata.
GradientBoostingClassifier (scikit-learn) — the "XGBoost-style" boosted-tree classifier referenced throughout. Stratified train/test split, 5-fold stratified CV, ROC-AUC, PR-AUC, and gain-based feature importances.
Difference-in-Differences estimating the effect of a simulated week-52 algorithm change on posting frequency (TikTok = treatment, Instagram = control), with significance from a 500-iteration permutation test. Implemented directly in numpy.
Tenure cohorts (0–1 / 1–3 / 3–5 / 5yr+) with burnout-rate progression.
creatorpulse_analysis.py (np.random.seed=42)
│
┌────────────────────────────┴────────────────────────────┐
│ §1 Synthetic data generation │
│ 500 creator profiles ──► 52,000 weekly records │
└────────────────────────────┬────────────────────────────┘
│
┌──────────────┬──────────────┬───┴──────────┬──────────────┬──────────────┐
▼ ▼ ▼ ▼ ▼ ▼
§2 EDA §3 Stat tests §4 A/B test §5 Survival §6 Feature §8 Causal
(MWU/χ²/t + (t-test, (KM + log-rank engineering DiD + 500×
Bonferroni) Cohen's d, + Cox PH, (weeks 1–52, permutation
power) lifelines) leakage-free) test)
│
▼
§7 ML classifier
(GBM, 5-fold CV,
ROC/PR-AUC)
│
▼
§10 6-panel dashboard ──► creatorpulse_dashboard.png + plt.show()
All packages are required — the script hard-imports them at module load and will not run if any is missing.
| Package | Role | Required? |
|---|---|---|
pandas |
data wrangling / time-series shaping | Required |
numpy |
simulation, RNG, DiD math | Required |
scipy |
Shapiro–Wilk, Mann–Whitney, t-test, chi-square, normal CDF | Required |
scikit-learn |
GradientBoostingClassifier, CV, ROC/PR-AUC |
Required |
lifelines |
Kaplan–Meier, log-rank, Cox PH | Required — hard-imported (KaplanMeierFitter, CoxPHFitter, logrank_test) |
matplotlib |
dashboard rendering | Required |
seaborn |
dashboard styling | Required |
statsmodels |
listed in requirements.txt |
Listed, but not imported by the current script — the DiD is implemented in pure numpy. Safe to keep installed; not load-bearing today. |
Earlier docs described
lifelines(andstatsmodels) as "optional." That was incorrect: the survival-analysis section will crash on import withoutlifelines. Install everything inrequirements.txt.
These are representative values produced by the seeded simulation. They are not validated against real creators and should not be read as real-world effect sizes. With the fixed seed they are reproducible run-to-run.
| Output | Typical value | What it is |
|---|---|---|
| Classifier ROC-AUC (test) | ~0.84 | Boosted-tree AUC recovering the injected burnout signal |
| 5-fold CV ROC-AUC | ~0.83 ± 0.02 | tight CV variance on synthetic features |
| A/B retention lift | ~+20pp, p < 0.001 | effect baked into the intervention simulator |
| DiD estimate | ~+1.2 posts/week | causal effect of the simulated algorithm change |
| Cox / KM separation | solo > team risk | hazard direction matching the generator's risk bumps |
Any "297× ROI" or "2.8× faster burnout" style numbers that may appear in older copy are back-of-envelope illustrations driven by the simulation's own parameters — keep them out of any real-world or résumé claim.
Running the script writes this 6-panel summary (Kaplan–Meier curves, ROC curve, feature importances, burnout by archetype, burnout by platform, burnout by tenure cohort):
creatorpulse/
├── creatorpulse_analysis.py # single self-contained pipeline (data gen → analysis → dashboard)
├── creatorpulse_dashboard.png # generated output (overwritten on each run)
├── requirements.txt # pinned-minimum dependencies
├── .gitignore # ignores .venv/, __pycache__/, .DS_Store
└── README.md
- Synthetic by design. The model recovers a signal that the generator inserted; this validates that the pipeline works, not that burnout is predictable in the wild.
- No train/serve split, no real labels, no external validation. This is a portfolio demonstration of method breadth and correctness, not a production system.
- Reproducible. Fixed seeds make every figure and metric deterministic across runs.
MIT.
Koutilya Yenumula — M.S. Computer Science, University of South Florida · AWS Certified Data Engineer – Associate github.com/koutilyaY · LinkedIn