Growing Viral

The problem

Brands spend over $200B a year on social media advertising, but platform ranking algorithms are moving targets and “going viral” remains mostly unpredictable. Our team’s brief was to take a 2,000-post dataset with engagement metrics, content attributes, and a binary is_viral label, and build a model that could predict virality reliably enough to inform marketing decisions — while staying interpretable enough that content teams could actually act on the output. Black-box accuracy wasn’t the goal; defensible, repeatable predictions were.

How the team worked

The project ran in three stages. Stage 1 produced a baseline C5.0 decision tree at ~93% average accuracy. Stage 2 split the team across five algorithmic philosophies — each of us independently tested a different family of methods against the same problem, so we could compare what each lens revealed about the data. Stage 3 was synthesis: take the strongest findings from all five Stage 2 explorations and build one final solution.

My piece: K-Means clustering

I owned the unsupervised lens. I ran K-Means at multiple values of k using both the Lloyd and Hartigan-Wong algorithms on the normalized engagement features, varying configuration to test whether clusters would naturally separate viral from non-viral posts. My best configuration (k = 4, Lloyd) reached 71.6% average accuracy — about 21 percentage points below the C5.0 baseline.

That gap was the finding.

At k = 2 and k = 3, both clusters mapped to the majority “viral” class, meaning the model functioned as nothing more than a majority-class predictor. This wasn’t a tuning problem. It told us something structural about the data: viral and non-viral posts don’t form natural spherical groups in Euclidean engagement space. The virality boundary is threshold-driven, not distance-driven — exactly the kind of boundary axis-aligned tree splits are designed for, and exactly the kind distance-based methods can’t represent.

This empirically disqualified distance-based classifiers (KNN and similar) for the final solution and locked in tree-based ensembles. We stopped wasting cycles on a methodology that the data’s geometry was never going to reward.

K-Means did produce one durable artifact, though. At k = 4, the clusters turned out to be interpretable engagement archetypes:

• Viral Powerhouse — 97.4% viral rate
• Moderate Performers — middle of the engagement distribution
• High-Share Medium-Reach — disproportionate shares relative to views
• Efficient But Limited — high engagement rate, small audience

In Stage 3, we engineered these archetypes back into the supervised model as a cluster_id feature on each post. The unsupervised pass became a compact summary of where every post sat in overall engagement space, which the Random Forest could then consume alongside the raw engagement features. K-Means was refit per train/test fold rather than globally — fitting on the full dataset would have leaked test-set positions into the cluster boundaries.

The final model

The Stage 3 production model was a Random Forest (1,000 trees, mtry = 2) trained on three high-importance features — likes.group, engagement.group, shares.group — plus the engineered cluster_id.

At a 90/10 train-test split, averaged over ten independent runs:

Metric	RF3 baseline	RF3 + cluster_id
Accuracy	91.9%	92.8%
Sensitivity	93.9%	95.1%
F1-Score	94.2%	94.9%
OOB error	—	6.94%

Two things worth surfacing from those numbers. First, the three-feature model beat the six-feature model at every split ratio. Adding lower-importance predictors like region and comments.group actually hurt performance by injecting noise that diluted the signal from likes and engagement rate. Lean and engineered beat full and raw. Second — and this matters more than any single accuracy number — three independent algorithms converged on the same two dominant predictors: Macy’s Random Forest cforest importance, Evan’s neural network Olden’s-method importance, and my K-Means archetype boundaries all pointed at likes and engagement_rate. That cross-algorithm triangulation was the strongest evidence we had that the signal was real, not a single-method artifact.

The honest caveats

Two limitations we put in the final report and that I’d repeat to anyone reading this:

1. The dataset is synthetic (Kaggle-generated). Evan’s neural network hit 98.7% accuracy in Stage 2, which he flagged himself as suspicious — synthetic data tends to over-inflate model performance. Our final Random Forest’s 92.8% is more credible because of the cross-method convergence, but it’s still synthetic.
2. The model is post-hoc, not pre-publication. The dominant predictors (likes, engagement rate) can’t be measured until a post is already accumulating engagement. Static pre-publication features like region and content type contribute almost nothing. So the model is best understood as a trajectory classifier — useful for deciding whether to amplify an already-trending post — not as an oracle for unpublished content.

Architecture / data flow

Raw dataset (2,000 posts, 15 attributes)
        ↓
Preprocessing
  - cap engagement_rate at 0.5 (preserves all 2,000 records + original class balance)
  - bin numerics into low/med/high using Stage 1 cutoffs
  - engineer hashtag_count
        ↓
        ├── Binned features → Random Forest input
        └── Normalized features
                  ↓
              K-Means refit per fold (k=4, Hartigan-Wong)
                  ↓
              cluster_id → joined back to binned features
        ↓
Random Forest (ntree=1000, mtry=2) on RF3 + cluster_id
        ↓
Predicted is_viral + variable importance

What I’d do differently

The most useful change to my K-Means setup would be to drop likes and engagement_rate from the clustering input entirely. Because those two features dominate the Euclidean distance calculation, the resulting clusters end up being a compressed restatement of “likes × engagement_rate” — which is exactly the signal the Random Forest already has from those features as predictors. The cluster_id ends up partly redundant with what the supervised model can already see.

Re-running K-Means on only the non-dominant features (comments, shares, sentiment_score, hashtag_count) would force the unsupervised pass to find structure that isn’t already captured by likes and engagement rate. The resulting cluster_id would be genuinely additive to the Random Forest — encoding novel information about secondary engagement patterns — rather than restating headline metrics the supervised model already trains on directly. Same archetypes, earned independently of the dominant predictors instead of leaning on them.

What I learned

Evan’s neural network hit 98.7% accuracy on this dataset, and our first instinct was to celebrate; our second was to worry. The data was synthetic, and high accuracy on synthetic data is often a sign of overfitting to artifacts that won’t exist in the real world. The lesson that stuck with me: a number is only as trustworthy as the data behind it, and getting comfortable being suspicious of your own results — especially when they’re flattering — is a skill worth practicing.

Stack

R, kmeans (Lloyd / Hartigan-Wong), randomForest, party::cforest (for unbiased conditional variable importance), caret. Visualizations in base R.