Why standard regression fails for creative element scoring. Learn how Shapley Additive Explanations (Lundberg & Lee) calculate the precise marginal contribution of visual text overlays and audio transitions without multicollinearity bias.
1. The Multicollinearity Trap in Creative Tag Analysis
Performance marketers frequently attempt to evaluate creative attributes by running correlation tables: e.g. correlating the presence of a "Green Background" or "30% Off Badge" with overall ROAS. In practice, this approach produces severe false positives due to multicollinearity and confounding variables.
For example, if high-performing UGC creators frequently shoot in natural daylight and use fast-paced speech, simple correlation will attribute high ROAS to both factors equally, failing to identify whether it was the creator authenticity or the auditory cadence that drove the conversion.
Standard correlation cannot disentangle co-occurring creative features. A principled game-theoretic framework is required to compute true marginal value.
2. Cooperative Game Theory & Shapley Formulation
To solve this attribution challenge, AdHelix employs TreeSHAP (SHapley Additive exPlanations), an algorithm rooted in cooperative game theory (Lundberg & Lee, NIPS 2017; Shapley, 1953). In this framework, an ad creative is modeled as a cooperative coalition of discrete features (e.g. 5500K Natural Daylight, Problem-Agitation Hook, Karaoke Captions, Female Presenter 25-34).
The Shapley value assigns a payout (marginal lift in ROAS, CTR, or CPA) to each feature by averaging its marginal contribution across all possible subsets of creative features.
\phi_i(x) = \sum_{S \subseteq F \setminus \{i\}} \frac{|S|!(|F| - |S| - 1)!}{|F|!} \left[ f_x(S \cup \{i\}) - f_x(S) \right]3. Empirical Bayes L2 Regularized Ridge Regression for Small Cohorts
In ad accounts with small sample sizes (N < 30 active video creatives), local data variance can introduce noise into the feature attribution model. To maintain statistical stability, AdHelix applies Empirical Bayes L2 Regularized Ridge Regression.
The model shrinks sparse local brand coefficients toward vertical category priors derived from aggregated cross-account benchmarks. As the brand runs more ads and collects more impression telemetry, the local empirical data progressively overrides the prior.
\hat{\beta}_{\text{Bayes}} = \left( X^T X + \lambda I \right)^{-1} X^T YEmpirical Bayes shrinkage protects media buyers from overreacting to noisy creative experiments with low sample sizes.
4. Non-Linear Feature Interactions in Short-Form Video
Unlike linear regression models, TreeSHAP natively captures non-linear feature interactions. For instance, our research demonstrates that a high-contrast text overlay produces a +14% lift in Hook Rate when paired with fast editing tempo (<1.2s cuts), but produces a -8% drag when paired with slow, static panning shots.
By modeling these multi-variable interactions, creative strategists receive concrete rules on which combinations of production, talent, and narrative tags generate positive synergy.
- Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems (NeurIPS 2017), 30, 4765–4774.https://papers.nips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67d0f-Abstract.html
- Shapley, L. S. (1953). A Value for n-Person Games. In H. W. Kuhn & A. W. Tucker (Eds.), Contributions to the Theory of Games (Vol. 2, pp. 307–317). Princeton University Press.
- Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd ed.). Chapman and Hall/CRC.https://doi.org/10.1201/b16018