Interpretable State and Time Dependent Multi-Touch Attribution
Published:
Multi-touch attribution (MTA) aims to assign credit to the sequence of ads that influence a customer’s decision to make a purchase. Existing state-of-the-art models often rely on complex black-box predictors with post-hoc attribution (e.g., Shapley values), which can be unstable and difficult for industry to act on. We propose an interpretable, state and time-dependent MTA framework that explicitly models how advertising exposures accumulate and decay in a customer’s latent willingness to purchase. When coupled across customers the resulting problem is formulated as a mixed-integer problem, which we tackle by proposing the application of a family of scalable ADMM and quadratic-penalty based algorithms plus a knowledge distillation step. On synthetic data with known ground truth, our method is robust to noise and recovers accurate period valence and purchase patterns. On a real world Capital One dataset, it matches or outperforms logistic regression, gradient-boosted trees, and LSTMs while preserving white-box attribution.