CausalMTA: Unbiased Multi‑Touch Attribution via Causal Inference

Abstract: Multi‑Touch Attribution (MTA) aims to reconstruct user touch‑point trajectories and fairly allocate conversion credit. Existing data‑driven MTA methods train a conversion‑prediction model and distribute credit using Shapley Value, assuming the model is unbiased. This assumption is violated because user preferences introduce confounding bias and out‑of‑distribution (OOD) issues in counterfactual prediction.

The paper proposes CausalMTA , a causal‑inference‑based unbiased MTA framework that eliminates both static and dynamic confounding biases. The framework consists of two key modules: (1) Journey Reweighting, which re‑weights user paths using a variational recurrent auto‑encoder and inverse‑probability‑weighting to remove static user‑attribute bias; (2) Causal Conversion Prediction, which employs a gradient‑reversal‑layer‑augmented RNN to learn representations disentangled from ad exposure, thus correcting dynamic preference bias. After bias removal, an unbiased conversion model is trained and Shapley Values are computed for attribution.

Background: Advertising products (search, display, feed, live, video, interactive) generate multiple touch points in a customer journey. Traditional rule‑based attribution (first‑touch, last‑touch, linear, etc.) ignores true causal effects, leading to suboptimal budget allocation. Data‑driven MTA methods improve on this but ignore confounding variables, causing biased predictions.

Methodology: Journey Reweighting models the generation of ad‑channel sequences with a variational recurrent auto‑encoder (VRAE). Sample weights are estimated via density‑ratio estimation and applied using inverse‑probability‑weighting (IPTW) to obtain a re‑weighted dataset free of static confounding. Causal Conversion Prediction builds on the Counterfactual Recurrent Network (CRN) by adding a gradient‑reversal layer, enabling the model to learn balanced representations of dynamic features and ad exposure. The combined pipeline yields an unbiased conversion predictor, and Shapley Value is used to allocate credit to each touch point.

Theoretical Analysis: Under the independence assumption, total confounding bias can be decomposed into static and dynamic components. The paper proves that re‑weighting eliminates static bias and that the causal prediction module removes dynamic bias by minimizing a loss that forces the learned representation to be independent of ad exposure. Consequently, the unbiased model minimizes counterfactual prediction error.

Experiments: CausalMTA is evaluated on (1) synthetic datasets with controllable bias levels, (2) the public Criteo dataset, and (3) real‑world Alibaba‑Mama data. Compared with eight baselines (statistical learning, deep learning, and causal methods) and two ablations, CausalMTA consistently outperforms in conversion‑rate prediction (higher AUC) and yields more reasonable attribution values. Data‑replay experiments demonstrate lower CPA and higher CVR under budget constraints, confirming superior ROI potential.

Conclusion: CausalMTA provides a principled, bias‑free solution for multi‑touch attribution, improving both conversion prediction accuracy and the fairness of channel credit allocation. The approach bridges causal inference theory with practical advertising systems, delivering state‑of‑the‑art performance on simulated, public, and proprietary datasets.