DCRMTA: Unbiased Causal Representation for Multi-touch Attribution

Jiaming Tang

TL;DR

DCRMTA is a new end-to-end framework for Multi-touch Attribution (MTA) that uses causal representation learning to disentangle user preferences from ad effectiveness. By combining adversarial training with counterfactual attention mechanisms, it achieves state-of-the-art conversion prediction and more efficient marketing budget allocation.

What Problem Does This Paper Address?

In the world of online advertising, "Multi-touch Attribution" (MTA) is the holy grail: it’s the process of determining which specific ads (touchpoints) actually caused a user to buy a product. However, this is notoriously difficult due to confounding bias.

Imagine a user who loves high-end sneakers. They are likely to click on many sneaker ads and likely to buy sneakers regardless of the ads. A simple model might give too much credit to the ads, failing to realize the user's pre-existing preference was the real driver.

Previous attempts to fix this usually tried to "delete" user preference features entirely to achieve an unbiased model. The authors of this paper argue that this is a mistake: user features actually contain valuable causal information. The challenge is to eliminate the bias (the confounding part) while keeping the causal signal (the part that explains how different users respond to different ads).

Methods

The authors propose DCRMTA (Deep Causal Representation for MTA), which consists of three main components:

  1. Causal Journey Representation: This module uses an LSTM with a hierarchical attention mechanism to process the sequence of ads a user sees. To ensure the model doesn't just "memorize" biased paths, they incorporate a Gradient Reversal Layer (GRL). This sets up an adversarial game where the model is forced to learn representations that are useful for prediction but invariant to the specific sequence distribution, effectively "unbinding" dynamic confounding variables.
  2. Causal User Representation: This is the core innovation. It uses a Causal Attention Module (CAM) to perform counterfactual reasoning. The model generates "fake" (counterfactual) attention maps by zeroing out or randomly inserting noise into user features. By comparing the "real" prediction with the "fake" one, the model can mathematically isolate the true causal effect of specific user attributes.
  3. Fusion & Attribution: The causal features from the journey and the user are fused to predict the final conversion. Once the model is trained, it uses Shapley Values—a concept from cooperative game theory—to fairly distribute credit among the various advertising channels.

Datasets

The researchers evaluated DCRMTA on two primary datasets:

  • Synthetic Dataset: A controlled environment (originally proposed in the CausalMTA paper) used to measure how well the model eliminates known biases.
  • Criteo-custom: The authors derived this from the famous Criteo-Raw dataset. They applied a heuristic filtering algorithm to create a more balanced dataset with 12 distinct ad campaigns and over 547,000 users, making it more representative of complex real-world MTA problems.

Key Insights and Ideas

  • Don't Throw the Baby Out with the Bathwater: Unlike previous "debiasing" methods that stripped away user data, DCRMTA recognizes that user attributes (age, gender, history) are essential for understanding why an ad works. The goal is to filter the data, not delete it.
  • Counterfactuals via Perturbation: The use of "fake" attention maps is a clever way to simulate "what if" scenarios. By perturbing the attention weights, the model learns which features are actually driving the conversion versus those that are just along for the ride.
  • Adversarial Unbiasing: Using a GRL to create "invariant representations" is a powerful transfer from the domain of domain adaptation. It ensures the model's logic holds up even when the underlying distribution of ads changes.

Key Results

  • State-of-the-Art Performance: On the synthetic dataset, DCRMTA achieved an AUC of 0.8009, outperforming the previous best model (CausalMTA) by roughly 2.6%.
  • Real-World Accuracy: On the Criteo-custom dataset, it maintained the highest AUC (0.7991) and the lowest Cross-Entropy Loss (0.1489) compared to baselines like DNAMTA and Logistic Regression.
  • Economic Impact: In a "Data Reply" simulation (reallocating budget based on the model's suggestions), the authors found they could achieve 32% of the original conversions using only 10% of the original cost, demonstrating massive potential for marketing efficiency.
  • Ablation Success: Removing the Causal User Representation caused the AUC to drop from 0.8009 to 0.7450, proving that their method of handling user features is the primary driver of the model's success.

Limitations and Future Work

The authors identify a few areas for improvement:

  • Computational Intensity: Calculating causal effects and counterfactuals during backpropagation is computationally expensive, which might limit real-time deployment for massive datasets.
  • Data Sparsity: In real-world settings, many users only have a single touchpoint. The temporal (LSTM) part of the model struggles to provide extra value in these "short-sequence" cases.
  • Future Directions: The authors suggest exploring more sophisticated sampling methods to handle the extreme imbalance between "converted" and "non-converted" journeys in typical marketing data.

Summary updated

Abstract

Multi-touch attribution (MTA) currently plays a pivotal role in achieving a fair estimation of the contributions of each advertising touchpoint to-wards conversion behavior, deeply influencing budget allocation and advertising recommenda-tion. Previous works attempted to eliminate the bias caused by user preferences to achieve the unbiased assumption of the conversion model. The multi-model collaboration method is not ef-ficient, and the complete elimination of user in-fluence also eliminates the causal effect of user features on conversion, resulting in limited per-formance of the conversion model. This paper re-defines the causal effect of user features on con-versions and proposes a novel end-to-end ap-proach, Deep Causal Representation for MTA (DCRMTA). Our model focuses on extracting causa features between conversions and users while eliminating confounding variables. Fur-thermore, extensive experiments demonstrate DCRMTA's superior performance in converting prediction across varying data distributions, while also effectively attributing value across dif-ferent advertising channels.
BibTeX
@misc{tang2024dcrmta,
  author = {Jiaming Tang},
  title = {{DCRMTA: Unbiased Causal Representation for Multi-touch Attribution}},
  year = {2024},
  eprint = {2401.08875},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.08875},
}
Figure