Deep Learning Models Compared for Multi-Touch Attribution in Online Advertising

The digital advertising industry runs on a deceptively simple question: when a consumer clicks “buy,” which of the many advertisements they encountered along the way actually deserves the credit? A new comparative study, published in Information Systems Frontiers by researchers at Huazhong University of Science and Technology and Zhengzhou University, tackles that question with unusual rigor, evaluating nine multi-touch attribution (MTA) models—ranging from classical heuristics to cutting-edge deep learning frameworks—across a synthetic dataset and two large public datasets. The findings carry practical weight for an industry that poured hundreds of billions of dollars into internet advertising last year, and they arrive at a conclusion that is already turning heads among marketing scientists: attribution models that perform better on prediction tasks also produce better, more efficient budget-allocation decisions, and the very best of them share a common design philosophy borrowed from causal inference.

The challenge of attribution stems from the nature of modern consumer journeys. A typical conversion—a purchase, a sign-up, an app install—is preceded not by a single ad exposure but by an extended sequence of touchpoints: a display banner seen on a news site, a search engine ad clicked a day later, a social media impression, a retargeting email. Assigning conversion credit to these touchpoints is the core task of multi-touch attribution. For years the industry relied on crude rules such as “last touch” (all credit to the final ad) or “first touch,” alongside the historically persistent but data-driven approaches that model conversion paths statistically. The problem, as the authors led by Tao Xue, Panyu Zhai, Yanwu Yang, and Yanling Li emphasize, is that naive models conflate correlation with causation. Users who are already inclined to buy are exposed to more ads; a model that simply measures the co-occurrence of ads and conversions will systematically overestimate the effect of advertising on users who needed no persuasion—a distortion known in the literature as user preference confounding bias.

The study’s methodological contribution lies in its systematic, head-to-head comparison. The researchers assembled nine representative MTA models spanning three generations of technique. The first generation includes heuristic and probabilistic baselines such as last-touch attribution and survival-theory-based approaches, which assign credit using positional rules or hazard-rate-style decay rather than learned representations. The second generation comprises data-driven deep learning models, including attention-based architectures—dual-attention networks in which the model learns to weight each touchpoint in a user’s click sequence according to its inferred contribution—and frameworks built on Shapley values, the cooperative game-theory construct that distributes a “payout” among players according to their marginal contributions across all possible orderings. In deep MTA, the Shapley framework is typically approximated with recurrent or convolutional neural networks, such as temporal convolutional networks (TCNs), which process the ordered touchpoint sequence through stacks of dilated causal convolutions so that the model can capture temporal dependencies—how the effect of an ad exposure depends on what came before it and how much time has elapsed.

The third, and in this study the most consequential, generation is causal attribution modeling. Models such as CausalMTA, first presented at the KDD 2022 conference, attempt to eliminate confounding bias explicitly. The mechanism works roughly as follows: the network learns a representation of each user’s intrinsic preferences—the stable, unobserved traits that drive both their exposure to certain ads and their propensity to convert—and then applies a de-confounding transformation, conceptually analogous to inverse propensity weighting or back-door adjustment in causal graph theory, so that attribution scores reflect the causal contribution of each touchpoint rather than mere association with high-intent users. The study frames these architectures through two attribution mechanisms that recur across deep learning frameworks: attention, which assigns credit through learned weighting of the sequence, and Shapley values, which assign credit through marginal-contribution estimation. Figs. 5, 6, and 8 of the paper depict how these two mechanisms slot into the overall pipelines.

The experimental design is notable for its realism and breadth. The team used a synthetic dataset, where ground-truth attribution effects can be known by construction, and two widely used public benchmarks: the Criteo attribution modeling and bidding dataset, drawn from real-time-bidding display advertising logs, and the Ali-CCP dataset from Alibaba, which pairs click streams with rich post-click behavior sequences. Together these datasets span the scale and messiness of production advertising systems—millions of user journeys, sparse conversions, high-dimensional categorical features, and severely imbalanced click-to-conversion ratios. Performance was assessed along two axes. The first is conversion rate (CVR) prediction, the standard supervised task of estimating the probability that a user journey ends in conversion, evaluated with established metrics. The second, and more decision-relevant, axis is advertising effectiveness: whether the attribution scores produced by each model translate into better decisions, measured through downstream tasks such as budget allocation across channels.

The headline results are striking in their consistency. Across all three datasets, causal attribution models significantly outperformed their non-causal counterparts, precisely because they correct the confounding bias introduced by user preferences. When a model attributes credit only for the incremental effect of an exposure—what the ad actually changed about the user’s behavior, rather than what the user would have done anyway—the resulting attribution scores are both more accurate against synthetic ground truth and more stable on real data. Just as important for practitioners, the study finds a tight coupling between attribution quality and predictive quality: models that achieve better attribution results also deliver better CVR prediction. This alignment matters because in operational settings the same underlying representation often serves both tasks, and because accurate conversion prediction is a prerequisite for the real-time bidding systems that decide, in milliseconds, how much an impression is worth.

The connection to downstream decisions is where the study’s practical stakes become clearest. Attribution outputs feed directly into budget-allocation models: if a channel is credited with a larger share of conversions, it receives a larger share of the next quarter’s spend. The authors demonstrate that models with more accurate attribution lead to more efficient advertising decisions—budgets that track true incremental impact rather than spurious correlations. This finding echoes, and empirically sharpens, a long-standing warning in the marketing literature. Earlier work, including Berman’s analysis in Marketing Science and Danaher and van Heerde’s study of “delusion in attribution,” showed that attribution-based budget rules can go badly wrong when the attribution itself is biased. The new study provides a constructive answer: the delusion is not inevitable, but avoiding it requires models that are causal by design, not merely associative.

The technical details of the winning architectures repay attention. In the attention-based frameworks, a user’s touchpoint sequence is embedded into a dense representation, after which attention modules compute compatibility scores between each touchpoint and the conversion target; these scores are normalized into weights that both produce the attribution and inform the conversion prediction. In Shapley-based frameworks, the network estimates each touchpoint’s marginal contribution by comparing predicted conversion probabilities with and without the touchpoint in the journey—a combinatorially intractable computation in principle, made feasible through sampling-based or learned approximations. The causal models layer a de-confounding module on top: user preference embeddings are extracted and either reweighted or conditioned away, so that the attention or Shapley mechanism operates on a representation stripped of the confounding signal. The paper’s sixteen figures trace these pipelines in detail, and the supplementary appendix provides complete notation tables and summary statistics of clicks and conversions on the Criteo and Ali-CCP datasets, alongside feature-level association analyses that confirm the statistical significance of categorical predictors of conversion.

The study does not pretend that causal deep MTA is a solved problem. The authors note that the performance and improvement of deep learning-based MTA models reported in academia had, until now, been unclear—partly because papers evaluated different models on different datasets with different metrics, making genuine comparison impossible. By running nine models through a common evaluation harness on common data, the study supplies something the field has lacked: a controlled benchmark. It also highlights the resource constraints facing practitioners, since many of the strongest models are computationally demanding and their approximations of Shapley values introduce their own biases. The authors point toward future work on better counterfactual estimation, on integrating large pre-trained sequence models, and on extending attribution beyond clicks to the full spectrum of post-click behaviors that Alibaba’s dataset makes visible.

For an industry in the midst of a measurement crisis—driven by cookie deprecation, privacy regulation, and the collapse of cross-device tracking—the timing of this research is apt. Multi-touch attribution has long been criticized as fragile, dependent on individual-level identifiers that are disappearing. But the study’s central lesson transcends the identity-tracking debate: any attribution method, however the data is collected, will mislead advertisers if it fails to separate the causal effect of advertising from the pre-existing intentions of the users it reaches. The Huazhong-Zhengzhou team’s evidence that causal deep learning models close that gap—while simultaneously improving conversion prediction and budget efficiency—offers a technically grounded path forward. Advertisers, platforms, and the academic marketing community now have a benchmark against which the next generation of attribution models can be measured, and a clear signal that the causal revolution in machine learning has reached the advertising budget spreadsheet.