CausalMMM: Learning Causal Structure for Marketing Mix Modeling

The Paper

"CausalMMM: Learning Causal Structure for Marketing Mix Modeling" was published in 2024 by Chang Gong, Di Yao, Lei Zhang, Sheng Chen, Wenbin Li, Yueyang Su, and Jingping Bi, all affiliated with the Institute of Computing Technology, Chinese Academy of Sciences and the University of Chinese Academy of Sciences. It appeared at WSDM 2024, the ACM International Conference on Web Search and Data Mining, one of the leading venues for applied machine learning on large-scale commercial data.

The paper targets Marketing Mix Modeling (MMM) - the quantitative practice of attributing sales outcomes to advertising channel investments. MMM underpins budget allocation decisions at scale across e-commerce, retail, and brand advertising. The core problem: every prior approach to causal MMM requires a human to pre-specify which advertising channels causally affect which others. CausalMMM replaces that manual specification with a learned, data-driven process that discovers shop-specific causal structures automatically while simultaneously modeling the marketing response dynamics (carryover and saturation) that make advertising attribution non-trivial.

Read the Paper on arXiv →

The Problem Before This Paper

Marketing Mix Modeling as a discipline is decades old. Early MMM was purely regression-based: fit a linear model from advertising spend to sales, read off the coefficients as attribution weights. The obvious limitation is that regression does not distinguish correlation from causation, and advertising channels frequently confound each other - a brand awareness TV campaign shifts the baseline for search-click conversions in ways that regression cannot cleanly separate.

The causal MMM line of work addressed this by introducing Directed Acyclic Graphs (DAGs) over advertising channels to model how spend on one channel influences outcomes through others. But every prior causal MMM method shared a structural assumption: the causal graph is pre-specified. A domain expert or analyst decides in advance that, say, TV causes Search-Click and Search-Click causes GMV. The model then conditions on that graph.

This creates two compounding problems that practitioners building production MMM systems deal with constantly:

• Pre-specified graphs are wrong, and wrongly-specified graphs degrade prediction. The actual causal relationships between advertising channels differ by brand category, shop maturity, seasonality, and competitive context. A graph built for a luxury cosmetics brand does not transfer to a grocery shop on the same platform.
• Causal heterogeneity across shops is not handled. An e-commerce platform with thousands of brand shops cannot afford to manually specify a separate causal graph for each one. Existing approaches either use a single shared graph (wrong) or require per-shop manual specification (unscalable).

Granger causality methods - including NGC, GVAR, and the attention-based InGRA - had been developed for causal discovery in time series, but none of them incorporated marketing-specific constraints such as the saturation effect (diminishing returns on ad spend) or carryover (advertising effect that decays over time), which are structural features of how advertising actually works. Applying those methods directly to marketing data produces causal structures that are statistically valid but practically unreliable for budget allocation.

What They Built

CausalMMM is a variational inference framework with two coupled components: a Causal Relational Encoder that discovers causal structures across advertising channels, and a Marketing Response Decoder that predicts sales outcomes while enforcing the saturation and carryover dynamics known to govern advertising response.

The problem is formally defined as follows. Given a marketing dataset with N shops, each with advertising spend matrix X_n ∈ ℜ^T×d (T time steps, d channels), target values y_n ∈ ℜ^T, and contextual variables C_n, the model must simultaneously infer shop-specific causal graphs {G₁,...,G_N} and predict future GMV. The key formulation uses non-linear Granger causality: a channel x_j Granger-causes channel x_i if access to the history of x_j improves the prediction of x_i over using x_i's own history alone.

x_j^(t+1) = g_j(x₁^(1:t), ..., x_d^(1:t)) + ε_j^(t+1)

The causal adjacency matrix is treated as a discrete latent variable z ∈ {0,1}^d×d, where z_ij=1 indicates that channel i causally influences channel j. Because discrete variables are not differentiable, the model uses Gumbel-Softmax sampling to enable end-to-end gradient-based training:

z_ij ~ Softmax( (h_ij + g) / τ )

where g is Gumbel noise and τ is a temperature hyperparameter controlling how close the samples are to one-hot. The encoder produces pairwise edge embeddings h_ij via a graph neural network operating over channel histories, then samples discrete edges from this relaxed distribution.

The decoder takes the sampled causal graph and propagates information through it to model temporal dynamics. For each channel j, a message-passing step aggregates hidden states from causally upstream channels weighted by the graph edges:

MSG_j^t = ∑_i z_ij · h̃_ij^t     h̃_j^(t+1) = f_seq(MSG_j^t, h̃_j^(1:t))

The saturation module applies a learned S-curve transformation to the temporal output, modeling the diminishing returns structure of advertising spend. The shape parameters α and γ are themselves functions of the contextual variables C_n, so saturation behavior is heterogeneous across shops:

Saturation(x) = x^α / (x^α + γ)     where α = f_α(C_n), γ = f_γ(C_n)

The full model is trained by optimizing a variational lower bound that jointly maximizes predictive accuracy and regularizes the learned causal structure:

L = E[log p_θ((X, y) | z)] - λ · KL[ q_φ(z | (X, y)) || p(z) ]

The KL term penalizes deviation from a sparse prior over the causal graph, biasing the model toward recovering compact structures. The penalty factor λ controls this trade-off: too small and the model learns dense uninformative graphs; too large and it collapses to no edges. Training complexity scales as O(WT) where W is the number of parameters and T is the sequence length - linear in data size, which matters for the platform-scale deployment this work targets.

Key Findings

• Automatic causal structure discovery outperforms expert-specified graphs. On the AirMMM real-world dataset, learned structures produced better GMV predictions than baselines that use pre-defined causal assumptions, across all forecasting horizons tested.
• 5.7%-7.1% AUROC improvement over best causal discovery baseline. CausalMMM achieves 0.935 AUROC on Simulation 1 versus 0.863 for InGRA, the strongest prior method, on the causal structure recovery task over synthetic ground-truth data.
• Marketing-specific constraints matter for causal discovery. Ablations show that removing the saturation module (CM-rw) degrades GMV prediction from MSE 1.80 to 3.09 at M=7 steps - a 71.7% increase in error - even when the causal structure is otherwise correct.
• Heterogeneity modeling is robust across different numbers of latent structures. At R=10 latent shop-type clusters, CausalMMM achieves AUROC 0.925 versus 0.865 for InGRA, maintaining the performance gap as structural complexity increases.
• Performance is stable as channel count and sequence length scale. At d=20 advertising channels and T=720 time steps, the model maintains 0.941 accuracy - a configuration that approximates real deployment conditions.
• Gumbel-Softmax temperature requires calibration. At very low temperatures, samples approach one-hot and gradient flow collapses. The model is sensitive to τ but not fragile: a moderate temperature range yields consistent results.

Results

The evaluation covers two tracks: causal structure recovery on synthetic datasets with known ground-truth DAGs, and GMV prediction on the real-world AirMMM dataset from an e-commerce platform (50 shops, 12 advertising channels, 5 contextual variables, 676 days of observations from January 30, 2021 to December 6, 2022).

Causal Structure Learning (Simulation 1 / Simulation 2):

• CausalMMM AUROC: 0.935 ± 0.012 / 0.903 ± 0.015
• InGRA (best prior method) AUROC: 0.863 ± 0.082 / 0.854 ± 0.083
• GVAR AUROC: 0.821 / 0.800 (approximate from reported figures)
• NGC AUROC: below 0.80 in both simulations
• Improvement over InGRA: 5.7% on Sim 1, 7.1% on Sim 2, with substantially lower variance - the error bars on InGRA (±0.082) are nearly 7x wider than CausalMMM (±0.012)

GMV Prediction on AirMMM (MSE, lower is better):

• M=1 step: CausalMMM 0.29 vs BTVC 0.47, GVAR 0.37, LSTM 0.52
• M=7 steps: CausalMMM 1.80 vs BTVC 2.07, GVAR 1.91
• M=30 steps: CausalMMM 9.55 vs BTVC 9.43, GVAR 12.41

At M=30 steps, BTVC slightly outperforms CausalMMM (9.43 vs 9.55 MSE). The authors note this is expected: at very long horizons, Bayesian temporal models with strong structural priors can outperform learned approaches where distributional shift accumulates. The gap is narrow and does not affect the overall ranking.

Ablation Study (M=7 GMV MSE):

• CM-full (causal encoder removed, no graph): MSE 2.65
• CM-markov (temporal carryover module removed): MSE 2.85
• CM-rw (saturation module removed): MSE 3.09
• Full CausalMMM: MSE 1.80

The ablation ordering is notable: saturation contributes more to prediction quality than carryover, and carryover contributes more than the causal graph itself in the short-term prediction window. The causal structure's primary value is in medium-to-long horizon prediction and in producing interpretable attribution for budget allocation.

Why This Matters for AI and Automation

This paper sits at the intersection of two fields that rarely collaborate directly: causal discovery and applied marketing science. The implications extend well beyond MMM.

• Automated causal graph learning for business data. The CausalMMM framework generalizes to any setting where multiple correlated time series jointly cause a target outcome and the causal structure is unknown or heterogeneous. Supply chain demand modeling, multi-touch attribution in digital analytics, and operational efficiency optimization are all structurally similar problems.
• Heterogeneous structure learning at scale. The variational inference approach allows the model to share statistical strength across shops while recovering shop-specific graphs. This is a pattern applicable to any large-scale platform where entities share domain structure but differ in specifics - financial risk modeling, healthcare patient subtyping, recommendation systems with segment-level treatment effects.
• Interpretability with quantifiable constraints. The learned causal graph is a direct output of the model, not a post-hoc explanation. Budget allocation decisions derived from these graphs are auditable: you can inspect which channel relationships the model recovered and whether they align with domain knowledge.
• Variational inference + discrete structure learning as a design pattern. The Gumbel-Softmax trick for learning discrete latent structures end-to-end - without relying on REINFORCE or other high-variance gradient estimators - is directly applicable to neural architecture search, program synthesis, and any task requiring differentiable selection over discrete combinatorial spaces.
• Domain constraints improve data-driven models. The saturation and carryover modules are not generic regularizers - they encode prior knowledge about how advertising works. The ablation results show they contribute more to prediction quality than the causal structure alone. This is a reminder that for applied ML in structured domains, baking in domain constraints beats larger models.

My Take

The strongest contribution here is not the AUROC improvement over InGRA - it is the demonstration that causal structure learning and domain-constrained response modeling can be trained jointly within a single variational framework without either component degrading the other. Prior work in causal discovery treated structure learning and downstream prediction as separate stages. CausalMMM makes them co-dependent and shows that joint training strictly dominates.

The variance reduction is underappreciated in the headline numbers. InGRA achieves AUROC 0.863 with standard deviation 0.082. CausalMMM achieves 0.935 with standard deviation 0.012. For a system you are deploying to make budget allocation decisions at platform scale, the reliability of the causal structure matters as much as its average quality. A method that is occasionally right but frequently wrong about channel relationships is not deployable.

The M=30 horizon result (where BTVC slightly outperforms) is worth watching. The authors frame it as a known trade-off, but it points at a real limitation: learned causal structures are fit on historical patterns and do not extrapolate well when marketing dynamics shift seasonally or through campaign strategy changes. The Bayesian prior in BTVC provides robustness that a learned latent graph does not. A hybrid approach - strong structural priors informed by the learned graph, rather than either/or - seems like the logical next step.

One gap not addressed: the evaluation is on a single platform (the AirMMM dataset from one unnamed e-commerce operator). MMM applications span retail, CPG, financial services, and gaming verticals with very different channel mixes and saturation dynamics. Whether the contextual variable mechanism for heterogeneous saturation parameters generalizes across these verticals is an open question the paper does not answer.

Discussion question: CausalMMM learns shop-specific causal graphs but still relies on historical spend-to-outcome data to do so - meaning it cannot recover causal structure for new shops or new channels without observation history. Compared to approaches like DoWhy or Double Machine Learning that can handle experimental (A/B test) data directly, how should practitioners think about when to use learned causal discovery methods versus intervention-based causal estimation for marketing attribution?