NNN: N ext-Generation N eural N etworks for Marketing Measurement

Thomas Mulc, Mike Anderson, Paul Cubre, Huikun Zhang, Ivy Liu, Saket Kumar · 2025

Read NNN: N ext-Generation N eural N etworks for Marketing Measurement by Thomas Mulc, Mike Anderson, Paul Cubre, Huikun Zhang, Ivy Liu, Saket Kumar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-06212.txt.

Introduction

Thomas Mulc and co-authors present, in twenty twenty-five, a neural network method to measure marketing effect. Classic marketing mix models use single numbers per channel and fixed decay curves for ad effects. Their method, called NNN, uses rich data packs that hold both numbers and quality traits, such as search terms and ad designs. An attention mechanism, which lets the model weigh what matters, can then catch complex links between channels and long-term effects.

A penalty that favors sparse models lets this rich model work on small data sets, which are the norm in marketing. Tests on simulated and real data showed strong gains in prediction. Probing the trained model can also rate keywords and ad designs.

Catalog record

CREATORThomas Mulc, Mike Anderson, Paul Cubre, Huikun Zhang, Ivy Liu, Saket Kumar
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE95 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-06212

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NNN: N ext-Generation N eural N etworks for Marketing Measurement Thomas Mulc, Mike Anderson, Paul Cubre, Huikun Zhang, Ivy Liu, Saket Kumar Abstract 1 Abstract We present NNN, an experimental Transformer-based neural network approach to marketing measurement. Unlike Marketing Mix Models (MMMs) which rely on scalar inputs and parametric decay functions, NNN uses rich embeddings to capture both quantitative and qualitative aspects of marketing and organic channels (e.g., search queries, ad creatives). This, combined with its attention mechanism, potentially enables NNN to model complex interactions, capture long-term effects, and improve sales attribution accuracy. We show that L1 regularization permits the use of such expressive models in typical data-constrained settings. Evaluating NNN on simulated and real-world data demonstrates its efficacy, particularly through considerable improvement in predictive power. In addition to marketing measurement, the NNN framework can provide valuable, complementary insights through model probing, such as evaluating keyword or creative effectiveness. 2 Introduction In today’s data-driven marketing landscape, accurately attributing sales to specific campaigns remains a paramount challenge. Marketers grapple with quantifying the return on investment for their marketing spend, seeking reliable estimates of the sales uplift generated by each channel. While randomized controlled experiments offer causal rigor, they are often impractical or prohibitively expensive to implement across a full media mix and typically cannot predict effectiveness at untested spend levels. Marketing Mix Models (MMMs), with roots tracing back to the 1950s [ 19 , 4 ] , have seen a resurgence in recent years [ 7 , 23 , 13 , 45 ] , offering a means to estimate the effectiveness of diverse media channels using aggregated, observational data. This renewed interest is partly fueled by the increasing difficulty of long-term tracking for attribution and user-level experiments [ 32 ] , especially given the deprecation of third-party cookies and stricter privacy regulations like GDPR, CCPA [ 6 ] and Apple’s Intelligent Tracking Prevention (ITP) [ 1 ] . MMMs circumvent these tracking needs by operating on aggregated data (e.g., weekly impressions and sales per region). However, current MMM methods face well-documented limitations [ 7 , 33 ] . First, MMM datasets are often limited in samples relative to the number of factors influencing sales, which can lead to statistically overdetermined models [ 7 ] which has historically limited the feasibility of employing more complex, data-hungry techniques like deep neural networks. Second, the common assumption 1 1 1 There are a few exceptions such as [ 30 ] . of time-invariant response curves makes it challenging to account for variations in marketing effectiveness due to changes in ad creative quality, targeting strategies, or external factors like seasonality 2 2 2 Such as changes in effectiveness during the holiday shopping season. . Finally, capturing long-term brand effects is notoriously difficult; parametric adstock functions typically used to model carryover effects often lose identifiability from the baseline beyond a few months. To address these challenges, this paper introduces NNN (Next-Generation Neural Networks for Marketing Measurement), a novel paradigm leveraging recent advancements in deep learning and representation learning. This is an experimental approach which uses a different foundational modeling method compared to current MMM methodology. Some key differences include: • Rich Data Representation. Instead of relying on scalar inputs (e.g., spend), NNN utilizes high-dimensional embeddings to represent both marketing activities (e.g., Search Ads, YouTube Ads) and crucial organic signals (e.g., Google Search queries). These embeddings capture not only quantitative volume but also qualitative aspects like ad creative content or search query semantics, allowing the model to directly address the challenge of time-varying creative effectiveness. • Modeling Long-Term Effects via Intermediary Signals. NNN explicitly models the influence of marketing activities on intermediary signals highly predictive of consumer behavior, specifically Google Search query patterns (as depicted in Figure 1 ). This, combined with a Transformer architecture capable of attending to long historical sequences, provides a mechanism for capturing marketing effects that unfold over extended time horizons, potentially overcoming the limitations of standard adstock models. • Neural Network Power with Regularization. The framework employs neural networks and a custom Transformer to learn complex non-linear interactions between channels and over time. We demonstrate that L1 regularization makes it feasible to train these highly parameterized models effectively, even within the often data-constrained settings typical of MMM, mitigating concerns about overfitting. • Objective Model Selection. NNN facilitates model selection based on rigorous evaluation of predictive performance (e.g., MAPE, R 2 ) across multiple held-out datasets, particularly temporally distinct test sets, providing a robust alternative or complement to traditional criteria based on parameter plausibility. Our paper is structured as follows. We give an overview of related working influencing our approach (Related Work). Next, we discuss the data used (Datasets), and how they are formatted in our method (Data Structure). We give an overview of the computation of the model (Model Overview), then go into further details regarding the how traditional data is augmented and embedded (Input Data Representations). We discuss the intricacies of our Transformer (Transformer) and details about how the model makes the final predictions (Prediction Heads). We discuss training details (Model Training) and attribution methods (Sales Attribution). We give an overview of our experiments and then discuss the results and analysis (Experiments). Finally, we give a discussion about how our method compares to existing measurement approaches (Discussion), and give a concluding overview (Conclusion). Figure 1 : High-level causal structure in our models. Red arrows represent the links from organic Search to sales, modeled in [ 28 ] . Yellow arrows represent new links from previous search to current sales, which are implemented through our Transformer. Blue arrows represent the new causal links between marketing and sales. Dashed arrows represent the links the search predictive component of the model. Note that this component is autoregressive, whereas the sales is not. More complicated causal structures are possible with our framework as well, but mostly outside the scope of this work. 3 Related Work Our proposed NNN framework builds upon and diverges from several strands of research in marketing measurement, primarily Media Mix Modeling, the application of neural networks in this domain, and the use of embeddings to represent media signals. This section situates NNN within this existing literature. 3.1 Media Mix Modeling Evolution The field of Media Mix Modeling (MMM) provides the foundational context for NNN, offering established methods for assessing advertising ROI using historical, aggregate data [ 19 , 4 ] . Recognizing limitations in early linear models, Bayesian Media Mix Modeling (BMMM) emerged, introducing techniques to model ad carryover and saturation (shape effects) more flexibly, often using Hill and adstock functions [ 23 ] . BMMM leverages Bayesian inference (e.g., MCMC methods) to incorporate prior knowledge into model estimation, a crucial aspect given the often data-constrained nature of MMM problems. This Bayesian framework was further extended to handle geographical variations through hierarchical modeling [ 34 ] . Meridian, a Google Bayesian MMM framework integrated advancements of MMM research over the past decade. It incorporates features like ROI priors and calibration [ 44 ] , reach and frequency data integration [ 43 ] , and paid search modeling. While these approaches are powerful and innovative, here we are exploring an entirely new research direction that does not rely on scalar inputs or

Read the complete text