TRANSFORMING CUSTOMER JOURNEYS: MODELING, ATTRIBUTION AND OPTIMIZATION

Abstract

When observing a sequence of customer interactions, it is important for a firm to understand how these interactions relate to key objectives–qualified customer leads, customer conversion events, or churn. Inspired by recent development in Large Language Models (LLM), in Chapter 1, I propose a transformer-based framework that models the time series of customer interactions in a way that resembles modeling a sequence of words in a sentence. I propose a heterogeneous mixture multi-head attention mechanism that captures the individual heterogeneity in the touchpoint effects. The proposed model accounts for the varying effects of an interaction when it occurs at different times and the complex interactions between the various touchpoints on other touchpoints and events. I present an application in the multi-channel marketing context where the firm observes sequences of customer-firm encounters across multiple touchpoints. I compare the proposed model to the existing models such as hidden Markov model, point process model and LSTM, and show that the model has superior performance over these models. Extensive simulations further establish the model’s superiority over competing approaches. Beyond multichannel marketing, the transformer-based model also has broad applicability in customer journeys across other domains.

In Chapter 2, I propose a multi-modal transformer that integrates journey data with marketing mix model. Emerging consumer privacy regulations, such as GDPR and CCPA, along with platform-level restrictions on tracking technologies, have significantly reduced the availability and completeness of individual-level data, thereby limiting the effectiveness of traditional multi-touch attribution approaches. At the same time, aggregate data on media spending and outcomes remain widely available, motivating renewed interest in marketing mix models (MMM). I develop a transformer-based framework that jointly models individual-level journeys and aggregate media spending. The model uses self-attention to capture temporal dependencies within each modality and cross-attention to link media spending with customer behavior, allowing it to learn how customer responsiveness varies with both past interactions and marketing inputs. The media encoder serves as a flexible ad-stock function, while the journey encoder captures dynamic customer states and cross-channel interactions. To address data fragmentation, I introduce a pattern-mixture approach to account for untracked customers and show that the model remains robust under partially observed journeys. In addition, the model can generate synthetic customer journeys under counterfactual media allocations, allowing firms to optimize media mix using policy gradient methods. I demonstrate the effectiveness of the proposed approach through simulations and an empirical application, showing strong predictive performance and substantial improvements in media allocation decisions.