Predictive Customer Journey Intelligence: An Explainable Machine Learning Framework for Modeling Dynamic Omnichannel Consumer Behavior

Abstract

The growth of omni-channel retail environments has made the consumer decision making process more complicated, dynamic and interconnected, with continuous interactions across various digital and physical touch points. Existing customer journey models and customer attribution methods are generally based on using a fixed structure of channels, defined customer attributes and not sufficiently understanding the sequential dependencies. In this study, Predictive Customer Journey Intelligence (PCJI) is presented, an explainable machine learning framework to tackle these limitations to model dynamic omnichannel consumer behavior. The proposed framework combines temporal behavioral modeling, graph-based touchpoint representation, and explainable artificial intelligence to model changing customer interactions and forecast behavioral changes between nodes in the graph representing the complex structure of retail channels. Specifically, a hybrid Temporal Graph Attention–Long Short-Term Memory (TGA-LSTM) architecture is designed to capture temporal dependencies, next-touchpoint transition, channel-switching probabilities and purchase conversion. This framework is assessed on an empirically simulated sample of 48,600 customer journeys from six ‘omnichannel touchpoints' (in-store, mobile apps, websites, social commerce, online marketplaces and call centres) observed over one and a half years. It is evaluated against three models: Markov chain attribution, logistic regression, and extreme gradient boosting (XGBoost) as well as against traditional LSTM models. The proposed TGA-LSTM framework achieves an area under the receiver operating characteristic curve (AUC) of 0.934 for predicting purchase conversion, which is a reported 9.7-percentage-point higher than XGBoost, and the cross-entropy loss for predicting next-touchpoint is reduced by 21.4% over the best sequence-only baseline. The combination of attention mechanisms and Shapley-based explainability facilitates interpretable explanations for the relative impact of behavioural, temporal, and channel-specific aspects of the customer's journey on individual predictions, allowing for dynamic analysis of customer preferences and changes in journey stages. The results show the potential of XAI to improve the understanding of the customer journey and incorporate temporal dependence. Relationships between touch points that are interconnected and under a single predictive framework. Theoretically, the study adds by conceptualizing omnichannel consumer behaviour as a dynamic. Temporally Changing Customer Interaction Network; Methodologically by proposing a hybrid deep-learning approach to predictive modeling of customer journeys;. Managerially, by providing data-driven customer segmentation, personalization and adaptive marketing resource allocation. But the findings are not directly generalizable to actual retail contexts, due to using simulated data. The framework should be validated in future studies with longitudinal. Real world consumer data and explore causal inference, cross-industry applicability and privacy preserving machine learning, to enable responsible and scalable adoption of predictive customer journey intelligence. Keywords: Customer journey analytics, omnichannel marketing, deep learning, temporal graph networks, attribution modelling, consumer behaviour prediction