CatBoost-based context-aware purchase-decision prediction on tabular customer journey sequences

Abstract

Building on the premise that real-time marketing decisions require models that are both accurate and resource-efficient, our study proposes a context-aware, tabular pipeline using CatBoost with ordered target encoding and greedy feature combinations to learn from customer journey sequences. In feature engineering, we derive period-level aggregates such as touch counts and mean dwell times across daily, weekly or monthly intervals, compute channel-usage entropy to measure diversity, and quantify recency relative to the most recent interaction, capturing intensity and diversity of user behavior. These sequences are further enriched with context signals such as dwell time, device type, temporal gaps and channel entropy, along with demographic attributes, to provide a rich representation of customer behavior. We apply the method to the Netherlands Travel dataset (May 2015–Oct 2016), trimming each journey to the first ten and last twenty touchpoints, engineering row-level and period-level aggregates, and collapsing them to the purchase level. CatBoost, a gradient boosting algorithm, leverages ordered target encoding and greedy feature combinations to capture nonlinear interactions while remaining interpretable and CPU-friendly. Stratified five-fold cross-validation yields an AUC of 0.9479, and tuning the F0.5 = 0.869 score produces a threshold of 0.723 with precision 0.897, recall 0.773 and accuracy 0.895. These results demonstrate that structured tabular representations of customer journeys, combined with tree-based learning, can achieve strong predictive performance while maintaining interpretability and computational efficiency, offering a practical modeling framework for decision-support systems in travel marketing.