An Intelligent Conversion Modeling Approach Leveraging Machine Learning and Deep Learning to Predict User Behavior from Clickstream Data

Abstract

The rapid expansion of hyperlocal e-commerce platforms has generated vast amounts of clickstream data that capture complex consumer navigation and purchasing behaviors. The capability to convert these multi-dimensional behavioral cues into credible forecasts of purchase desire is becoming a central requirement in modern studies to the planning of uniquely tailored marketing intervention, real-time pricing models, and smart recommendation systems. The presented manuscript presents an extensive ensemble machine-learning model that will predict consumer buying behaviour, based on 137 different behavioral attributes derived from 1,000 customer sessions in various hyperlocal e-commerce platforms. The methodology framework also includes a collection of state-of-the-art algorithms, such as Cat Boost and XG Boost, LightGBM, the TabNet deep-learning model and a weighted ensemble of AutoGluon, which combines the merits of each of the underlying learners. A systematic ablation experiment with twelve feature types demonstrated that intent-based predictors, namely cart additions and checkout business, provide a predictive signal of the order of 8.71%, which is approximately thirty times as much as that derived by demographic variables alone. CatBoost had a marginally better discriminative capacity with a ROC-AUC of 0.9710, while the weighted ensemble had the highest classification performance, with an accuracy of 91.0% and an ROC-AUC of 0.9663. The model was robust, and XG Boost provided an accuracy of 92.80% with a standard deviation of 1.91%. The feature-importance diagnostics has detected the Reached_Checkout (importance=0.291), Added-To-Cart (0.019) and CartAdditions (0.015) behavioral predictors as the most important ones. Altogether, these results highlight the importance of the dynamic signals of engagement as opposed to more traditional demographic factors in predicting the success of the purchase and provide a practical set of recommendations that can be implemented to reduce the rate of cart abandonment and increase the rate of conversion in hyperlocal e-commerce platforms.