Details of Research Outputs

Web sites and online services increasingly engage with users through live chats to provide support, advice, and offers. Such approaches require reliable methods to predict the user’s intent and make an informed decision when and how to intervene during an active session. Prior work on predicting purchase intent involved clickstream data mining and feature construction in an ad-hoc manner with a moderate success (AUC 0.70 range). We demonstrate the use of the consumer Purchase Decision Model (PDM) and a principled way of constructing features predictive of the purchase intent. We show that the Logistic Regression (LR) classifiers, trained with multi-action motifs, perform on par with the state-of-the-art LSTM sequence model achieving comparable AUC (0.95 vs 0.96) and performing better for the sparse purchase sessions, with higher recall (0.85 vs 0.61) and higher F1 score (0.73 vs 0.66). While LSTM performs better than LR in terms of weighted averages of F1, precision, and recall, it requires 7 times longer to train and offers no insights about the predictive model in terms of the user actions and the purchase decision stages. The LR predictors are robust and effective in simulating real-time interventions, achieving F1 of 0.84 and AUC of 0.85 after observing only 50% of an active session. For non-purchase sessions that leaves room for live intervention, on average within 8 actions before the session ends.