Using contextual information to decrease the cost of incorrect predictions in online customer behavior modeling

The performance of user profiling models depends on both the predictive accuracy and the cost of incorrect predictions. In this paper we study whether including contextual information leads to a decrease in the misclassification cost. Several experimental analyses were done by varying the cost ratio, the market granularity and the granularity of context. The experimental results show that context leads to a decrease in the misclassification cost under particular conditions. These findings have significant implications for companies that have to decide whether to gather contextual information and make it actionable: how deep it should be and which unit of analysis to consider in market research.