2003 · 29 citations · 0 references
ComputingComputer ScienceData MiningPredictive AnalyticsWeb AnalyticsCustomer Behavior AnalyticsCustomer SegmentationOnline Consumer BehaviorVisitor ClassificationWeb MiningIncremental ProcessingProbability EstimationPurchase PredictionSite VisitsUsage-based Customer ProfilingWeb Content MiningWeb Log MiningWeb Usage Mining
We explore methods for dynamic classification of visitors to an e-commerce web site based on visit sequences of page accesses. The time aspect is important in the processing of such data, and we require techniques that yield information before a customer's full sequence is realized. Further, we recognize that the timing of classification decisions may be important. We focus on prediction of purchases based on site navigation paths, and explore two related problems. The first is incremental estimation of purchase probabilities. We develop a probability estimation model based on mixtures of Markov chains, and develop several extensions. Second, we consider dynamic classification of visits into “buy” and “non-buy” classes. We assume that at each click the merchant has three options: classify the visit as a “buy” visit, classify the visit as a “non-buy” visit, or await further information to be revealed. We examine dynamic decision rules-derived using dynamic programming-for generating these classifications from estimated probabilities, and compare them to schemes based on fixed probability thresholds. We illustrate our methodologies on a real web log data set from a large retailer of computers. We demonstrate that probability estimation models based on second- and higher-order transition information outperform models of lower order. We show that both the fixed thresholds and the dynamic decision rules outperform a simple classification heuristic, and can be tuned to trade off the speed and accuracy of detection of both purchase visits and non-purchase visits.