A map-based gender prediction model for big E-commerce data

Gender differences have been found in costumer behavior and perceptions of the benefits, concerns and important attributes of e-commerce, although related information is in general not available. In this paper, a map based gender predictive model is proposed for predicting the gender types of e-commerce participants from their product viewing records by exploiting the hierarchy of the online products classification structure at several hierarchy levels from top categories down to individual products. In our method, the gender types are estimated by mapping the category transition probability of viewed products based on the assumption that the transition process satisfies first-order Markov property. A comparative adjustment is applied for probability smoothing in order to avoid discontinuity with zero values in probability estimates. An iterative learning scheme with incremental data training by adding 'reliable' testing data instances into training in an iterative way is developed to strengthen the model learning and improve the robustness of model predictions. The proposed method was evaluated by using the data provided in PAKDD'2015 Data Mining Competition. The balanced accuracy we achieved in the 1st iteration is 80.73%, and finally through ten iterations of learning process, is increased to 81.07% scoring the 10th place out of 330 international teams' submissions.