Neural Network Survival Analysis

Preface I wish to thank my promoter Prof. Dirk Van den Poel who gave me the opportunity to take this thesis subject and guided me to pave through these uncharted areas. My thanks specially go to my co-promoter-Prof. Els Goetghebeur. Her suggestions and advices finally made me grasp the key to the gate of hope for my thesis. I'd also like to thank Jozefien Buyze. Thank you for sharing knowledge with me and encouraging me when I ran into problems that look overwhelming. Also many thanks go to my friends and family for keeping me looking forward, particularly to my one year old. Although you cannot talk yet, your smile is the most powerful source of strength for me. I also like to thank my husband for sharing the burden on both my thesis work and housekeeping. Summary In this study, we analyzed a data set from real commercial data on the purchase behaviors of 168 customers to predict the next purchase time. The data were grouped to the training set and test set, and analyzed by a piecewise standard Cox PH model, a piecewise marginal Cox model and the PLANN neural network approach. The effects of the following five factors were studied: the previous purchase interval, the type of a customer, the region and size of the city where a customer lives, and the season of the last purchase. The three models (two Cox's PH models and the ANN model) were used to predict the survival of the test set. In total eight subgroups of the test set were selected and their predicted survivals were compared to the KM survival estimates. The comparison shows that the ANN methods displayed similar predictability performance with that of the piecewise standard Cox PH model. Thus, the hypothesis that the ANN method is superior to the conventional Cox's PH models does not valid. The study reveals the following patterns in the purchase behaviors: 1) the next purchase interval approximately proportional to previous interval， while the output of the marginal Cox model indicates that for a customer the marginal effect of the previous interval on the next purchase interval is not significant. 2) The purchase interval of a customer living in big or medium city is not significant different with that of a customer in small or tiny city. 3) The customers whose types are 'catering' or 'horeca' have similar purchase trends and have …