Conversion Prediction in User Click Streams

In the field of online marketing online-marketing, the goal is to maximize the number of conversions. Here, a conversion means that a user, who is visiting a website, for example, becomes a client who purchases anything. Being able to predict whether a current user is likely to convert or not would provide a valuable tool for marketers. In this thesis, we investigated several different machine learning models for this task, namely: Association Rules, Markov Models, kNN-prediction models, tree classifiers, and RNN to solve the tasks of predicting the next action for a current sequence of events and classifying an unfinished session to whether this session will become a conversion session or not. All tests are done on two separate data sets. The first one contained sessions collected on hotel websites, and the other one was the publicly available Retail Rocket data set. It turned out that the RNN models had the best overall performance, however, the tree classifiers and kNN could also achieve good results with shorter training and prediction times.