Implementing co-visitation networks to detect fraud in online advertising

This bachelor thesis presents an open-source implementation of the co-visitation networks technique, first presented by m6d Research [1], which helps identifying internet domains that contain Non-Intentional Traffic. This technique is presented in two different platforms: a Python module and a Jupyter notebook version. This implementation was tested and then applied into a real dataset. First, this data had to be collected and filtered. Then selected some of those records where selected to avoid problems, due to its original size volume. Finally, after applying the mentioned implementation into a real data set, some results could be obtained after the outputs were transformed into graphs. Among these results, it was possible to differentiate two categories: the popular websites that had a large volume of visitations due to their usual high traffic, and the fraudulent websites that overlapped in visits with other non-related websites. Cases belonging to the latter category were difficult to observe manually as data was collected two years ago and most of the plausible fraudulent websites were already closed. After analyzing these results, it could be assured that the technique seemed to work better when working with subsets of smaller time frames rather than larger ones. The reason behind this appears to be that the manual filtration of thousand of potential fraudulent sites mixed up with the visits to popular websites is more tedious than the few dozens obtained in smaller time frames. Whatsoever, this technique’s objective is to detect users patterns in the internet and how domains are related with each other.Ingeniería Informátic