The ‘traditional’ machine learning models rely on a tabular input that is feature engineered. This means that we, as researchers, decide what gets turned into a feature. In these cases features could be: amount of homepages visited, amount of searches done, total amount of pixels scrolled. However, it is very difficult to capture the spatial (time) dimension when doing feature-engineering. By using deep learning and embedding layers we can efficiently capture this spatial dimension by supplying a sequence of user behavior (as indices) as input for the model.
…g. They allow us to capture relationships in language that are very difficult to capture otherwise. However, embedding layers can be used to embed many more things than just words. In my current research project I am using embedding layers to embed online user behavior. In this case I am assigning indices to user behavior like ‘page view on page type X on portal Y’ or ‘scrolled X pixels’. These indices are then used for constructing a sequence of user behavior.