Predicting Future User Activities with Constrained Non-Linear Models

Prediction of future user activities from their history, all past activities, is a challenging problem. One reason is that the number of all potential histories grows exponentially with the length of the 4 history. Recently, deep-learning models have been proposed for solving this problem. It is easy to learn a simple predictor of future user activities, by averaging all past activities of the user and then learning an activity classifier from this average representation, for instance by logistic regression. This approach tends to have a high bias, due to using a simple feature representation and model. It is also easy to apply sequence models in deep learning to learn a predictor of future user activities. This approach tends to have a high variance, because of the large number of parameters in the neural network.