Thomas, Sherin and Srijith, P K
(2018)
A Bayesian Point Process Model for User Return
Time Prediction in Recommendation Systems.
Masters thesis, Indian Institute of Technology Hyderabad.
Abstract
In order to sustain the user-base for a web service, it is important to know the return time of a
user to the service. In this work, we propose a point process model which captures the temporal
dynamics of the user activities associated with a web service. The time at which the user returns to
the service is predicted, given a set of historical data. We propose to use a Bayesian non-parametric
model, log Gaussian Cox process (LGCP), which allows the latent intensity function generating the
return times to be learnt non-parametrically from the data. It also allows us to encode prior domain
knowledge such as periodicity in users return time using Gaussian process kernels. Further, we cap-
ture the similarities among the users in their return time by using a multi-task learning approach
in the LGCP framework. We compare the performance of LGCP with different kernels on a real-
world last.fm data and show their superior performance over standard radial basis function kernel
and baseline models. We also found LGCP with multitask learning kernel to provide an improved
predictive performance by capturing the user similarity.
Actions (login required)
 |
View Item |
Altmetric
Altmetric