Join us on October 16th to continue our Fall 2024 CTML Seminar Series! Yi Li's talk "Towards Estimation of the Intensity" will take place at 11:00AM at Berkeley Way West, 5th Floor, Room 5101. Please note the conference room change for this seminar.
Intensity estimation serves as a key building block to understand the jumping process. In statistics, we can even model the density through intensity. However, intensity estimation usually involves partial log likelihood that takes the form of integration of the intensity and log of the intensity itself for a single observation of the jumping process. Thus, how to handle such complex partial log likelihood is of interest. Unlike the square error loss where over-fitting can be easily defined, here, the over-fitting of the jumping processes through the intensity estimation using the partial log likelihood is less clearly understood. In this paper, we propose to handle the complex partial log likelihood through its various approximations. In addition, we propose different modeling techniques including highly adaptive lasso and multi-layer neural networks for the intensity model. We treat different configurations of the approximated losses and models paired together as different learners for the intensity, we then use a very finely approximated true loss to do the selection of the learner to avoid the over-fitting.