First speaker: Christina Zhou
Title: Deep Survival for Frailty Data
Abstract: In traditional survival analysis, common methods include fully parametric models such as the Weibull distribution and the semi-parametric Cox Proportional Hazards (CPH) model. In recent years, with the rise of deep learning and the increase of complex, high dimensional data (such as from EHRs), deep survival methods have been developed. By incorporating deep learning aspects, methods can perform faster computation on large data; and neural nets are better able to capture complicated relationships in covariate data (instead of assuming linear relationships as necessary in CPH). Incorporating deep learning into survival frailty data is less explored. When survival data assumes that there is an unobserved heterogeneity, such as within individuals in in clusters, this effect cannot be explained by observed covariates. Thus, frailty methods such as the shared frailty model were developed to account for these unmeasured characteristics. One problem with traditional methods is that when the dimension of latent variable is large, computation can become complicated. For some methods, the posterior distribution can be difficult to compute, if not intractable. For others, having a high number of parameters to estimate might be computationally expensive and difficult to keep track of. We are interested in designing a modern method that offers faster computation and results in better prediction than existing traditional methods for clustered data. For each fixed cluster, we assume a latent variable Z to explain the heterogeneity among clusters. Since this posterior distribution is often intractable, we estimate it through Monte Carlo samples and use a variation of the evidence lower bound (ELBO) as our loss function. We create a type of EM algorithm that uses the ADAM optimizer to obtain parameter estimates. In this talk, I will provide background on traditional survival analysis; a brief overview of and motivation for neural nets and deep learning; and discuss our method design.
Second speaker: Daniel De Marchi
Title: Shapley Marginal Surplus for Strong Models
Abstract: Shapley value approaches have seen widespread use in machine learning as a way to explain model predictions. The goal in explaining these predictions is to inform real-world decisions and to see which variables influence the outcome of interest. In this paper, we demonstrate that while model-based Shapley values might be accurate explainers of model predictions, machine learning models themselves are often poor explainers of the data and understanding their predictions can be misleading to understanding the data itself. Particularly in the presence of interrelated variables the output of a highly predictive model may fail to account for these relationships and wildly distort the true variable importance. In this paper we introduce a novel metric, Shapely Marginal Surplus for Strong Models, that samples the space of possible models to come up with a truly explanatory measure of feature importance that takes into account potential feature interrelation. We compare this method to other popular feature importance metrics and demonstrate significant outperformance relative to other methods.