source: arxiv statistics ml: deep multitask learning for mixed-type outcomes with shared sparsity
level: research
most multitask learning methods struggle when tasks have different outcome types, like continuous, binary, or count data, because they need separate loss functions that are hard to combine. this paper introduces a framework that uses unknown monotone transformations to make outcomes comparable across tasks. the approach assumes that only a common subset of predictors is important, which is useful in high-dimensional settings like genomics where many variables are measured but few are truly relevant.
the method works by optimizing a smoothed rank-based criterion with a group-lasso penalty. this encourages the model to select the same predictors for all tasks. the implementation uses a multitask deep neural network with a shared first layer, allowing the network to learn a common representation while handling different outcome types through task-specific transformations. the rank-based loss avoids needing to specify distributions for each outcome, making it flexible for mixed data.
the framework is designed for high-dimensional biological applications where the number of predictors can grow with the sample size. by sharing sparsity patterns, the model can borrow strength across tasks, improving prediction and variable selection when individual tasks have limited data. the deep learning component captures complex nonlinear relationships, while the monotone transformations preserve the ordering of outcomes, ensuring interpretability.
why it matters: this approach enables more accurate predictive models in fields like genomics or personalized medicine, where data often includes diverse outcome types and many irrelevant variables.
source: arxiv statistics ml: deep multitask learning for mixed-type outcomes with shared sparsity