van der Schaar Lab

Conference Season 2023: 7 papers at AISTATS & 4 papers at ICLR

Advancing best-in-class reinforcement learning, symbolic regression, and representation learning for tabular data, and transforming synthetic data, clinical trials, treatment effect estimation, and missing data imputation 

We are very proud to consistently be represented at the world’s largest and most prestigious AI and machine learning conferences with our cutting-edge research, impactful papers, and participating in fruitful workshops as part of the conferences.

The 2023 conference season starts with Artificial Intelligence and Statistics (AISTATS), running from 25 – 27 April, one of the most prominent annual gatherings of researchers at the intersection of artificial intelligence, machine learning, statistics, and related areas. This year, the van der Schaar Lab will present 7 papers, the largest number of accepted papers we have had at AISTATS thus far. This shows the ongoing strong efforts of our team to produce increasingly impactful and creative work.

Following close behind will be the Eleventh International Conference on Learning Representations (ICLR), running from 1 – 5 May. ICLR is globally renowned for presenting and publishing cutting-edge research on all aspects of deep learning used in the fields of artificial intelligence, statistics, and data science. The van der Schaar Lab will be part of the event with 4 papers.

Collectively, these papers touch on some of the most important areas within the lab’s extensive research agenda, including reinforcement learning, symbolic regression and representation learning for tabular data, synthetic data, clinical trials, treatment effect estimation, and missing data imputation.

When a decision-maker should break a commitment

Clinical trials are costly and time intensive necessities in the development of new drugs. Committing to the process bears the inherent risk of potential failure and the loss of investment. Our team asks the question: When should a decision-maker break a commitment that is likely to fail – either to make an alternative commitment or to make no further commitments at all? In this ground-breaking paper, we formulate this question as a novel type of optimal stopping/switching problem called the optimal commitment problem (OCP), and propose a practical algorithm for solving it.

Find the paper here.

Improving symbolic regression

In this work, we address the complex problem of discovering concise closed-form mathematical equations from data by leveraging pre-trained deep generative models to capture the intrinsic regularities of equations. This novel approach fundamentally changes symbolic regression as it unifies several prominent approaches and offers a new perspective to justify and improve on the previous ad hoc designs. We lay the groundwork for superior recovery rates of true equations while also being computationally more efficient at inference time than previous solutions.

Find the paper here.

Realistic synthetic tabular data by exploiting relational structure

From computer vision to natural language processing – deep generative models have shown notable success in learning highly complex and non-linear representations to generate realistic synthetic data. So far, the complex challenges characterising tabular data, such as heterogeneous relationships, a limited number of samples, and difficulties in incorporating prior knowledge made the generation of suitable synthetic data problematic. The van der Schaar lab is the first to solve this issue by introducing GOGGLE, an end-to-end message passing scheme that jointly learns the relational structure and functional relationships as the basis of generating synthetic samples.

Find the paper here.

Shining the spotlight on competing risks in inferring heterogeneous treatment effects

Although competing risks present great practical relevance, this problem has seen very little attention in studying treatment effect estimation. We are the first to theoretically analyse and empirically illustrate when and how competing risks play a role in using generic machine learning prediction models for the estimation of heterogeneous treatment effects. In our investigation, we find that competing risks can act as an additional source of covariate shift, refocussing the lens through which treatment effect estimations should be observed.

Find the paper here.

Breaking new ground: Neural Laplace Control

Many real-world offline reinforcement learning problems involve continuous-time environments with delays. Observations are made at irregular time intervals and, in return, actions take effect with unknown delays as well. Existing offline reinforcement learning algorithms have shown to be effective with irregularly observed states in town or known delays. However, we are the first to attempt to solve the challenge of environments involving both irregular observations in time AND unknown delays. We succeed with our proposed solution: Neural Laplace Control – a continuous-time model-based offline reinforcement learning method that combines a Neural Laplace dynamics model with a model predictive control planner. In doing so, our model achieves near expert policy performance.

Find the paper here.

You can find the full list of papers for AISTATS and ICLR below.

AISTATS: 25 – 27 April 2023

ICLR: 1 – 5 May 2023

Andreas Bedorf