Deep learning models (such as transformer models) have obtained great advances in various domains, such as object recognition, natural language processing and generative applications. However, training such a model requires some data science experience. Various hyperparameters need to be set correctly, including those that relate to the optimal architecture. Additionally, these methods are known to require plenty of data. In this presentation, we will deal with various techniques (i.e., neural architecture search, metalearning and the extrapolation of learning lurves) that support the data scientist in further automating such training processes.

The area of Meta-learning (MtL) leverages knowledge from problems for which successful Machine Learning (ML) solutions are known to support automated algorithm selection for new problems. But far more meta-knowledge can be extracted by relating data properties to algorithmic performance. This topic remains under-explored compared to using MtL for automated algorithm selection. For instance, one may reveal the competencies and limitations of different ML algorithms and highlight data quality issues worth investigating. By deepening such understanding, we expect to contribute to improving the comprehensibility and reliability of the usage of ML models. We also expect to generate contributions in areas that can directly benefit from data and algorithm understanding, such as data pre-processing. The idea is to guide the solution of the previous tasks using meta-knowledge extracted about the dichotomous relationship between data properties and algorithmic performance.

Automated data science has established itself as a key technology in both research and industry to customize data science and machine learning pipelines to learning tasks. While it supports its users by automating repetitive, tedious, and time-consuming tasks of selecting algorithms for preprocessing data, learning models from data, and ensembling the resulting models to achieve peak performance, automated data science remains a computationally expensive endeavor. On the one hand, the demand for computational resources required to use methods from automated data science incurs financial costs, which are also linked to energy costs. On the other hand, consuming a substantial amount of energy also significantly impacts the environment in terms of CO2 footprint and increased hardware attrition.

This session will focus on the interface between automated data science and sustainability. To this end, we will first discuss different types of efficiency when dealing with automated data science systems and, with that, the concept of energy efficiency. Furthermore, we will discuss that we can consider sustainability on different levels in an automated data science system, ranging from optimizing for sustainable data science or machine learning pipelines to designing sustainable automated data science systems to designing automated data science systems to learn across tasks and leverage their knowledge to become more efficient optimizers after all. Furthermore, we will discuss an example where "green thinking" even helps to design better AutoML systems, thus leveraging "green thinking" as an incentive for creative thinking in research. We conclude the session with a call to action to establish sustainability in both research on and applying automated data science systems. As a practical takeaway, we have a brief live-coding session on a tool named PyExperimenter, which facilitates distributing computational experiments, e.g., on high-performance computers. Moreover, PyExperimenter logs the energy consumption of conducted experiments and, in addition to that, the estimated carbon footprint, integrating with the carbon emission measuring framework CodeCarbon.io. While PyExperimenter can be used to evaluate automated data science systems, its interface is quite generic, so it can also be used for any other type of computational experiments, eliciting the carbon footprint and, therefore, the sustainability of computational experiments in general.