AStA Wirtschafts- und Sozialstatistisches Archiv最新文献

In this paper we discuss the use and potential advantages and disadvantages of machine learning driven models in rental guides. Rental guides are a formal legal instrument in Germany for surveying rents of flats in cities and municipalities, which are today based on regression models or simple contingency tables. We discuss if and how modern and timely methods of machine learning outperform existing and established routines. We make use of data from the Munich rental guide and mainly focus on the predictive power of these models. We discuss the “black-box” character making some of these models difficult to interpret and hence challenging for applications in the rental guide context. Still, it is of interest to see how “black-box” models perform with respect to prediction error. Moreover, we study adversarial effects, i.e. we investigate robustness in the sense how corrupted data influence the performance of the prediction models. With the data at hand we show that models with promising predictive performance suffer from being more vulnerable to corruptions than classic linear models including Ridge or Lasso regularization.

Machine learning (ML) is increasingly being used in official statistics with a range of different applications. The main focus of ML models is to accurately predict attributes of new, unlabeled cases whereas the focus of classical statistical models is to describe the relations between independent and dependent variables. There is already a lot of experience in the sound use of classical statistical models in official statistics, but for ML models this is still under development. Recent discussions concerning the quality aspects of using ML in official statistics have concentrated on its implications for existing quality frameworks. We are in favor of the use of ML in official statistics, but the main question remains as to what factors need to be considered when using ML models in official statistics. As a means of raising awareness regarding these factors, we pose ten propositions regarding the (sensible) use of ML in official statistics.

Despite the fact that National Statistical Offices (NSOs) continue to embrace and adopt Machine Learning (ML) methods and tools in a variety of areas of their operations, including data collection, integration, and processing, it is still not clear how these complex and prediction-oriented approaches can be incorporated into the quality standards and frameworks within NSOs or if the frameworks themselves need to be modified. This article focuses on and builds upon two of the quality dimensions proposed in the Quality Framework for Statistical Algorithms (QF4SA): model explainability and accuracy (including uncertainty). The implications of the current methods for explainable ML and uncertainty quantification will be examined in further detail, as well as their possible uses in statistical production, such as continuous model monitoring in intermediate ML classifications and auto-coding phases. This strategy will ensure that human subject-matter experts, who are an essential component of every statistical program, are effectively integrated into the life cycle of ML projects. It will also guarantee to maintain the quality of ML models in production, adhere to the current quality frameworks within NSOs, and ultimately boost confidence and trust in these emerging technologies.

The quality of Machine Learning (ML) applications is commonly assessed by quantifying how well an algorithm fits its respective training data. Yet, a perfect model that learns from and reproduces erroneous data will always be flawed in its real-world application. Hence, a comprehensive assessment of ML quality must include an additional data perspective, especially for models trained on human-annotated data. For the collection of human-annotated training data, best practices often do not exist and leave researchers to make arbitrary decisions when collecting annotations. Decisions about the selection of annotators or label options may affect training data quality and model performance.

In this paper, I will outline and summarize previous research and approaches to the collection of annotated training data. I look at data annotation and its quality confounders from two perspectives: the set of annotators and the strategy of data collection. The paper will highlight the various implementations of text and image annotation collection and stress the importance of careful task construction. I conclude by illustrating the consequences for future research and applications of data annotation. The paper is intended give readers a starting point on annotated data quality research and stress the necessity of thoughtful consideration of the annotation collection process to researchers and practitioners.

In this article, we present a framework to assess the challenges in the climate-related data landscape. From our perspective, we describe challenges and opportunities for innovation to close existing gaps in selected data quality dimensions for central banks and statistical offices. We further examine the role of networks in promoting innovation and international collaboration, highlighting practical experiences with a case study of the Sustainable Finance Data Hub at the Deutsche Bundesbank. Finally, we discuss how these lessons can provide avenues to enhance data quality in central banks and official statistics and outline directions for future research.

Zusammenfassung