Data Acquisition: A New Frontier in Data-centric AI

Datasets, the cornerstone of modern machine learning (ML) systems, have been increasingly sold and purchased for different ML pipelines [2]. Several data marketplaces have emerged to serve different stages of building ML-enhanced data applications. For example, NASDAQ Data Link [3] offers financial datasets cleaned and structured for model training, Amazon AWS data exchange [4] focuses on generic tabular datasets, and Databricks Marketplace [5] integrates raw datasets and ML pipelines to deliver insights. The data-as-a-service market size was more than 30 billions and is expected to double in the next five years [6]. While the data marketplaces are increasingly expanding, unfortunately, data acquisition for ML remains challenging, partially due to its ad-hoc nature: Based on discussions with real-world users, data acquirers often need to negotiate varying contracts with different data providers first, then purchase multiple datasets with different formats, and finally filtering out unnecessary data from the purchased datasets.