Data-Centric Artificial Intelligence

2 Model-centric and Data-centric AI

• Model-only optimization has plateaued on many datasets, while practitioners often face unique problems without public datasets or suitable pretrained models.Data quantity and quality, including relevance and label quality, influence AI-system performance.Source: §2 · p. 2
• Data-centric AI emphasizes systematic data design and engineering, generally holding the ML model fixed while improving the data.This contrasts with model-centric AI, which focuses on choosing the model type, architecture, and hyperparameters.Source: §2 · p. 2
• Domain-specific data work is integral to data-centric AI, supported by methods and semi-automated tools to accelerate development of successful AI systems.The paper also describes data-centric AI as involving data work and domain knowledge.Source: §2 · p. 2
• Changes in ML performance metrics can indicate whether data adjustments are effective, offering a way to approximate data quality.Data-centric performance gains are pursued through more appropriate data.Source: §2 · p. 2
• Data-centric and model-centric AI are complementary paradigms, and AI-system development should ultimately incorporate both.The article notes that AI includes ML as well as other methods, so data-centric AI is broader than data-centric machine learning alone.Source: §2 · p. 3

3 Dimensions of Data-centric AI

• The framework has two dimensions: refining existing data (“better data”) and extending it by acquiring additional data (“more data”).Figure 2 presents the framework for systematic data design and engineering.Source: §3 · p. 3, Figure 2 · p. 4
• Refinement improves existing data quality by addressing instance features and labels, representativeness, low-quality instances, and the relevance and fairness of features.Relevant underrepresented instances may be augmented, while outliers should be distinguished from edge cases; refinement work has often been manual.Source: §3 · p. 3
• Extension acquires data for dataset blind spots through new instances, features, or target labels, supporting initial performance requirements and responses to distribution shifts.The framework characterizes extension as increasing data quantity, in contrast to refinement of existing data.Source: §3 · p. 4
• Widespread adoption requires methods and tools for data versioning, labeling, exploration, visualization, and identifying special instances; systematic data design remains an identified methods gap.Semi-automated data exploration may contribute to data understanding across instance-level refinement, dataset-level refinement, and extension; Table 1 summarizes examples of tools and meta tools.Source: §3 · pp. 4–5, Table 1 · p. 5

4 Delimitations of Data-centric AI from Related Concepts

• Unlike Big Data’s general emphasis on collecting, storing, and processing large volumes, data-centric AI systematically acquires better or more data and may remove deficient or irrelevant data.The distinction is especially relevant in specialized domains where large data volumes may not be available.Source: §4 · pp. 5–6
• MLOps emphasizes production engineering practices, while data-centric AI focuses on systematic dataset refinement and extension, including dataset monitoring and versioning.The paper notes little MLOps focus so far on monitoring and versioning datasets adapted by data-centric methods.Source: §4 · p. 6
• Data-driven versus model-driven methods concern how information is generated, whereas data-centric versus model-centric AI distinguish paradigms for developing an ML model.The paper states that the choice between these two pairs of paradigms does not necessarily determine the other.Source: §4 · p. 6

5 Implications for BISE Research

• BISE research can advance data-centric AI by improving data work and its incorporation into AI-based systems.The proposed research areas span individual, organizational, and cross-organizational levels.Source: §5 · p. 6, Figure 3 · p. 7

5.1 Individual Level

• High-quality data understanding calls for relevant subset selection, metadata, visualization, profiling, and domain-specific knowledge.Interactive exploration and profiling can help users identify patterns and data-quality issues such as missing values, outliers, and inconsistencies.Source: §5.1 · p. 7
• Human-in-the-loop data work should balance the value of additional data against expert review time and cost.Semi-automated tools can preselect potentially mislabeled images, edge cases, or informative instances for physicians to review.Source: §5.1 · pp. 7–8

5.2 Organizational Level

• Organizations need continuous data-quality monitoring, verification, and validation, including for real-time or sensor-generated data.Assessing completeness, feature accuracy, and label accuracy alongside model performance can help evaluate data enhancement.Source: §5.2 · p. 8
• Data-centric development makes data understanding and preparation iterative, calls for data versioning, and requires ongoing data work as data changes.Model choice should reflect data understanding and domain knowledge; initial tests may prompt revisiting whether data quantity is sufficient.Source: §5.2 · pp. 8–9

5.3 Cross-Organizational Level

• Cross-organizational data sharing needs controlled platforms, quality standards, and governance for versioning, licensing, and citation.The proposed research also includes incentives and trust or reputation systems for assessing data-source reliability.Source: §5.3 · pp. 9–10
• Federated learning still requires distributed data understanding and preparation, shared processing standards, and attention to each organization’s context.When organizations share models instead of data, assessing label quality and instance relevance is challenging without access to the full dataset.Source: §5.3 · p. 10

• The article introduces data-centric AI, contrasts it with related concepts, and highlights literature gaps for BISE guidance.Source: §6 · p. 10
• Data-centric AI has the potential to improve AI-based system performance in research and practice.Source: §6 · p. 10
• The article identifies data-centric AI as a promising field for BISE research.Source: §6 · p. 10