Quality Model for Machine Learning Components

Technical Summary: Quality Model for Machine Learning Components

Problem Statement
Despite the increasing adoption of machine learning (ML) and the availability of MLOps infrastructures, a significant gap remains between ML model development and successful production deployment. Studies indicate that only approximately 40% of ML prototypes reach production. A primary cause of failure is that testing practices are largely limited to evaluating model properties, such as predictive accuracy, while neglecting requirements derived from the broader system context (e.g., throughput, resource consumption, robustness).

Furthermore, a disconnect often exists between model developers and system stakeholders. Model developers frequently lack the system context or the necessary skills to test beyond accuracy, while system requirements are often not shared with them. Existing quality standards, such as ISO 25010 and the newer ISO 25059 for AI systems, present a challenge: they combine system-level attributes with component-level attributes. Attributes like "co-existence" (the ability to perform efficiently while sharing resources) cannot be assessed at the component level by a model developer, yet they impose critical constraints on the component. This conflation hinders effective requirements elicitation and testing for ML components, particularly in scenarios where development is outsourced, conducted by separate data science teams, or involves pre-trained models.

Methodology
The authors developed a quality model specifically for ML components through a rigorous, multi-step process designed to isolate testable component attributes from system-level attributes.

1. Source Identification: The process began with three standards documents (ISO 25010, ISO 25059, and ISO 25019) and two relevant academic/industry studies.
2. Extraction and Deduplication: A team of four researchers (with backgrounds in software engineering and data science) extracted 163 quality attributes (QAs) verbatim. Using card sorting, they grouped these into 80 QA groups to eliminate duplicates and resolve naming inconsistencies.
3. Annotation and Filtering: Each QA group was annotated by type:
   • Type 1: Applies to the ML component (testable at the component level).
   • Type 2: Applies to the ML-enabled system but places requirements on the ML component (system-derived).
   • Type 3: Applies strictly to the system (not testable at the component level).
     Through consensus building, 44 groups were identified as Type 1 (or Type 1 & 2), while 10 were Type 3.
4. Definition and Categorization: The team refined the 44 groups into 35 distinct QAs with definitions focused on ML capabilities. A final card sorting exercise grouped these into 7 categories, resulting in a final model of 30 Quality Attributes.

Key Contributions
The primary contribution is a Quality Model for ML Components that serves as a guide for requirements elicitation and negotiation. It provides a common vocabulary for ML component developers and system stakeholders to define system-derived requirements and focus testing efforts. The model distinguishes between attributes that can be tested at the component level and those that are system-level, addressing the limitations of current standards like ISO 25059.

The model is structured into seven categories:

1. Behavior Analysis: Analyzability, Monitorability, Testability, Understandability.
2. Confidence: Explainability, Functional Correctness, Interpretability.
3. Consistency: Repeatability, Reproducibility.
4. Continued Operation: Deployability, Resilience, Reliability, Resource Utilization, Robustness, Time Behavior.
5. Maintenance and Evolution: Domain Adaptability, Maintainability, Modifiability, Portability, Replaceability, Retrainability, Reusability.
6. Responsible AI: Compliance, Economic Risk Consideration, Fairness, Health and Safety Risk Consideration, Inclusivity, Societal and Ethical Risk Consideration.
7. Security and Privacy: Privacy, Resistance.

Results and Validation
The relevance and value of the model were validated through an invitation-only practitioner survey conducted in August and September 2025, yielding 22 complete responses from government, research, and industrial organizations.

• Current Testing Practices (RQ1): The survey revealed that current testing is overwhelmingly dominated by "Predictive Quality" (specifically accuracy), which constituted 19% of responses. Other attributes like latency and bias were mentioned far less frequently (6%). This confirms the gap the model aims to fill.
• Applicability to LLMs vs. Traditional ML (RQ2): Participants indicated that 70% of the identified QAs are relevant to both traditional ML models and Large Language Models (LLMs). While some attributes were perceived as more relevant to one type (e.g., "calibration" for LLMs, "latent variables" for traditional ML), the model was found to provide broad guidance across model types.
• Importance and Difficulty (RQ3):
  • Importance: Reliability, Testability, Deployability, Functional Correctness, and Robustness were rated as the most important.
  • Difficulty: "Resistance" (security) was rated as the most difficult to test, followed by Resilience and Ethical Risk Considerations. Conversely, Resource Utilization and Time Behavior were rated as easier to test.
  • Consensus: 100% of participants agreed (strongly or otherwise) that considering these quality attributes leads to the early detection of problems that would otherwise remain undetected until production.

Significance and Claims
The paper claims that this is the first empirically developed and validated quality model that focuses specifically on testable quality attributes for ML components. Its significance lies in:

1. Bridging the Communication Gap: It provides a shared language for system stakeholders and model developers to negotiate requirements, particularly when teams are separated or development is outsourced.
2. Enabling System-Directed Testing: It allows model developers to assess how well their components satisfy system-derived requirements (e.g., resource limits, time behavior) before integration.
3. Practical Application: The model has been successfully integrated into MLTE (ML Test and Evaluation), an open-source tool. MLTE uses the model to guide requirements elicitation and includes a test catalog organized by these quality attributes, offering reusable test code examples for each attribute.

The authors maintain a modest stance, acknowledging that the model does not cover data quality (which they propose as a future step) and that scalability was intentionally excluded as a system-level property rather than a component property. The work aims to shift the industry focus from purely predictive metrics to a holistic view of component quality within the system context.