XAIport: Bringing Explainability to MLOps Pipelines

The Problem

Machine Learning Operations (MLOps) focuses on automating ML lifecycles - model development, validation, deployment, and monitoring. However, explainability is often treated as an afterthought, added only after models are deployed.

This creates problems:

Unexplainable models in production
Late discovery of model biases
Compliance issues in regulated industries
Difficulty debugging model failures

Our Solution: XAIport

XAIport enables early adoption of XAI - integrating explainability checks throughout the ML development cycle, not just at the end.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      XAIport Framework                       │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│   ┌─────────────┐   ┌─────────────┐   ┌─────────────┐     │
│   │ Data Service│   │Model Service│   │ XAI Service │     │
│   └──────┬──────┘   └──────┬──────┘   └──────┬──────┘     │
│          │                 │                  │            │
│          └────────────────┬┴─────────────────┘            │
│                           │                                │
│                    ┌──────▼──────┐                        │
│                    │  Open APIs  │                        │
│                    └──────┬──────┘                        │
│                           │                                │
│   ┌───────────────────────┼───────────────────────┐       │
│   │        Cloud AI Services Integration          │       │
│   │   [Azure]     [GCP]      [AWS]     [HuggingFace]     │
│   └───────────────────────────────────────────────┘       │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Key Features

Microservice Architecture: Each XAI method is encapsulated as an independent service
Open APIs: RESTful endpoints for easy integration
Cloud Agnostic: Works with Azure, GCP, AWS, and local models
Configurable Pipelines: Define XAI workflows through configuration
Provenance Tracking: Record how explanations are generated for reproducibility

Quick Start

from xaiport import XAIService, Pipeline

# Initialize XAI service
xai = XAIService()

# Create explanation pipeline
pipeline = Pipeline([
    xai.shap_explainer(background_samples=100),
    xai.lime_explainer(num_features=10),
    xai.consistency_evaluator()
])

# Run on your model
results = pipeline.explain(
    model=your_model,
    data=your_data,
    target_class=1
)

# Get consistency metrics
print(f"Explanation consistency: {results.consistency_score}")

Experiment Results

We tested XAIport with three major cloud AI services:

Cloud Service	Baseline Accuracy	After XAI Optimization
Azure Cognitive	0.82	0.89
Google Vertex AI	0.79	0.86
AWS Rekognition	0.81	0.87

Key finding: Early XAI adoption improves both model performance AND explainability metrics.

Use Cases

1. Model Development

Identify feature importance issues early in development.

2. Quality Assurance

Automated XAI checks in CI/CD pipelines.

3. Compliance

Generate explanation reports for regulatory requirements.

4. Debugging

Understand why models make specific predictions.

GitHub Repository

git clone https://github.com/ZeruiW/XAIport
cd XAIport
docker-compose up

Access the web interface at http://localhost:8080

Citation

@inproceedings{wang2024xaiport,
  title={XAIport: A Service Framework for the Early Adoption of XAI in AI Model Development},
  author={Wang, Zerui and Liu, Yan and Thiruselvi, Aadhitya and Hamou-Lhadj, Wahab},
  booktitle={Proceedings of the 46th IEEE/ACM International Conference on Software Engineering (ICSE)},
  pages={67--71},
  year={2024},
  doi={10.1145/3639476.3639759}
}

What's Next

XAIpipeline (IEEE SSE 2025): Automated orchestration for multi-cloud deployments
STAA Integration: Real-time video explanations in XAIport
Enterprise Features: Role-based access, audit logs, compliance reports