claude-flow

{ "name": "MLE-STAR Machine Learning Engineering Workflow", "version": "1.0.0", "description": "Complete Machine Learning Engineering workflow using MLE-STAR methodology (Search and Targeted Refinement)", "variables": { "dataset_path": "data/ml_dataset.csv", "target_metric": "accuracy", "model_output_dir": "models/", "experiment_name": "mle-star-experiment", "search_iterations": 3, "refinement_iterations": 5 }, "agents": [ { "id": "search_agent", "type": "researcher", "name": "Web Search & Foundation Agent", "config": { "capabilities": ["web_search", "model_research", "approach_validation"], "search_depth": "comprehensive", "domains": ["machine_learning", "data_science", "kaggle_solutions"] } }, { "id": "foundation_agent", "type": "coder", "name": "Foundation Model Builder", "config": { "capabilities": ["data_preprocessing", "initial_modeling", "baseline_creation"], "programming_languages": ["python", "jupyter"], "frameworks": ["scikit-learn", "pandas", "numpy", "matplotlib"] } }, { "id": "refinement_agent", "type": "optimizer", "name": "Targeted Refinement Specialist", "config": { "capabilities": ["feature_engineering", "model_optimization", "hyperparameter_tuning"], "optimization_methods": ["grid_search", "bayesian", "evolutionary"], "ablation_analysis": true } }, { "id": "ensemble_agent", "type": "architect", "name": "Ensemble & Meta-Learning Expert", "config": { "capabilities": ["model_stacking", "ensemble_methods", "meta_learning"], "ensemble_strategies": ["voting", "stacking", "blending", "dynamic_weighting"] } }, { "id": "validation_agent", "type": "tester", "name": "Validation & Debugging Agent", "config": { "capabilities": ["cross_validation", "error_detection", "data_leakage_prevention"], "validation_strategies": ["stratified_kfold", "time_series_split", "leave_one_out"], "debugging_tools": ["profiling", "error_correction", "data_integrity_checks"] } }, { "id": "orchestrator", "type": "coordinator", "name": "MLE-STAR Orchestrator", "config": { "capabilities": ["workflow_coordination", "performance_tracking", "resource_management"], "coordination_mode": "adaptive", "monitoring_enabled": true } } ], "tasks": [ { "id": "web_search_phase", "name": "Web Search for ML Approaches", "type": "research", "description": "Search web for state-of-the-art ML approaches, model architectures, and proven solutions for the given problem domain", "assignTo": "search_agent", "input": { "problem_type": "${dataset_analysis.problem_type}", "data_characteristics": "${dataset_analysis.characteristics}", "search_queries": [ "latest machine learning models ${problem_type}", "kaggle winning solutions ${problem_type}", "state-of-the-art ${problem_type} benchmarks", "feature engineering techniques ${problem_type}" ] }, "timeout": 900, "retries": 2, "output": { "recommended_approaches": "array", "model_cards": "array", "benchmark_results": "object", "implementation_examples": "array" } }, { "id": "dataset_analysis", "name": "Dataset Analysis & Profiling", "type": "analysis", "description": "Comprehensive analysis of dataset characteristics, quality, and problem type identification", "assignTo": "foundation_agent", "input": { "dataset_path": "${dataset_path}", "target": "${target_column}" }, "timeout": 300, "output": { "problem_type": "string", "characteristics": "object", "data_quality_report": "object", "recommended_preprocessing": "array" } }, { "id": "foundation_building", "name": "Foundation Model Creation", "type": "implementation", "description": "Build initial ML pipeline based on web search findings and dataset analysis", "assignTo": "foundation_agent", "depends": ["web_search_phase", "dataset_analysis"], "input": { "search_results": "${web_search_phase.output}", "dataset_info": "${dataset_analysis.output}", "approaches_to_implement": "${web_search_phase.output.recommended_approaches}" }, "timeout": 1800, "output": { "baseline_models": "array", "preprocessing_pipeline": "object", "initial_results": "object", "code_components": "object" } }, { "id": "ablation_analysis", "name": "Component Impact Analysis", "type": "analysis", "description": "Systematically analyze which pipeline components have the highest performance impact", "assignTo": "refinement_agent", "depends": ["foundation_building"], "input": { "foundation_pipeline": "${foundation_building.output}", "baseline_performance": "${foundation_building.output.initial_results}", "components_to_test": [ "data_preprocessing", "feature_engineering", "model_selection", "hyperparameters" ] }, "timeout": 1200, "output": { "component_rankings": "array", "impact_scores": "object", "highest_impact_component": "string", "improvement_opportunities": "array" } }, { "id": "targeted_refinement", "name": "Iterative Component Refinement", "type": "optimization", "description": "Deep, focused improvements on the highest-impact pipeline components identified through ablation", "assignTo": "refinement_agent", "depends": ["ablation_analysis"], "input": { "target_component": "${ablation_analysis.output.highest_impact_component}", "baseline_performance": "${ablation_analysis.output.impact_scores}", "improvement_strategies": "${ablation_analysis.output.improvement_opportunities}", "max_iterations": "${refinement_iterations}" }, "timeout": 2400, "output": { "optimized_components": "object", "performance_improvements": "object", "refinement_history": "array", "best_configuration": "object" } }, { "id": "ensemble_creation", "name": "Adaptive Ensemble Building", "type": "architecture", "description": "Intelligently combine models using advanced ensemble techniques, not just averaging", "assignTo": "ensemble_agent", "depends": ["targeted_refinement"], "input": { "base_models": "${targeted_refinement.output.optimized_components}", "individual_performances": "${targeted_refinement.output.performance_improvements}", "ensemble_strategies": [ "stacking_with_meta_learner", "dynamic_weighting", "bayesian_model_averaging", "mixture_of_experts" ] }, "timeout": 1800, "output": { "ensemble_model": "object", "ensemble_performance": "object", "combination_strategy": "string", "meta_learner_details": "object" } }, { "id": "validation_and_debugging", "name": "Comprehensive Validation & Error Correction", "type": "testing", "description": "Rigorous validation with data leakage prevention, error correction, and performance verification", "assignTo": "validation_agent", "depends": ["ensemble_creation"], "input": { "final_model": "${ensemble_creation.output.ensemble_model}", "full_pipeline": "${targeted_refinement.output.best_configuration}", "validation_config": { "cv_folds": 5, "test_size": 0.2, "stratification": true, "random_seed": 42 } }, "timeout": 900, "output": { "validation_results": "object", "cross_validation_scores": "array", "data_leakage_check": "boolean", "error_analysis": "object", "final_performance_metrics": "object" } }, { "id": "model_deployment_prep", "name": "Production Readiness & Deployment", "type": "deployment", "description": "Prepare model for production deployment with monitoring, versioning, and documentation", "assignTo": "orchestrator", "depends": ["validation_and_debugging"], "input": { "validated_model": "${validation_and_debugging.output}", "performance_benchmarks": "${validation_and_debugging.output.final_performance_metrics}", "deployment_target": "production" }, "timeout": 600, "condition": "${validation_and_debugging.output.final_performance_metrics.accuracy} > 0.85", "output": { "deployment_package": "object", "model_documentation": "object", "monitoring_config": "object", "deployment_instructions": "string" } } ], "dependencies": { "foundation_building": ["web_search_phase", "dataset_analysis"], "ablation_analysis": ["foundation_building"], "targeted_refinement": ["ablation_analysis"], "ensemble_creation": ["targeted_refinement"], "validation_and_debugging": ["ensemble_creation"], "model_deployment_prep": ["validation_and_debugging"] }, "settings": { "maxConcurrency": 3, "timeout": 7200, "retryPolicy": "exponential", "failurePolicy": "continue", "quality_threshold": 0.8, "enable_monitoring": true, "save_intermediate_results": true, "auto_documentation": true }, "metadata": { "methodology": "MLE-STAR", "version": "2.0.0", "created_by": "Claude Flow Automation", "tags": ["machine_learning", "automation", "mle_star", "ensemble", "optimization"], "description_detailed": "This workflow implements the MLE-STAR (Machine Learning Engineering via Search and Targeted Refinement) methodology. It starts with web search to find state-of-the-art approaches, builds a foundation model, performs ablation analysis to identify high-impact components, then deeply refines those components before creating intelligent ensembles and comprehensive validation.", "success_criteria": [ "Performance improvement over baseline > 5%", "No data leakage detected", "Cross-validation score variance < 0.05", "All components properly documented" ], "expected_runtime": "2-4 hours", "resource_requirements": { "memory": "8GB+", "storage": "10GB+", "compute": "4+ CPU cores recommended" } } }