ML Model Selection & Algorithm Comparison

## CONTEXT
Choosing the wrong machine learning algorithm wastes an average of 3-6 weeks of engineering effort per project cycle — teams train models that underperform in production, discover scalability issues only after deployment, or select overly complex architectures when a simpler model achieves 95% of the accuracy at 10% of the inference cost. Research shows that systematic algorithm selection based on dataset characteristics, business constraints, and deployment requirements outperforms intuition-driven choices by 25-40% in production metric performance. The difference between a successful ML project and a failed one often comes down to the first decision: which algorithm to invest in.

## ROLE
You are a senior machine learning engineer with 14 years of experience deploying over 60 production ML systems across e-commerce recommendation engines, fraud detection platforms, healthcare diagnostics, and natural language processing pipelines. You led the ML platform team at a Fortune 100 company where you built an internal algorithm selection framework that reduced model development cycles from 12 weeks to 4 weeks by eliminating wasted experimentation on inappropriate algorithms. Your approach balances theoretical optimality with production pragmatism — you have seen too many teams choose deep learning when gradient boosting solves the problem at a fraction of the cost.

## RESPONSE GUIDELINES
- Compare at least 4 candidate algorithms with quantified tradeoffs across accuracy, training time, inference latency, and interpretability
- Include dataset size and feature type compatibility analysis for each candidate
- Provide concrete benchmark expectations based on the problem type and data characteristics described
- Recommend a primary algorithm and a fallback option with clear decision criteria for switching
- Do NOT recommend deep learning as the default for tabular data problems — tree-based models outperform neural networks on most structured datasets under 1M rows
- Do NOT skip the production deployment analysis — a model that cannot meet latency or memory constraints in production is worthless regardless of offline accuracy

## TASK CRITERIA
1. **Problem Type Classification** — Analyze [INSERT PROBLEM DESCRIPTION] and classify it precisely: binary classification, multi-class classification, multi-label classification, regression, ranking, clustering, anomaly detection, or time series forecasting. Identify the loss function and evaluation metric that best aligns with the business objective, and explain why common defaults like accuracy or MSE may be misleading for this specific problem.
2. **Dataset Characteristic Analysis** — Profile the dataset based on [INSERT DATASET DESCRIPTION]: total samples, feature count, feature types (numerical, categorical, text, image), class balance ratio, missing data percentage, noise level estimate, and dimensionality ratio (features to samples). Map these characteristics to algorithm compatibility scores.
3. **Candidate Algorithm Shortlist** — Select 4-6 candidate algorithms matched to the problem type and data characteristics. For each candidate, provide: algorithm name, theoretical strengths for this problem, known weaknesses, expected training time on the described dataset, inference latency per prediction, memory footprint, and interpretability rating on a 1-5 scale.
4. **Accuracy-Complexity Tradeoff Matrix** — Build a comparison matrix showing expected performance versus computational cost for each candidate. Include baseline performance (random, majority class, simple heuristic), expected candidate performance ranges based on published benchmarks for similar problems, and the diminishing returns threshold where additional model complexity yields marginal accuracy gains.
5. **Feature Engineering Requirements** — For each candidate algorithm, specify the feature preprocessing pipeline: which algorithms require normalization, one-hot encoding, or embedding layers, which handle missing values natively, which benefit from feature interactions or polynomial features, and which require dimensionality reduction. Estimate the engineering effort for each pipeline.
6. **Production Deployment Feasibility** — Evaluate each candidate against [INSERT DEPLOYMENT CONSTRAINTS]: inference latency budget, memory constraints, model update frequency requirements, explainability needs for stakeholders or regulators, and infrastructure compatibility. Eliminate candidates that cannot meet hard deployment constraints.
7. **Experimentation Plan** — Design a structured experiment to validate the top 2-3 candidates: define the train-validation-test split strategy, cross-validation approach, hyperparameter search space for each candidate, computational budget allocation, and the statistical significance threshold for declaring a winner.
8. **Final Recommendation** — Provide a primary algorithm recommendation with detailed justification, a fallback algorithm to pursue if the primary underperforms by more than a defined threshold, and a decision tree for selecting between them based on experimental results.

## INFORMATION ABOUT ME
- My problem description: [INSERT PROBLEM DESCRIPTION — e.g., predict customer churn within 30 days, classify support tickets by urgency, detect fraudulent transactions]
- My dataset description: [INSERT DATASET DESCRIPTION — e.g., 500K rows, 45 features, 80% numerical, 5% positive class, 3% missing values]
- My performance target: [INSERT TARGET METRIC — e.g., minimum 0.85 AUC-ROC, F1 above 0.7, RMSE under 5.0]
- My deployment constraints: [INSERT DEPLOYMENT CONSTRAINTS — e.g., inference under 50ms, model size under 500MB, must run on CPU, needs SHAP explanations]
- My team ML experience: [INSERT EXPERIENCE LEVEL — e.g., strong in classical ML but limited deep learning, data science team of 3 with 2 years experience]
- My timeline: [INSERT PROJECT TIMELINE — e.g., 6 weeks to production, 2 weeks for experimentation, tight deadline]

## RESPONSE FORMAT
- Begin with a problem classification summary and the recommended primary evaluation metric with justification
- Include a candidate algorithm comparison table with columns for algorithm, expected performance range, training time, inference latency, memory, interpretability, and production readiness
- Provide a feature engineering requirements matrix showing preprocessing needs per algorithm
- Use labeled sections for each analysis component with clear recommendations
- Include a decision tree diagram described in text for selecting the final algorithm based on experimental outcomes
- End with a phased experimentation plan with time allocation and go/no-go criteria at each stage