ML techniques (blueprint)
Purpose: Catalog of machine learning techniques — algorithm families and deep-dive guides on features and evaluation. Each category describes the approach, common algorithms, trade-offs, and use cases.
Guide · Updated · Source
Audience: Teams adopting Data science & machine learning; technique choices for a specific project are documented in experiment logs and model cards.
Technique selection in the ML workflow
Algorithm choice depends on problem type, data scale, interpretability, and latency. Feature engineering usually dominates early iteration quality; evaluation design (splits, metrics, fairness) decides whether you can trust comparisons. Use this folder to map where you are in the workflow to the right reference: encoding and stores before or alongside modeling; metrics and validation while comparing candidates; algorithm rows below when narrowing model families. Deep dives Feature engineering & feature stores and Model evaluation, selection & validation are intentionally tool-agnostic — bind specific libraries and experiment IDs in project docs, not in these blueprints.
Core knowledge: Data science & machine learning body of knowledge — ML lifecycle (CRISP-DM aligned), statistics, model evaluation, MLOps, responsible AI.
ML technique selection flow
How teams map workflow position to technique references before binding libraries in experiment logs.
- Technique selection in the ML workflowFrames how algorithm, feature, and evaluation choices align with workflow position.
- Lane A ──► handoff ──► shared outcomePrimary lane delivers technique choices through governed handoffs to shared outcomes.
- Lane B ──► inspect / adapt ──► feedbackParallel lane inspects comparisons and adapts choices from reviewable evidence.
- Note: Feature engineeringSignals that feature work often dominates early iteration quality.
Technique selection in the ML workflow
Lane A ──► handoff ──► shared outcome
Lane B ──► inspect / adapt ──► feedback
Note: Feature engineering| Category | Core idea | Common algorithms | Deep dive |
|---|---|---|---|
| Feature engineering | Transform raw inputs into model-ready signals; train/serve consistency | Encoding, scaling, temporal & text features, feature stores | Feature engineering & feature stores |
| Evaluation & validation | Metrics, splits, tuning, fairness, explainability, production tests | Cross-validation, hyperparameter search, A/B and shadow tests | Model evaluation, selection & validation |
| Supervised — classification | Learn to assign labels from labeled examples | Logistic regression, decision trees, random forests, gradient boosting (XGBoost, LightGBM), SVM, neural networks | DATA-SCIENCE.md §1–3 |
| Supervised — regression | Learn to predict continuous values from labeled examples | Linear regression, polynomial regression, gradient boosting, neural networks | DATA-SCIENCE.md §1–3 |
| Unsupervised — clustering | Discover natural groupings in unlabeled data | K-means, DBSCAN, hierarchical clustering, Gaussian mixture models | DATA-SCIENCE.md §1–3 |
| Unsupervised — dimensionality reduction | Reduce feature space while preserving information | PCA, t-SNE, UMAP, autoencoders | — |
| Deep learning | Multi-layer neural networks for complex pattern recognition | CNNs (vision), RNNs/LSTMs (sequences), Transformers (language, multi-modal) | DATA-SCIENCE.md §1–3 |
| Natural language processing | Understanding and generating human language | Transformer models (BERT, GPT), word embeddings, named entity recognition, sentiment analysis | — |
| Computer vision | Understanding and analyzing images and video | CNNs, object detection (YOLO, Faster R-CNN), image segmentation, generative models | — |
| Time series | Forecasting and anomaly detection in temporal data | ARIMA, Prophet, LSTM, temporal convolutional networks, state-space models | — |
| Recommender systems | Predicting user preferences for items | Collaborative filtering, content-based filtering, matrix factorization, neural collaborative filtering | DATA-SCIENCE.md §1–3 |
| Reinforcement learning | Learning optimal actions through environment interaction | Q-learning, policy gradient, actor-critic, multi-armed bandits | — |
Selection guidance: Technique selection depends on problem type, data volume, interpretability requirements, latency constraints, and team expertise. See Data science & machine learning body of knowledge §3 (model evaluation) for the evaluation framework.
Keep project-specific model documentation in docs/product/ and experiment logs in docs/development/, not in this file.