Two taxonomies people often mix together
Supervised, unsupervised, semi-supervised and reinforcement learning describe how a system receives feedback. Linear models, trees, neural networks and transformers describe model architectures or families. A supervised task can use a tree or neural network, while a transformer can be trained with several learning objectives.
Keeping these dimensions separate makes comparison easier. Start by asking what examples and feedback exist, then ask which model family can represent the relationship at the required scale. Finally, add the product constraints: interpretability, latency, privacy, update frequency and cost.
Supervised prediction models
Supervised learning uses examples paired with desired outputs. Regression predicts continuous values such as demand or delivery time. Classification predicts categories or class probabilities such as fraud risk or defect type. Ranking models order alternatives, which is useful for search and recommendations.
The hardest part is often the label rather than the model. Historical decisions may encode bias, outcomes may arrive late and convenient proxies may not represent the real goal. Split data according to how the model will encounter the future, and keep a final test set isolated from model selection.
Unsupervised and representation models
Unsupervised methods look for structure without a labeled answer. Clustering groups similar examples, anomaly detection identifies unusual patterns and dimensionality reduction compresses or visualizes information. Representation learning produces embeddings that place related items near one another for search, recommendation or downstream prediction.
Because there is no single correct label, evaluation needs external evidence. A cluster should be stable and useful for a defined action; an embedding search system should retrieve relevant items for representative queries. Visual coherence alone is not enough.
Trees, ensembles and neural networks
Tree-based models partition feature space through rules and remain strong for many tabular datasets. Ensembles combine multiple trees to improve stability or correct errors. Neural networks learn layered representations and are especially effective for images, audio, language and other high-dimensional inputs.
The boundary is not ideological. A product can use a neural network to create embeddings and a boosted-tree model to make the final risk prediction. Hybrid systems often win because they assign each component the job it handles well.
Transformers and foundation models
Transformers use attention mechanisms to model relationships across sequences and now underpin many language and multimodal systems. Foundation models are trained on broad data and adapted through prompting, retrieval, tool use or fine-tuning. They can support many tasks without a separate model for each one.
Breadth creates new evaluation challenges. A foundation model may change behavior with prompt wording, context order or a provider update. Teams need versioned prompts, representative task suites and monitoring for regressions. The model’s general capability does not remove the need for application-specific evidence.
Choose by constraints, then prove the choice
Create a shortlist based on data volume, feature type, response time, explanation requirements and maintenance capacity. Establish a simple baseline, then compare more complex candidates on the same split and acceptance criteria. Include calibration when a probability drives action and segment-level results when performance may vary across users.
Before launch, test the surrounding pipeline and a rollback plan. A model that cannot be monitored or retrained with available staff is not operationally superior. Good model selection is a documented argument that connects data and error cost to a deployable system—not a preference for whichever architecture currently attracts the most attention.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.