Every card audits the supplied blueprint. Open the detail panel for mechanisms, tuning, limitations and validation. Business examples are proposed experiments, not documented deployments. Select up to three entries to compare.
01 / Regression
Linear regression
Estimate a number using an additive model.
ŷ = β₀ + Σ βⱼxⱼ
Blueprint auditAccurate, simplified: linear means linear in coefficients; transformed features can represent curves. The intercept is implicit in Xθ only if X includes a constant column.
Mechanism, trade-offs & application
How it works
Ordinary least squares minimizes squared residuals. Ridge penalizes large coefficients; lasso can set coefficients to zero.
What to tune
Regularization strength; transformations; interactions.
Where it fails
Outliers and correlated features can destabilize estimates. Normal residuals are not required to compute predictions.
Illustrative application
Ruz: estimate construction cost from floor area, specification and location; coefficients describe association, not causal effects.
How to validate
Compare MAE and RMSE with a median baseline; hold out newer projects and inspect residuals by project type.
02 / Classification
Logistic regression
Estimate class probabilities from a linear score.
P(y=1|x) = 1 / (1 + exp(−(β₀ + βᵀx)))
Blueprint auditAccurate for binary classification: multinomial logistic regression uses softmax. A linear boundary is in the chosen feature space.
Mechanism, trade-offs & application
How it works
Fits log loss, usually with regularization. A decision threshold converts probability into an action.
What to tune
Penalty strength; class weights; decision threshold.
Where it fails
Probabilities are not automatically well calibrated; class weighting changes their interpretation.
Illustrative application
ImageFirm: rank inbound leads by conversion probability using only information available at arrival.
How to validate
Check log loss, precision–recall and calibration; choose a threshold using the cost of missed and wasted follow-ups.
03 / Classification
K-nearest neighbors classification
Let nearby labeled examples vote.
ŷ = argmax_c Σᵢ∈Nₖ(x) wᵢ · 1[yᵢ=c]
Blueprint auditIncomplete: similarity depends on representation, scaling and distance. Raw-pixel image classification is only a limited example.
Mechanism, trade-offs & application
How it works
Stores training examples; queries retrieve k neighbors and aggregate labels. Weights may be uniform or distance based.
What to tune
k; metric; weighting; feature scaling.
Where it fails
Distances lose usefulness in high dimensions; brute-force prediction costs roughly O(nd) per query.
Illustrative application
ImageFirm: classify support requests using normalized text embeddings and reviewed neighboring examples.
How to validate
Tune k inside cross-validation; check minority-class recall and prediction latency.
04 / Classification
Naive Bayes
Combine class priors with feature evidence.
P(y|x) ∝ P(y) ∏ⱼ P(xⱼ|y)
Blueprint auditCorrection: independence is conditional on the class, not unconditional feature independence.
Mechanism, trade-offs & application
How it works
Gaussian, multinomial and Bernoulli variants use different likelihoods. Smoothing avoids zero likelihoods for unseen tokens.
What to tune
Likelihood family; smoothing; class priors.
Where it fails
Correlated evidence can produce overconfident probabilities even when labels are useful.
Illustrative application
ImageFirm: create a fast text-routing baseline before paying for a larger model.
How to validate
Track macro-F1 and per-class errors; assess calibration if probabilities drive decisions.
05 / Classification · Regression
Decision tree
Split the feature space into simple rules.
Choose splits that reduce impurity or prediction loss.
Blueprint auditAccurate, qualified: small trees are readable; deep trees are not automatically interpretable.
Mechanism, trade-offs & application
How it works
Greedy splits create leaves that predict classes or numeric values. A locally best split need not yield a globally best tree.
What to tune
Maximum depth; minimum leaf size; pruning.
Where it fails
Small data changes can alter the structure. Regression leaves do not extrapolate trends beyond observed target values.
Illustrative application
Ruz: build a compact feasibility triage rule set for human review.
How to validate
Compare shallow and pruned trees on held-out projects; inspect rare leaves and unstable rules.
06 / Classification · Regression
Random forest
Average many deliberately different trees.
Regression: ŷ = (1/B) Σᵦ Tᵦ(x)
Blueprint auditIncomplete: random feature subsets at splits are central, alongside bootstrap sampling in the usual formulation.
Mechanism, trade-offs & application
How it works
Averaging decorrelated trees reduces variance. Out-of-bag samples provide an internal evaluation when bootstrapping is used.
What to tune
Tree count; feature subset size; leaf size; depth.
Where it fails
High accuracy is not guaranteed. Impurity importance can favor high-cardinality variables.
Illustrative application
Ruz: estimate project overruns from historical project attributes.
How to validate
Use held-out permutation importance; test on later projects and measure memory and serving cost.
07 / Classification · Regression
Gradient boosting
Add weak models that reduce the current loss.
Fₘ(x) = Fₘ₋₁(x) + ηhₘ(x)
Blueprint auditOverstated: strong tabular performance is common, not universal state-of-the-art. XGBoost and LightGBM are implementations with distinct designs.
Mechanism, trade-offs & application
How it works
New learners follow negative loss gradients; residual fitting is the squared-error special case.
What to tune
Learning rate; rounds; leaf complexity; subsampling; early stopping.
Where it fails
Careless validation and tuning can overfit. More rounds are not always better.
Illustrative application
ImageFirm: benchmark lead scoring against logistic regression under identical splits.
How to validate
Compare profit at a fixed review capacity, calibration and latency—not accuracy alone.
08 / Classification · Regression
Support vector machine
Prefer a boundary with a wide margin.
min ½‖w‖² + CΣᵢ max(0, 1 − yᵢ(wᵀxᵢ+b))
Blueprint auditIncomplete: kernels create nonlinear boundaries in input space; support vector regression is also available.
Mechanism, trade-offs & application
How it works
Support vectors determine the decision surface. Kernel methods compare observations implicitly in another feature space.
What to tune
C; kernel; gamma for RBF; scaling.
Where it fails
Kernel training can become costly with many samples; scores are not probabilities without additional calibration.
Illustrative application
ImageFirm: benchmark a linear SVM for sparse text classification.
How to validate
Use consistent scaling in a pipeline; compare PR metrics and kernel serving latency.
09 / Clustering
K-means
Represent observations with k centroids.
min Σᵢ ‖xᵢ − μcᵢ‖²
Blueprint auditIncomplete: standard k-means minimizes squared Euclidean distance, not an arbitrary distance function.
Mechanism, trade-offs & application
How it works
Alternate assigning points to nearest centers and recomputing means. Initialization affects the local solution.
What to tune
k; scaling; initialization; multiple restarts.
Where it fails
Poor fit for curved groups, unequal density and outliers. Lower inertia alone cannot choose k because it decreases as k grows.
Illustrative application
ImageFirm: explore customer behavior segments using comparable, standardized features.
How to validate
Compare silhouette, stability across seeds and whether segments support a distinct useful action.
10 / Clustering
Hierarchical clustering
Build a nested hierarchy of groups.
Merge groups according to a linkage rule.
Blueprint auditMisleading: a hierarchy can be built without k, but obtaining final groups requires a cut rule such as k or a distance threshold.
Mechanism, trade-offs & application
How it works
Agglomerative clustering merges groups; divisive methods split them. Linkage defines how group distances are measured.
What to tune
Linkage; metric; cut threshold; connectivity.
Where it fails
Single linkage can chain groups. Ward linkage uses Euclidean variance; dense distance storage often scales quadratically.
Illustrative application
Ruz: explore a hierarchy of comparable developments by specification and location.
How to validate
Check whether groupings survive reasonable metric and linkage changes before naming segments.
11 / Clustering
DBSCAN
Connect dense neighborhoods and leave sparse points unassigned.
Core point: |Nε(x)| ≥ min_samples
Blueprint auditQualified: handles non-convex shapes, but one density threshold can fail when cluster densities vary. Noise does not automatically mean a business anomaly.
Mechanism, trade-offs & application
How it works
Core points connect dense regions; border points attach to them; remaining points receive a noise label.
What to tune
Neighborhood radius ε; minimum samples; distance metric.
Where it fails
Scaling and dimensionality strongly affect neighborhoods. Some implementations can require quadratic memory.
Illustrative application
Ruz: explore geographic concentration of comparable projects using an appropriate geographic distance.
How to validate
Inspect noise rate and sensitivity to ε; validate flagged records manually.
12 / Representation
Principal component analysis
Rotate data toward directions of greatest variance.
Xcentered = UΣVᵀ; Z = Xcentered Vᵣ
Blueprint auditIncomplete: high variance is not the same as predictive information. PCA is unsupervised and components are uncorrelated, not necessarily independent.
Mechanism, trade-offs & application
How it works
Projects centered features onto orthogonal directions; a truncated representation discards remaining variation.
What to tune
Component count; standardization; optional whitening.
Where it fails
A low-variance feature may contain the signal you need. Components can be hard to interpret.
Illustrative application
ImageFirm: compress correlated telemetry before a downstream model.
How to validate
Fit PCA only on training folds; compare downstream validation performance as well as retained variance.
13 / Representation
t-SNE
Make a local-neighborhood map of high-dimensional observations.
Minimize KL(P ‖ Q) between pairwise similarity distributions.
Blueprint auditIncomplete: local structure is emphasized, not perfectly preserved; island sizes and inter-island distances are not reliable measurements.
Mechanism, trade-offs & application
How it works
Uses heavy-tailed similarities in the embedding and a non-convex optimization objective.
What to tune
Perplexity; learning rate; initialization; seed.
Where it fails
Different runs may tell different visual stories. Standard scikit-learn t-SNE has no transform for new observations.
Illustrative application
ImageFirm: inspect document embeddings for possible labeling issues, then verify in the original space.
How to validate
Compare several seeds and perplexities; never treat a separated island as proof of a real class.
14 / Classification · Representation
Linear discriminant analysis
Model classes with a shared covariance.
δc(x) = xᵀΣ⁻¹μc − ½μcᵀΣ⁻¹μc + log πc
Blueprint auditIncomplete: Gaussian assumptions are class-conditional, with shared covariance. This is not latent Dirichlet allocation.
Mechanism, trade-offs & application
How it works
Produces linear class boundaries; supervised dimensionality reduction has at most C−1 components for C classes.
What to tune
Shrinkage; class priors; solver.
Where it fails
Covariance estimates can be unstable with many features and few examples.
Illustrative application
ImageFirm: try a compact supervised projection for labeled quality categories.
How to validate
Compare shrinkage LDA with logistic regression using a stratified holdout.
15 / Classification
Quadratic discriminant analysis
Allow each class its own covariance.
δc(x) = −½ log|Σc| − ½(x−μc)ᵀΣc⁻¹(x−μc) + log πc
Blueprint auditAccurate but incomplete: Gaussian class-conditional densities remain assumed; class-specific covariance creates quadratic boundaries.
Mechanism, trade-offs & application
How it works
Estimates separate covariance matrices; flexibility increases the number of fitted parameters.
What to tune
Covariance regularization; class priors.
Where it fails
Each class needs enough information to estimate covariance; singular matrices and overfitting are risks.
Illustrative application
ImageFirm: test whether quality classes differ in feature variability, not only their means.
How to validate
Compare regularized QDA with LDA; examine class-specific validation errors.
16 / Regression
K-nearest neighbors regression
Average the outcomes of similar observations.
ŷ = Σᵢ∈Nₖ wᵢyᵢ / Σᵢ∈Nₖ wᵢ
Blueprint auditQualified: time-series forecasting needs lag features and chronological evaluation; neighbor averaging is not inherently a temporal model.
Mechanism, trade-offs & application
How it works
Local averaging produces a piecewise prediction surface. More neighbors smooth predictions but can erase local differences.
What to tune
k; weights; metric; scaling.
Where it fails
Poor extrapolation; irrelevant features distort similarity. Training data must remain available at inference.
Illustrative application
Ruz: estimate apartment value from comparable sales after accounting for location and date.
How to validate
Use geographic or temporal holdouts; compare MAE against a local median baseline.
17 / Neural models
Multilayer perceptron
Compose learned transformations with nonlinear activations.
h = φ(W₁x+b₁); ŷ = g(W₂h+b₂)
Blueprint auditQualified: MLP is a neural architecture family. Large data requirements depend on capacity, regularization and the task.
Mechanism, trade-offs & application
How it works
Backpropagation computes gradients; an optimizer updates weights. Output activation and loss must match the target.
What to tune
Width; depth; learning rate; batch size; weight decay.
Where it fails
Training is non-convex; scaling, random seeds and early stopping matter.
Illustrative application
ImageFirm: benchmark nonlinear prediction on structured creative-performance features.
How to validate
Compare with boosted trees; repeat seeds and inspect learning curves before increasing capacity.
18 / Neural models
Convolutional neural network
Reuse local filters across structured inputs.
hᵢⱼ = φ(Σᵤᵥ Kᵤᵥ xᵢ₊ᵤ,ⱼ₊ᵥ + b)
Blueprint auditIncomplete: convolutions also serve 1D signals and 3D volumes. Cost depends on architecture, input size and hardware.
Mechanism, trade-offs & application
How it works
Local connectivity and shared weights exploit spatial structure. Convolution supports translation equivariance; full invariance is not automatic.
What to tune
Kernel size; stride; channels; augmentation; pretrained backbone.
Where it fails
Background shortcuts and camera changes can break performance despite good random-split scores.
Illustrative application
Ruz: assist defect screening from site photos, with expert review of uncertain cases.
How to validate
Hold out entire sites or cameras; test per-defect recall and deployment latency.
19 / Neural models
Recurrent neural network / LSTM
Carry a learned state through a sequence.
hₜ = φ(Wₓxₜ + Wₕhₜ₋₁ + b)
Blueprint auditIncomplete: LSTM is a gated RNN variant. Its state is a learned summary, not reliable long-term recall of every input.
Mechanism, trade-offs & application
How it works
Recurrent weights are shared across time. LSTM gates control information flow and help address vanishing gradients.
What to tune
Sequence length; hidden size; layers; gradient clipping.
Where it fails
Sequential computation limits parallelism. Future information in features or normalization causes leakage.
Illustrative application
Ruz: forecast occupancy from historical observations and genuinely known future covariates.
How to validate
Use rolling-origin evaluation against seasonal-naive and lagged tabular baselines.
20 / Reinforcement learning
Q-learning
Learn the long-term value of actions from transitions.
Q(s,a) ← Q(s,a) + α[r + γ maxₐ′ Q(s′,a′) − Q(s,a)]
Blueprint auditMisleading: reinforcement learning is a paradigm; Q-learning is one algorithm. Optimality is conditional, not a universal promise.
Mechanism, trade-offs & application
How it works
Off-policy temporal-difference learning updates action values. At terminal states, the bootstrap term is zero.
What to tune
Learning rate α; discount γ; exploration schedule.
Where it fails
Tabular convergence needs suitable finite-MDP conditions, sufficient visits and learning-rate conditions. Neural approximation does not inherit the guarantee.
Illustrative application
Ruz: explore simulated energy-control decisions before any real-world trial; reward must include comfort constraints.
How to validate
Evaluate unseen scenarios, constraint violations and return variability—not training reward alone.