Classical Machine Learning & Modeling

Anomaly detection identifies observations or patterns that depart from a chosen model of expected behavior. The skill includes defining what unusual means, selecting a detection method and calibrating alerts against investigation costs. An anomaly is a reason to examine an event, not proof of a defect, attack or fraudulent act.

Exploratory data analysis examines data structure, quality and relationships before committing to a model or inference. It uses summaries and visualizations to discover patterns and challenge assumptions. The skill is turning observations into testable questions and documented data decisions while keeping exploration separate from confirmatory evidence.

Model evaluation measures how well a model serves a defined task on appropriate data. It connects a test design, metrics and failure analysis to a deployment decision. The competence is selecting meaningful comparisons and interpreting uncertainty, rather than declaring a model good because one aggregate score is high.

Predictive analytics uses data to estimate unknown or future outcomes for a practical decision. It includes defining the target, building a prediction pipeline and translating scores into an action. The skill combines modeling with temporal availability, uncertainty and operational evaluation, so predictions remain useful when conditions differ from the training data.

Time series forecasting predicts future values from observations ordered in time and, where available, external information. The skill includes representing trend, seasonality and dependence, choosing a forecast horizon and testing predictions as they would have been made. Useful forecasts communicate uncertainty and outperform appropriate temporal baselines.

Hyperparameter optimization searches configurations that control model structure or training. It uses validation performance and a resource budget to choose among candidates. The competence is defining a meaningful search space and reliable objective while preventing leakage, overfitting to validation results and expensive searches that offer little improvement over sensible baselines.

Recommender systems rank or select items that may be useful to a user in a particular context. They learn from interactions, item information or both. The skill spans candidate generation, ranking and evaluation, including cold starts, exposure bias and the difference between predicting past engagement and improving the user experience.

Reinforcement learning learns an action policy from rewards obtained through interaction with an environment. Actions can affect future states as well as immediate outcomes. The skill includes defining the environment and reward, choosing a learning method and evaluating a policy's behavior under uncertainty, constraints and possible reward exploitation.

Multi-armed bandits learn which actions to choose while receiving feedback only for the selected action. They balance exploring uncertain alternatives with exploiting those that currently appear best. The competence includes reward design, exploration policy and evaluation of partially observed outcomes, especially when historical data were collected by a different selection policy.

Classical machine learning covers predictive and descriptive methods such as linear models, trees, kernels and clustering. The competence is selecting and validating a model appropriate to the data and decision. It emphasizes problem formulation, representation and reliable evaluation, often providing strong baselines before more complex neural approaches are considered.

Gradient boosting builds an additive model by repeatedly fitting learners to improve a specified loss. Decision-tree boosting is widely applicable to structured prediction problems. The skill includes choosing the objective, balancing learning rate and model capacity and validating whether additional stages improve generalization rather than merely fitting training residuals.

Regression analysis models how an outcome relates to explanatory variables and evaluates the uncertainty of that relationship. The skill includes choosing a functional form, interpreting parameters and checking assumptions. Prediction, association and causal estimation are different uses of regression and require different arguments beyond fitting the same numerical model.

Unsupervised learning finds patterns or representations in data without using task labels as training targets. It includes clustering, dimensionality reduction and density modeling. The skill is defining useful structure, choosing a representation and evaluating stability and usefulness, because a discovered pattern is not automatically a meaningful category or explanation.

Prophet is a forecasting library that models a time series through trend, seasonal components and optional event effects. The skill is preparing a suitable series, configuring those components and evaluating forecasts over realistic horizons. Its interpretable decomposition helps investigation, but automatic fitting does not replace temporal validation or domain knowledge.

AutoML automates parts of model and pipeline selection within a defined search space. It can compare preprocessing, estimators and configurations under a resource budget. The skill is setting the task and validation correctly, inspecting the selected pipeline and deciding whether automation has improved the result without concealing leakage, constraints or maintenance costs.

Optuna is a framework for defining and running hyperparameter optimization studies. A trial proposes values through an objective function, evaluates a configuration and may stop early when evidence suggests poor performance. Competence includes designing conditional spaces, selecting samplers and pruning rules and preserving a reliable record of the search.

CatBoost is a gradient-boosting library with mechanisms for categorical features and ordered training. The competence is preparing feature types, configuring an objective and evaluating the fitted model under the data's real boundaries. Native categorical support can simplify a pipeline, but it does not remove leakage or the need to examine generalization.

Classification predicts a discrete label or a distribution over labels from an observation. The skill includes defining classes, learning decision boundaries and selecting thresholds that match error costs. It also requires evaluating uncertainty and ambiguous cases, because a model's most likely class is not automatically an acceptable decision.

Decision trees predict by recursively partitioning feature space into regions with similar outcomes. Their conditional structure can be inspected as a sequence of tests. The skill includes controlling tree complexity, interpreting paths and validating stability, since a readable tree can still overfit or depend on misleading features.

Ensemble learning combines multiple models to produce one prediction. Different constructions reduce instability, correct errors sequentially or learn how to mix complementary predictors. The skill is creating genuinely useful diversity and evaluating the combined system, including the added computation and the danger of leakage when a second model learns from first-stage predictions.

LightGBM is a gradient-boosted tree library using histogram-based split search and, by default, leaf-wise tree growth. The competence includes preparing structured data, controlling tree capacity and validating the selected model. Its implementation choices affect memory, training behavior and overfitting, so they should be understood rather than reduced to a generic claim of speed.

Random forests combine many decision trees trained with sample and feature randomization. Averaging or voting reduces the instability of individual trees. The skill includes choosing forest capacity, evaluating predictions and interpreting importance cautiously, especially when correlated features, repeated entities or imbalanced labels affect the apparent result.

Supervised machine learning learns a mapping from inputs to labeled outcomes. Its central competence is constructing the learning problem so labels, available features and evaluation represent the intended decision. Model fitting comes after those choices, and cannot compensate for a target that is unreliable, leaked or disconnected from practical use.

Support vector machines learn a decision boundary using margin-based optimization, with kernels enabling nonlinear relationships. The skill includes choosing a representation, controlling regularization and evaluating a suitable kernel. Scaling and computational cost matter, and the decision score should not be mistaken for a calibrated probability without additional assessment.

XGBoost is a library for regularized gradient-boosted models, commonly using decision trees. The skill includes matching the objective to the task, controlling additive model complexity and building a consistent training and inference pipeline. Reliable use requires understanding validation, feature representation and iteration selection rather than relying on the library's reputation.

Cluster analysis groups observations according to a selected notion of similarity. The skill is choosing a representation and clustering model, interpreting groups and assessing their stability and usefulness. Clusters are outputs of assumptions and geometry; they are not automatically natural categories, causal mechanisms or ready-made labels for operational decisions.

Naive Bayes builds a probabilistic classifier using Bayes' rule and a conditional-independence approximation for features. It provides a compact baseline when the likelihood model matches the representation. Competence includes selecting the right variant, smoothing estimates and evaluating classification and probability quality separately.

HDBSCAN finds density-based groups through a hierarchy of density levels and selects persistent clusters. It can label observations as noise without fixing the number of clusters in advance. The skill is choosing a meaningful geometry, configuring cluster size and density requirements and evaluating whether stable-looking groups answer the substantive question.

SMOTE creates synthetic minority-class training examples by interpolating between neighboring minority observations. It is one option for handling class imbalance, not a guarantee of better detection. The skill includes choosing a valid representation, restricting resampling to training folds and evaluating performance on the original deployment-like class distribution.

SARIMA models a univariate time series using seasonal and nonseasonal autoregressive, differencing and moving-average components. The skill is identifying a defensible seasonal structure, selecting orders and checking residuals and forecasts. It is useful when recurring temporal dependence can be represented parsimoniously, but requires careful treatment of stationarity and structural change.

ARIMA models a time series through autoregressive terms, differencing and moving-average terms. The skill includes selecting a parsimonious specification, fitting it and checking whether residual behavior and held-out forecasts support its use. It provides a structured forecasting baseline rather than an automatic explanation of the process generating the series.

Class imbalance handling designs learning and decision rules when labels have unequal frequency or consequences. It can involve sampling, weighting, thresholds and better evaluation. The skill is choosing a response to the actual failure mode, rather than assuming that equalizing training counts will improve decisions at the population's original prevalence.

Logistic regression predicts class probabilities from a linear combination of features passed through a logistic or related multiclass link. The skill includes representation, regularization, coefficient interpretation and threshold choice. It is a transparent classification baseline, provided its probabilities and assumptions are checked for the population and decision in which it will be used.

K-nearest neighbors predicts from the labels or values of nearby stored examples under a chosen distance. It is a local, instance-based method rather than a compact fitted equation. The skill includes choosing a meaningful representation and neighborhood size and managing the quality and cost of retrieving neighbors at prediction time.

K-means partitions observations into a chosen number of groups by minimizing within-cluster squared Euclidean distances to centroids. The skill includes selecting features, scale and cluster count and assessing stability and interpretation. It is a useful geometric grouping method, but its objective does not establish that the resulting groups are meaningful categories.

Isolation Forest scores unusual observations by how readily randomized partitioning separates them from other data. It is an anomaly-detection method, not a supervised proof of harmful behavior. The skill includes feature preparation, threshold calibration and evaluation of alerts, especially when unusual but legitimate observations should not trigger action.

DBSCAN forms density-connected clusters using a neighborhood radius and a minimum density requirement. It can find irregular shapes and label observations outside selected dense regions as noise. The skill is choosing distance and density parameters and interpreting core, border and noise assignments without confusing algorithmic density with substantive meaning.

Cross-validation estimates predictive performance by repeatedly fitting on one part of the data and evaluating on another. The skill is selecting splits that reflect future use and keeping every learned transformation inside each training partition. It supports model comparison, but requires a separate view of the uncertainty and bias introduced by selection.

Linear regression estimates a continuous response through a linear combination of specified predictors. The skill includes selecting features and transformations, fitting coefficients and assessing assumptions, uncertainty and predictive error. Its simplicity supports inspection, but a linear coefficient is a conditional association unless a separate design justifies a causal interpretation.

Principal component analysis projects data onto orthogonal directions ordered by captured variance. It supports compression, visualization and removal of redundant linear dimensions. The skill includes choosing scale, fitting the projection without leakage and assessing lost information, because variance preserved by an unsupervised projection is not necessarily the information a downstream task needs.