H2O.ai Certified Machine Learning key terms

The balance between a model that is too simple and misses real patterns (high bias) and one that is too flexible and fits noise (high variance). Neither extreme generalizes well to unseen data.

A model learns patterns specific to the training data, including noise, so training performance is much better than performance on new data. Detected by comparing training and held-out validation metrics.

A resampling method that splits data into folds, trains on some folds and validates on the rest, then rotates. It produces a more stable estimate of generalization performance than a single split.

Data withheld from training and used only to estimate final model performance. It should be touched once at the end, since repeated tuning against it leaks information and inflates reported accuracy.

A table of predicted versus actual class labels showing true positives, true negatives, false positives, and false negatives. Most classification metrics, including precision and recall, can be derived from it.

Of all cases the model labeled positive, the fraction that truly are positive. High precision means few false alarms, which matters when acting on a positive prediction is costly.

Of all truly positive cases, the fraction the model correctly identified. High recall means few missed positives, which matters when failing to catch a case is costly.

Area under the receiver operating characteristic curve, which plots true positive rate against false positive rate across thresholds. It measures ranking quality independent of any single threshold; 0.5 is random.

The probability cutoff that converts a continuous model score into a class prediction. Moving it trades precision against recall without retraining the model, so it should be chosen from business costs.

Numeric scores summarizing prediction error for continuous targets. Common ones are RMSE, MAE, and R-squared; RMSE penalizes large errors more heavily, while MAE is more robust to outliers.

Information from the target or from the future leaks into training features, producing unrealistically good validation scores that collapse in production. Common causes are target-derived columns and fitting transforms before splitting.

A situation where one class is far rarer than another, so accuracy becomes misleading and models may ignore the minority. Remedies include resampling, class weights, and threshold tuning.

H2O.ai's open-source distributed machine learning platform, accessed through R, Python, or a REST API. It trains algorithms such as GLM, GBM, Random Forest, XGBoost, and Deep Learning on data held in an in-memory cluster.

A set of H2O worker nodes coordinated by a leader, sharing distributed data frames and running training in parallel. Cluster size and memory determine how much data can be handled in memory.

The distributed, column-oriented data structure inside an H2O cluster. It is created by importing or uploading data and is the input to H2O estimators; it is not a pandas DataFrame.

The web-based notebook interface bundled with H2O-3 for importing data, building models, and running experiments through a graphical UI. Useful for exploration and for users who prefer not to code.

H2O.ai's automated machine learning product that performs feature engineering, model selection, hyperparameter tuning, and interpretability automatically. It is aimed at rapid deployment and is distinct from the open-source H2O-3 library.

Model Object, Optimized: a lightweight, portable, low-latency artifact for scoring an H2O model in Java or another runtime without the full training environment. Preferred over the heavier binary model for production deployment.

Plain Old Java Object: a generated Java class that encapsulates a trained H2O model for scoring. It is larger and slower to load than a MOJO but historically supported more model types.

How an algorithm treats absent or unparseable values. H2O trees learn a default direction for missing values during training, while linear models may need imputation, so the choice affects results.

Converting a categorical column into several binary indicator columns, one per category. It lets linear models use categories but increases dimensionality and can create sparse data.

Replacing a categorical level with a statistic of the target, such as its mean, computed from training data. Powerful but prone to leakage unless computed within folds with smoothing.

Creating new input variables from raw data, such as ratios, date parts, or aggregations, to make patterns easier for models to learn. Good features often matter more than model choice.

Rescaling numeric features to zero mean and unit variance. It helps distance-based and regularized linear models converge and treats features comparably, but is unnecessary for tree models.

Automated model selection and tuning that trains many algorithms and hyperparameter settings, then ranks them on a validation metric. It reduces manual search but still requires a sensible metric and time budget.

Searching settings that control the learning process, such as tree depth, learning rate, or regularization strength, which are not learned from data. Methods include grid, random, and Bayesian search.

Evaluating every combination of a predefined hyperparameter grid. It is exhaustive and simple but scales poorly as dimensions grow, which is why random or Bayesian search is often preferred.

Halting training when validation performance stops improving for a set number of rounds. In gradient boosting it prevents overfitting and shortens run time by keeping the best iteration.

An ensemble method that builds trees sequentially, each new tree fitting the residual errors of the current ensemble. It is a strong default for tabular data when tuned and regularized carefully.

Penalizing model complexity, for example with L1 or L2 penalties, to reduce variance and improve generalization. Stronger penalties shrink coefficients or limit tree growth, trading some training fit for stability.

A ranking of how much each input contributes to model predictions, often computed from split gains in trees. It guides feature selection and sanity checks but can mislead when features are correlated.

Shapley additive explanations: a method that attributes a prediction to individual features using cooperative game theory. It gives consistent local explanations with sign and magnitude per row.

A plot showing the average predicted response as one feature varies, holding others fixed. It reveals the shape of a learned relationship but can be misleading when features are strongly correlated.

The difference between an observed value and the model's prediction. Plotting residuals against predictions or features reveals systematic error, non-linearity, and heteroscedasticity that aggregate metrics hide.

Degradation of model performance in production as the relationship between inputs and target changes or input distributions shift. It requires monitoring and periodic retraining to keep predictions reliable.

A versioned catalog of trained models with metadata, metrics, and artifacts, used to track lineage and control which version is deployed. It supports rollback and reproducibility across environments.

Generating predictions for a large stored dataset in one scheduled run rather than per request. It suits offline use cases and can use MOJO artifacts to avoid running the training cluster.

Serving predictions per incoming request with low latency, typically through a deployed REST endpoint. It requires a lightweight artifact, a defined input schema, and monitoring for drift and errors.