Algorithms
H2O-3 Secure provides the following supervised and unsupervised machine learning algorithms.
Supervised learning
In supervised learning, the dataset is labeled with the answer that the chosen algorithm should come up with. Supervised learning takes input variables (x) along with an output variable (y). The output variable represents the column that you want to predict on. The algorithm then uses these variables to learn and approximate the mapping function from the input to the output. Supervised learning algorithms support classification and regression problems.
H2O-3 Secure supports the following supervised algorithms:
| Algorithm | Description |
|---|---|
| AdaBoost | Combines many weak learners into a single classifier by increasing the weight of misclassified points on each iteration. Solves binary classification problems only. |
| ANOVA GLM | Calculates Type III sum of squares to show how much each predictor or interaction contributes to a model. |
| Cox Proportional Hazards (CoxPH) | Models time-to-event data using a hazard function built from a baseline hazard and a risk score. |
| Decision tree | Builds a single tree of tests on numeric features to classify or predict a binary target. |
| Deep Learning (Neural Networks) | Trains a multi-layer feedforward neural network with stochastic gradient descent and back-propagation. |
| Distributed Random Forest (DRF) | Builds a forest of classification or regression trees on row and column subsets and averages their predictions. |
| Generalized additive models (GAM) | Extends a generalized linear model with smooth functions of predictor variables. |
| Gradient Boosting Machine (GBM) | Builds regression trees sequentially, with each tree improving on the approximation of the previous one. |
| Generalized linear model (GLM) | Estimates regression models for outcomes that follow exponential distributions, including Gaussian, Poisson, binomial, and gamma. |
| Hierarchical Generalized Linear Model (HGLM) | Extends GLM with cluster-level effects for data collected in groups, such as students within schools. |
| Isotonic Regression | Fits a non-decreasing free-form line to an ordered sequence of observations for single-variable regression. |
| ModelSelection | Selects the best predictor subset for a GLM model using one of four search modes. |
| Naïve Bayes classifier | Classifies data by applying Bayes' theorem under an assumption of independence between predictors. |
| RuleFit | Fits a tree ensemble, builds a rule set from it, then fits a sparse linear model to the resulting rule and feature set. |
| Stacked ensembles | Finds the optimal combination of a collection of prediction algorithms by training a second-level model on their predictions. |
| Support Vector Machine (SVM) | Builds a model that assigns new examples to one of two categories. Solves binary classification problems only. |
| GAM thin plate regression spline | Extends GAM smoothers to work with one or more predictors using thin plate regression splines. |
| Distributed uplift random forest (Uplift DRF) | Trains uplift trees to model the incremental impact of a treatment using treatment and control group assignment. Supports binomial classification only. |
| XGBoost | Builds decision trees sequentially, with each new tree correcting the deficiencies of the previous model. |
Unsupervised learning
In unsupervised learning, the model is provided with a dataset that isn't labeled (i.e. without an explicit outcome that the algorithm should return). In this case, the algorithm attempts to find patterns and structure in the data by extracting useful features. The model organizes the data in different ways, depending on the algorithm (clustering, anomaly detection, autoencoders, etc).
H2O-3 Secure supports the following unsupervised algorithms:
| Algorithm | Description |
|---|---|
| Aggregator | Reduces a numerical or categorical dataset to fewer rows by clustering dense regions into exemplars, while keeping outliers as outliers. |
| Extended Isolation Forest | Generalizes Isolation Forest by randomizing the branch-cut slope, which reduces the bias introduced by axis-aligned splits. |
| Generalized low rank models (GLRM) | Reduces the dimensionality of a dataset by decomposing it into two smaller numeric matrices. |
| Isolation Forest | Detects anomalies by isolating observations through random feature and split-value selection, which produces shorter paths for outliers. |
| K-Means clustering | Partitions observations into groups so that points within a group resemble each other more than they resemble points in other groups. |
| Principal Component Analysis (PCA) | Transforms a set of possibly collinear features into a new set of uncorrelated features. |
| TF-IDF | Measures how important a word is to a document within a collection of documents. |
| Word2vec | Learns vector representations of words from a text corpus. |
Related topics
- Supported data types: Which data types each algorithm accepts, and how to handle timestamp columns.
- Early stopping: Stop a model build or grid search once it meets a stopping condition.
- Permutation variable importance: Measure how much a feature affects prediction error by permuting it and scoring the model again.
- Quantiles: Retrieve and display quantiles for parsed data.
- Target encoding: Replace a categorical value with the mean of the target variable.
- Submit and view feedback for this page
- Send feedback about H2O-3 Secure to cloud-feedback@h2o.ai