XGBoost

31.1 About XGBoost

Oracle Machine Learning for SQL XGBoost prepares training data, invokes XGBoost, builds and persists a model, and applies the model for prediction.

OML4SQL XGBoost is a scalable gradient tree boosting system that supports both classification and regression. It makes available the open source gradient boosting framework.

You can use XGBoost as a stand-alone predictor or incorporate it into real-world production pipelines for a wide range of problems such as ad click-through rate prediction, hazard risk prediction, web text classification, and so on.

The OML4SQL XGBoost algorithm takes three types of parameters: general parameters, booster parameters, and task parameters. You set the parameters through the model settings table. The algorithm supports most of the settings of the open source project.

Through XGBoost, OML4SQL supports a number of different classification and regression specifications, ranking models, and survival models. Binary and multiclass models are supported under the classification machine learning technique while regression, ranking, count, and survival are supported under the regression machine learning technique.

XGBoost also supports partitioned models and internalizes the data preparation. Currently, XGBoost is available only on Oracle Database Linux platform.

Related Topics

31.2 XGBoost Feature Constraints

Feature interaction constraints allow users to specify which variables can and cannot interact. By focusing on key interactions and eliminating noise, it aids in improving predicting performance. This, in turn, may lead to more generalized predictions.

The feature interaction constraints are described in terms of groupings of features that are allowed to interact. Variables that appear together in a traversal path in decision trees interact with one another because the condition of a child node is dependent on the condition of the parent node. These additional controls on model fit are beneficial to users who have a good understanding of the modeling task, including domain knowledge. OML4SQL supports more of the available XGBoost capabilities once these constraints are applied.

Monotonic constraints allow you to impose monotonicity constraints on the features in your boosted model. There may be a strong prior assumption that the genuine relationship is constrained in some way in many circumstances. This could be owing to commercial factors (just specific feature interactions are of interest) or the type of scientific subject under investigation. A typical form of constraint is that some features have a monotonic connection to the predicted response. In these situations, monotonic constraints may be employed to improve the model's prediction performance. For example, let X be the feature vector with features [x1,…, xi , …, xn] and ƒ(X) be the prediction response. Then ƒ(X) ≤ ƒ(X’) whenever xi ≤ xi’ is an increasing constraint; ƒ(X) ≥ ƒ(X’) whenever xi ≤ xi’ is a decreasing constraint. These feature constraints are listed in DBMS_DATA_MINING — Algorithm Settings: XGBoost.

31.3 XGBoost AFT Model

Survival analysis is a field of statistics that examines the time elapsed between one or more occurrences, such as death in biological organisms and failure in mechanical systems.

The goals of survival analysis include evaluating patterns of event times, comparing distributions of survival times in different groups of people, and determining if and how much certain factors affect the likelihood of an event of interest. The existence of censored data is an important feature of survival analysis. If a person does not experience an event within the observation period, they are labeled as censored. Censoring is a type of missing data problem in which the time to event is not recorded for a variety of reasons, such as the study being terminated before all enrolled subjects have demonstrated the event of interest, or the subject leaving the study before experiencing an event. Right censoring is defined as knowing only the lower limit l for the genuine event time T such that T > l. Right censoring will take place, for example, for those subjects whose birth date is known but who are still living when they are lost to follow-up or when the study concludes. We frequently come upon data that has been right-censored. The data is said to be left-censored if the event of interest occurred before the subject was included in the study but the exact date is unknown. Interval censoring occurs when an occurrence can only be described as occurring between two observations or examinations.

The Cox proportional hazards model and the Accelerated Failure Time (AFT) model are two major survival analysis methods. OML4SQL supports both these models.

Cox regression works for right censored survival time data. The hazard rate is the risk of failure (that is, the risk or likelihood of suffering the event of interest) in a Cox proportional hazards regression model, assuming that the subject has lived up to a particular time. The Cox predictions are returned on a hazard ratio scale. A Cox proportional hazards model has the following form:

h (t,x) = h₀(t)e^βx

Where h(t) is the baseline hazard, x is a covariate, and β is an estimated parameter that represents the covariate's effect on the outcome. A Cox proportional hazards model's estimated amount is understood as relative risk rather than absolute risk.

The AFT model fits models to data that can be censored to the left, right, or interval. The AFT model, which models time to an event of interest, is one of the most often used models in survival analysis. AFT is a parametric (it assumes the distribution of response data) survival model. The outcome of AFT models has a physical interpretation that is intuitive. The model has the following form:

ln Y = < W, X> + σZ

Where X is the vector in R^d representing the features. W is a vector consisting of d coefficients, each corresponding to a feature. <W, X> is the usual dot product in R^d. Y is the random variable modeling the output label. Z is a random variable of a known probability distribution. Common choices are the normal distribution, the logistic distribution, and the extreme distribution. It represents the “noise”. σ is a parameter that scales the size of noise.

AFT model that works with XGBoost or gradient boosting has the following form:

ln Y = T(x) + σZ

Where T(x) represents the output of a decision tree ensemble, using the supplied input x. Since Z is a random variable, you have a likelihood defined for the expression lnY=T(x)+σZ. As a result, XGBoost's purpose is to maximize (log) likelihood by fitting a suitable tree ensemble T(x).

The AFT parameters are listed in DBMS_DATA_MINING — Algorithm Settings: XGBoost.

31.4 Ranking Methods

Oracle Machine Learning supports pairwise and listwise ranking methods through XGBoost.

For a training data set, in a number of sets, each set consists of objects and labels representing their ranking. A ranking function is constructed by minimizing a certain loss function on the training data. Using test data, the ranking function is applied to get a ranked list of objects. Ranking is enabled for XGBoost using the regression function. OML4SQL supports pairwise and listwise ranking methods through XGBoost.

Pairwise ranking: This approach regards a pair of objects as the learning instance. The pairs and lists are defined by supplying the same case_id value. Given a pair of objects, this approach gives an optimal ordering for that pair. Pairwise losses are defined by the order of the two objects. In OML4SQL, the algorithm uses LambdaMART to perform pairwise ranking with the goal of minimizing the average number of inversions in ranking.

Listwise ranking: This approach takes multiple lists of ranked objects as learning instance. The items in a list must have the same case_id. The algorithm uses LambdaMART to perform list-wise ranking.

31.5 Scoring with XGBoost

Learn how to score with XGBoost.

The SQL scoring functions supported for a classification XGBoost model are PREDICTION, PREDICTION_COST, PREDICTION_DETAILS, PREDICTION_PROBABILITY, and PREDICTION_SET.

The scoring functions supported for a regression XGBoost model are PREDICTION and PREDICTION_DETAILS.

The prediction functions return the following information:

PREDICTION returns the predicted value.
PREDICTION_COST returns a measure of cost for a given prediction as an Oracle NUMBER. (classification only)
PREDICTION_DETAILS returns the SHAP (SHapley Additive exPlanation) contributions.
PREDICTION_PROBABILITY returns the probability for a given prediction. (classification only)
PREDICTION_SET returns the prediction and the corresponding prediction probability for each observation. (classification only)

Related Topics

SQL Scoring Functions