Machine Learning - Enhancements
Automated Time Series Model Search
This feature enables the Exponential Smoothing algorithm to select the forecasting model type automatically - as well as related hyperparameters - when you do not specify the EXSM_MODEL setting. This can lead to more accurate forecasting models.
This feature automates the Exponential Smoothing algorithm hyperparameter search to produce better forecasting models without manual or exhaustive search. It enables non-expert users to perform time series forecasting without detailed understanding of algorithm hyperparameters while also increasing data scientist productivity.
Explicit Semantic Analysis Support for Dense Projection with Embeddings in OML4SQL
The unstructured text analytics algorithm Explicit Semantic Analysis (ESA) is able to output dense projections with embeddings, which are functionally equivalent to the popular doc2vec (document to vector) representation.
Producing a doc2vec representation is useful as input to other machine learning techniques, for example, classification and regression, to improve their accuracy when used solely with text or in combination with other structured data. Use cases include processing unstructured text from call center representative notes on customers or physician notes on patients along with other customer or patient structured data to improve prediction outcomes.
GLM Link Functions
The in-database Generalized Linear Model (GLM) algorithm now supports additional link functions for logistic regression: probit, cloglog, and cauchit.
These additional link functions expand the set available to match standard Generalized Linear Model (GLM) implementations. They enable increasing model quality, for example, accuracy, by handling a broader range of target column data distributions and expand the class of data sets handled. Specifically, the probit link function supports binary (for example, yes/no) target variables, such as when predicting win/lose, churn/no-churn, buy/no-buy. The asymmetric link function complementary-log-log (cloglog) supports binary target variables where one outcome is relatively rare, such as when predicting time-to-relapse of medical conditions. The cauchit link function supports handling data with, for example, data recording errors, more robustly.
Improved Data Prep for High Cardinality Categorical Features
This feature introduces the setting ODMS_EXPLOSION_MIN_SUPP to allow more efficient, data-driven encoding for high cardinality categorical columns. You can adjust the threshold (define the minimum support required) for the categorical values in explosion mapping or disable the feature, as needed.
This feature introduces a more efficient, data-driven encoding of high cardinality categorical columns, allowing users to build models without manual data preparation of such columns.
It efficiently addresses large datasets with millions of categorical values by recoding categorical values to include only those with sufficient support, enabling you to overcome memory limitations.
Lineage: Data Query Persisted with Model
This feature enables users to identify the data query that was used to provide the training data for producing a model. The BUILD_SOURCE column in the ALL/DBA/USER_MINING_MODELS view enables users to access the data query used to produce the model.
This feature records the query string that is run to specify the build data, within the model's metadata to better support the machine learning lifecycle and MLOps.
Multiple Time Series
The Multiple Time Series feature of the Exponential Smoothing algorithm enables conveniently constructing Time Series Regression models, which can include multivariate time series inputs and indicator data like holidays and promotion flags. It enables constructing Time Series Regression models to include multivariate time series inputs and indicator data like holidays and promotion flags.
This feature automates much of what a data scientist would perform manually by generating backcasts and forecasts on one or more input time series, where the target time series also receives confidence bounds. The result is used as input to other ML algorithms, for example, to support time series regression using XGBoost, with multivariate categorical, numeric, and time series variables.
OML4Py and OML4R Algorithm and Data Type Enhancements
The Oracle Machine Learning for Python (OML4Py) API exposes additional in-database machine learning algorithms, specifically Non-negative Matrix Factorization (NMF) for feature extraction, Exponential Smoothing Method (ESM) for time series forecasting, and Extreme Gradient Boosting (XGBoost) for classification and regression. OML4Py introduces support for date, time, and Integer datatypes.
The Oracle Machine Learning for R (OML4R) API exposes additional in-database machine algorithms, specifically Exponential Smoothing Method (ESM) for time series forecasting, Extreme Gradient Boosting (XGBoost) for classification and regression, Random Forest for classification, and Neural Network for classification and regression.
The enhancements to OML4R and OML4Py further enable Oracle Database as a platform for data science and machine learning, providing some of the most popular in-database algorithms from Python and R.
The additional in-database algorithms enable use cases such as demand forecasting using ESM, churn prediction and response modeling using Random Forest, and generating themes from document collections using NMF. As a feature extraction algorithm, NMF supports dimensionality reduction and as a data preparation step prior to modeling using other algorithms. XGBoost is a popular classification and regression algorithm due to its high predictive accuracy and also supports the machine learning technique survival analysis. Random Forest is a popular classification algorithm due to its high predictive accuracy. Neural Network is a classification and regression algorithm that is well-suited to data with noisy and complex patterns, such as found in sensor data, and provides fast scoring.
The OML4Py support for date, time, and integer data types enables operating on database tables and views that contain those data types, for example, to transform and prepare data at scale in the database.
Outlier Detection using Expectation Maximization (EM) Clustering
The Expectation Maximization algorithm is expanded to support distribution-based anomaly detection. The probability of anomaly is used to classify an object as normal or anomalous. The EM algorithm estimates the probability density of a data record, which is mapped to a probability of an anomaly.
Using Expectation Maximization (EM) for anomaly detection expands the set of algorithms available to support anomaly detection use cases, like fraud detection. Since different algorithms are capable of identifying patterns in data differently, having multiple algorithms available is beneficial when addressing machine learning use cases.
Partitioned Model Performance Improvement
This feature improves the performance for a high number of partitions (up to 32K component models) in a partitioned model and speeds up the dropping of individual models within a partitioned model.
Machine learning use cases often require building one model per subset of data, e.g., a model per state, region, customer, or piece of equipment. The partitioned models capability already automated the building of such models - providing a single model abstraction for simplified scoring - and this enhancement improves overall performance when using larger number of partitions.
XGBoost Support for Constraints and for Survival Analysis in OML4SQL
The in-database XGBoost algorithm is enahnced to support the machine learning technique survival analysis, as well as feature interaction constraints and monotonic constraints. The constrainsts allow you to choose how variables are allowed to interact.
Survival analysis is an important machine learning technique for multiple industries. This enhancement enables increased model accuracy when predicting, for example, equipment failures and healthcare outcomes. Specifically, this supports data scientists with the Accelerated Failure Time (AFT) model - one of the most used models in survival analysis – to complement the Cox proportional hazards regression model.
Interaction and monotonic constraints provide for greater control over the features used to achieve better predictive accuracy by leveraging user domain knowledge when specifying interaction terms.