The glossary provides brief definitions of terms relevant to the Arkangel AI platform. These terms span all phases of machine learning, from data to deployment.


Accuracy Over Space

A model Leaderboard tab and Location AI insight that provides a spatial residual mapping within an individual model.

Accuracy Over Time

A model Leaderboard tab that visualizes how predictions change over time.

ACE scores

Also known as Alternating Conditional Expectations. A univariate measure of the correlation between the feature and the target. ACE scores detect non-linear relationships, but as they are univariate, they do not detect interaction effects.

Advanced Tuning

The ability to manually set model parameters after the model build, supporting experimentation with parameter settings to improve model performance.


(Predictions) Actual values for an ML model that let you track its prediction outcomes. To generate accuracy statistics for a deployed model, you compare the model's predictions to real-world actual values for the problem. Both the prediction dataset and the actuals dataset must contain association IDs, which let you match up corresponding rows in the datasets to gauge the model's accuracy.

Aggregate image featur

(Visual AI) A set of image features where each individual element of that set is a constituent image feature. For example, the set of image features extracted from an image might include a set of features indicating:

  1. The colors of the individual pixels in the image.

  2. Where edges are present in the image.

  3. Where faces are present in the image.

From the aggregate it may be possible to determine the impact of that feature on the output of a data analytics model and compare that impact to the impacts of the model's other features.

AI Catalog

A browsable and searchable collection of registered objects that contains definitions and relationships between various objects types. Items stored in the catalog include: data connections, data sources, data metadata.


The second phase of Exploratory Data Analysis (i.e., EDA2), that determines feature importances based on cross-correlation with the target feature. That data determines the “informative features” used for modeling during Autopilot.

Alternating Conditional Expectations

See ACE scores.

Anomaly detection

A form of unsupervised learning used to detect anomalies in data. Anomaly detection, also referred to as outlier or novelty detection, can be useful with data having a low percentage of irregularities or large amounts of unlabeled data.


(Data Prep) The published result of your data prep steps. You can export the results of all of your data prep steps to an AnswerSet or you can create a lens to specify the set of steps to export as an AnswerSet.

ARIMA (AutoRegressive Integrated Moving Average)

A class of time series model that projects the future values of a series based entirely on the patterns of that series.

Association ID

(MLOps) An identifier that functions as a foreign key for your prediction dataset so you can later match up actual values (or "actuals") with the predicted values from the deployed model. An association ID is required for monitoring the accuracy of a deployed model.

AUC (Area Under the Curve)

A common error metric for binary classification that considers all possible thresholds and summarizes performance in a single value on the ROC Curve. It works by optimizing the ability of a model to separate the 1s from the 0s. The larger the area under the curve, the more accurate the model.

Augmented Intelligence

Arkangel AI's enhanced approach to artificial intelligence, which expands current model building and deployment assistance practices. The Arkangel AI platform fully automates and governs the AI lifecycle from data ingest to model training and predictions to model-agnostic monitoring and governance. Guardrails ensure adherence to data science best practices when creating machine learning models and AI applications. Transparency across user personas and access to data wherever it resides avoids lock-in practices and operate strategically across clouds and on-premises.

Automatic Project Flows (APF)

(Data Prep) Functionality that allows you to operationalize curated data flows. Use APFs to schedule a sequence of data prep steps across projects, datasets, and AnswerSets. Then you manage runs using APF’s monitoring capabilities.

AutoML (Automated Machine Learning)

A software system that automates many of the tasks involved in preparing a dataset for modeling and performing a model selection process to determine the performance of each with the goal of identifying the best performing model for a specific use case.

Autopilot (full Autopilot)

The Arkangel AI "survival of the fittest" modeling mode that automatically selects the best predictive models for the specified target feature and runs them at ever-increasing sample sizes. In other words, it runs more models in the early stages on a small sample size and advances only the top models to the next stage. By default, Arkangel AI runs models at 16% of total data and advances the top 16 models, then runs those at 32%. Taking the top 8 models from that run, Arkangel AI runs on 64% of the data (or 500MB of data, whichever is smaller).

AutoTS (Automated time series)

A software system that automates all or most of the steps needed to build forecasting models, including featurization, model specification, model training, model selection, validation, and forecast generation.

Average baseline

(Time-aware) The average of the target in the Feature Derivation Window.



(Time-aware) The time series equivalent of cross-validation. Unlike cross-validation, however, backtests allow you to select specific time periods or durations for your testing instead of random rows, creating “trials” for your data.

Base dataset

(Data Prep) The data imported into a Data Prep project on which all actions are performed.

Baseline model

(AutoML, Time-aware) Also known as a naive model. A simple model used as a comparison point to confirm that a generated ML model is learning with more accuracy than a basic non-ML model.

For example, generated ML models for a regression project should perform better than a baseline model that predicts the mean or median of the target. Generated ML models for a time series project should perform better than a baseline model that predicts the future using the most recent actuals (i.e., using today's actual value as tomorrow's prediction).

For time series projects, baseline models are used to calculate the MASE metric (the ratio of the MAE metric over the baseline model).

Batch predictions

(MLOps, Predictions) A method of making predictions with large datasets, in which you pass input data and get predictions for each row; predictions are written to output files. Users can make batch predictions with MLOps via the Predictions interface or can use the Batch Prediction API for automating predictions. Schedule batch prediction jobs by specifying the prediction data source and destination and determining when the predictions will be run.

Bias Mitigation

Augments blueprints with a pre- or post-processing task intended to reduce bias across classes in a protected feature. Bias Mitigation is also a model Leadboard tab (Bias and Fairness > Bias Mitigation) where you can apply mitigation techniques after Autopilot has finished.

Bias vs Accuracy

A Leaderboard tab that generates a chart to show the tradeoff between predictive accuracy and fairness, removing the need to manually note each model's accuracy score and fairness score for the protected features.

"Blind History"

(Time-aware) “Blind history" captures the gap created by the delay of access to recent data (e.g., “most recent” may always be one week old). It is defined as the period of time between the smaller of the values supplied in the Feature Derivation Window and the forecast point. A gap of zero means "use data up to, and including, today;" a gap of one means "use data starting from yesterday" and so on.


A model that potentially increases accuracy by combining the predictions of between two and eight models. Arkangel AI can be configured to automatically create blender models as part of Autopilot, based on the top three regular Leaderboard models (for AVG, GLM, and ENET blenders). You can also create blenders manually (aka ensemble models).


"Can't operationalize" period

(Time-aware) The "can't operationalize" period defines the gap of time immediately after the Forecast Point and extending to the beginning of the Forecast Window. It represents the time required for a model to be trained, deployed to production, and to start making predictions—the period of time that is too near-term to be useful. For example, predicting staffing needs for tomorrow may be too late to allow for taking action on that prediction.


The center of a cluster generated using unsupervised learning. A centroid is the multi-dimensional average of a cluster, where the dimensions are observations (data points).

CFDS (Customer Facing Data Scientist)

An Arkangel AI employee responsible for the technical success of user and potential users. They assist with tasks like structuring data science problems to complete integration of Arkangel AI. CFDS are passionate about ensuring user success.

Challenger models

(MLOps) Models that you can compare to a currently deployed model (the "champion" model) to continue model comparison post-deployment. Submit a challenger model to shadow a deployed model and replay predictions made against the champion to determine if there is a superior Arkangel AI model that would be a better fit.

Champion model

(Time-aware, MLOps) A model recommended by Arkangel AI—for a deployment (MLOps) or for segmented modeling.

In MLOps, you can replace the champion selected for a deployment yourself or you can set up automatic retraining (Continuous AI) where Arkangel AI compares challenger models with the champion model and replaces the champion model if a challenger outperforms the champion.

In the segmented modeling workflow, Arkangel AI builds a model for each segment. Arkangel AI recommends the best model for each segment—the segment champion. The segment champions roll up into a Combined Model. For each segment, you can select a different model as champion, which is then used in the Combined Model.


The connection between an output port of one module and an input port of another module. Data flows from one module's output port to another module's input port via a channel, represented visually by a line connecting the two.


A type of prediction problem that classifies values into discrete, final outcomes or classes. Binary classification problems are those datasets in which what you are trying to predict can be one of two classes (for example, "yes" or "no"). Multiclass classification is a classification problem that results in more than two outcomes (for example, "buy", "sell", or "hold"). Unlimited multiclass is the ability to handle projects with a target feature containing an unlimited number of classes, with support for both a high threshold of individual classes and multiclass aggregation to support an unlimited number of classes above the threshold.

ClicktoPrep link

(Data Prep) A link to Data Prep components from business intelligence (BI) and data visualization tools, (for example Tableau®). ClicktoPrep links in a BI or visualization tool can take you to the last step in a Data Prep project, a Data Prep Filtergram, or to a specific Data Prep project step. You can make changes to the Data Prep data, then republish and refresh the visualization or report in your BI or visualization tool.


A form of unsupervised learning used to group similar data and identify natural segments.


A model Leaderboard tab (Describe > Coefficients) that provides a visual indicator of information that can help you refine and optimize your models.

Combined Model

(Time-aware) The final model generated in the segmented modeling workflow. With segmented modeling, Arkangel AIArkangel AI builds a model for each segment and combines the segment champions into a single Combined Model that you can deploy.

Common event

(Time-aware) A data point is a common event if it occurs in a majority of weeks in data (for example, regular business days and hours would be common, but an occasional weekend data point would be uncommon).

Compliance documentation

Automated model development documentation that can be used for regulatory validation. The documentation provides comprehensive guidance on what constitutes effective model risk management.

Composable ML

A code-centric feature, designed for data scientists, that allows applying custom preprocessing and modeling methods to create a blueprint for model training. Using built-in and custom tasks, you can compose and then integrate the new blueprint with other Arkangel AI features to augment and improve machine learning pipelines.


A modeling mode that runs all Repository blueprints on the maximum Autopilot sample size to ensure more accuracy for models.

Computer vision

(Visual AI) Use of computer systems to analyze and interpret image data. Computer vision tools generally use models that incorporate principles of geometry to solve specific problems within the computer vision domain. For example, computer vision models may be trained to perform object recognition (recognizing instances of objects or object classes in images), identification (identifying an individual instance of an object in an image), detection (detecting specific types of objects or events in images), etc.

Computer vision tools/techniques

(Visual AI) Tools—for example, models, systems—that perform image preprocessing, feature extraction, and detection/segmentation functions.

Confusion matrix

A table that reports true versus predicted values. The name “confusion matrix” refers to the fact that the matrix makes it easy to see if the model is confusing two classes (consistently mislabeling one class as another class). The confusion matrix is available as part of Arkangel AI's ROC Curve, Eureqa, and Confusion Matrix for multiclass model visualizations.


A model Leaderboard tab (Describe > Constraints) that allows you to review monotonically constrained features if feature constraints were configured in Advanced Options prior to modeling.

Continuous AI

(MLOps) Retraining strategies for MLOps that refresh production models based on a schedule or an event (for example, a drop in accuracy or data drift). Continuous AI also uses Arkangel AI's AutoML to automatically create and recommend new challenger models. When combined, these strategies maximize accuracy and timely predictions.


(AI Catalog) Information used to authenticate and authorize actions against data connections. The most common connection is through username and password, but alternate authentication methods include LDAP, Active Directory, and Kerberos.

Cross-Class Accuracy

A model Leadboard tab (Bias and Fairness > Cross-Class Accuracy) that helps to shows why the model is biased, and where in the training data it learned the bias from. Bias and Fairness settings must be configured.

Cross-Class Data Disparity

A model Leadboard tab (Bias and Fairness > Cross-Class Data Disparity) that calculates, for each protected feature, evaluation metrics and ROC curve-related scores segmented by class. Bias and Fairness settings must be configured.


Also known as CV. A type of validation partition that is run to test (validate) model performance. Using subsets ("folds") of the validation data, Arkangel AI creates one model per fold, with the data assigned to that fold used for validation and the rest of the data used for training. By default, Arkangel AI uses five-fold cross-validation and presents the mean of those five scores on the Leaderboard. See also validation.

Custom inference models

(MLOps) User-created, pre-trained models uploaded as a collection of files via the Custom Model Workshop. Upload a model artifact to create, test, and deploy custom inference models to Arkangel AI's centralized deployment hub. An inference model can have a predefined input/output schema or it can be unstructured. To customize prior to model training, use custom tasks.

Custom model workshop

(MLOps) In the Model Registry, a location where you can upload user-created, pre-trained models as a collection of files. You can use these model artifacts to create, test, and deploy custom inference models to Arkangel AI's centralized deployment hub.

Custom task

A data transformation or ML algorithm, for example, XGBoost or One-hot encoding, that can be used as a step in an ML blueprint inside Arkangel AI and used for model training. Tasks are written in Python or R and are added via the Custom Model Workshop. Once saved, the task can be used when modifying a blueprint with Composable ML. To deploy a pre-trained model where re-training is not required, use custom inference models.


See Cross Validation.


Data drift

(MLOps) The difference between values in new inference data used to generate predictions for models in production and the training data initially used to train the deployed model. Predictive models learn patterns in training data and use that information to predict target values for new data. When the training data and the production data change over time, causing the model to lose predictive power, the data surrounding the model is said to be drifting. Data drift can happen for a variety of reasons, including data quality issues, changes in feature composition, and even changes in the context of the target variable.

Data management

The umbrella term related to loading, cleaning, transforming, and storing data within Arkangel AI. It also refers to the practices that companies follow when collecting, storing, using, and deleting data.

Data Prep

When you gather, explore, and prepare data from multiple sources for machine learning.

Data preparation

The process of transforming raw data to the point where it can be run through machine learning algorithms to uncover insights or make predictions. Also called “data preprocessing.”

Data Prep library

(Data Prep) The Data Prep component (and page) where you add and manage datasets, including AnswerSets that you publish from your Data Prep projects. Select Library on the top left of the Data Prep window to access the library. In the Data Prep library, you can also export datasets, set them up for automation, add new versions, and create profiles for your datasets.

Data Prep project

(Data Prep) The Data Prep component (and page) that contains your projects. Select Projects on the top left of your Data Prep window to access all of your projects. The Projects page is where you access and manage your projects, as well as those of the other users of your Data Prep instance. You can create new projects on the Projects page or you can create projects by uploading datasets on the Library page.

Data Quality Handling Report

A model Leaderboard tab (Describe > Data Quality Handling Report) that analyzes the training data and provides the following information for each feature: feature name, variable type, row count, percentage, and data transformation information.


Data, a file or the content of a data source, at a particular point in time. A data source can produce multiple datasets; an AI Catalog dataset has exactly one data source. In Data Prep, a dataset can be generated from multiple data sources. In AI Catalog, a dataset is materialized data that is stored with a catalog version record. There may be multiple catalog version records associated with an entity, indicating that Arkangel AI has reloaded or refreshed the data. The older versions are stored to support existing projects, new projects use the most recent version. A dataset can be in one of two states:

  • A "snapshotted" (or materialized) dataset is an immutable snapshot of data that has previously been retrieved and saved.

  • A “remote” (or unmaterialized) dataset has been configured with a location from which data is retrieved on-demand (AI Catalog).

Data connection

A configured connection to a database—it has a name, a specified driver, and a JDBC URL. You can register data connections with Arkangel AI for ease of re-use. A data connection has one connector but can have many data sources.

Data source

A configured connection to the backing data (the location of data within a given endpoint). A data source specifies, via SQL query or selected table and schema data, which data to extract from the data connection to use for modeling or predictions. Examples include the path to a file on HDFS, an object stored in S3, and the table and schema within a database. A data source has one data connection and one connector but can have many datasets. It is likely that the features and columns in a datasource do not change over time, but that the rows within change as data is added or deleted.

Data stage

Intermediary storage that supports multipart upload of large datasets, reducing the chance of failure when working with large amounts of data. Upon upload, the dataset is uploaded in parts to the data stage, and once the dataset is whole and finalized, it is pushed to the AI Catalog or Batch Predictions. At any time after the first part is uploaded to the data stage, the system can instruct Batch Predictions to use the data from the data stage to fill in predictions.

Deep learning

A set of algorithms that run data through several “layers” of neural network algorithms, each of which passes a simplified representation of the data to the next layer. Deep learning algorithms are essential to Arkangel AI's Visual AI capabilities, and their processing can be viewed from the Training Dashboard visualization.

Deployment inventory

(MLOps) The central hub for managing deployments. Located on the Deployments page, the inventory serves as a coordination point for stakeholders involved in operationalizing models. From the inventory, you can monitor deployed model performance and take action as necessary, managing all actively deployed models from a single point.


(Visual AI) A computer vision technique that involves the selection of a subset of the input image data for further processing (for example, one or more images within a set of images or regions within an image).

Downloads tab

A model Leaderboard tab (Predict > Downloads) where you can download model artifacts.


See Smart downsampling.


(AI Catalog) The software that allows the Arkangel AI application to interact with a database; each data connection is associated with one driver (created and installed by your administrator). The driver configuration saves the JAR file storage location in Arkangel AI and any additional dependency files associated with the driver. Arkangel AI supports JDBC drivers.


EDA (Exploratory Data Analysis)

The Arkangel AI approach to analyzing and summarizing the main characteristics of a dataset. Generally speaking, there are two stages of EDA:

  • EDA1 provides summary statistics based on a sample of data. In EDA1, Arkangel AI counts, categorizes, and applies automatic feature transformations (where appropriate) to data.

  • EDA2 is a recalculation of the the statistics collected in EDA1 but using the entire dataset, excluding holdout. The results of this analysis are the criteria used for model building.

Ensemble models

See blender.


A Docker container where a custom task runs.


Exploratory Spatial Data Analysis (ESDA) is the exploratory data phase for Location AI. Arkangel AI provides a variety of tools for conducting ESDA within the Arkangel AI AutoML environment, including geometry map visualizations, categorical/numeric thematic maps, and smart aggregation of large geospatial datasets.


Model blueprints for Eureqa generalized additive models (Eureqa GAM), Eureqa regression, and Eureqa classification models. These blueprints use a proprietary Eureqa machine learning algorithm to construct models that balance predictive accuracy against complexity.

EWMA (Exponentially Weighted Moving Average)

A moving average that places a greater weight and significance on the most recent data points, measuring trend direction over time. The "exponential" aspect indicates that the weighting factor of previous inputs decreases exponentially. This is important because otherwise a very recent value would have no more influence on the variance than an older value.

External stage

An external stage is a cloud location outside of the Snowflake environment used for loading and unloading data for Snowflake. The cloud location can be either Amazon S3 or Microsoft Azure storage.


Fairness score

(Bias and Fairness) A numerical computation of model fairness against the protected class, based on the underlying fairness metric.

Fairness Threshold

(Bias and Fairness) The measure of whether a model performs within appropriate fairness bounds for each protected class. It does not affect the fairness score or performance of any protected class.

Fairness Value

(Bias and Fairness) Fairness scores normalized against the most favorable protected class (i.e., the class with the highest fairness score).

Favorable Outcome

(Bias and Fairness) A value of the target that is treated as the favorable outcome for the model. Predictions from a binary classification model can be categorized as being a favorable outcome (i.e., good/preferable) or an unfavorable outcome (i.e., bad/undesirable) for the protected class.


See Feature Derivation Window.


A column in a dataset, also called "variable" or "feature variable." The target feature is the name of the column in the dataset that you would like to predict.

Feature Derivation Window

(Time-aware) Also known as FDW. A rolling window of past values that models use to derive features for the modeling dataset. Consider the window relative to the Forecast Point, it defines the number of recent values the model can use for forecasting.

Feature Discovery

A Arkangel AI capability that discovers and generates new features from multiple datasets, eliminating the need to perform manual feature engineering to consolidate multiple datasets into one. A relationship editor visualizes these relationships and the end product is additional, derived features that result from the created linkages.

Feature Effects

A model Leaderboard tab (Understand > Feature Effects) that shows the effect of changes in the value of each feature on the model’s predictions. It displays a graph depicting how a model "understands" the relationship between each feature and the target, with the features sorted by Feature Impact.

Feature engineering

The generation of additional features in a dataset, which as a result, improve model accuracy and performance. Time series and Feature Discovery both rely on feature engineering as the basis of their functionality.

Feature extraction

(Visual AI) Models that perform image preprocessing (or image feature extraction and image preprocessing) are also known as “image feature extraction models” or “image-specific models.”

Feature Extraction and Reduction (FEAR)

(Time series) Feature generation for time series (e.g., lags, moving averages). It extracts new features (now) and then reduces the set of extracted features (later). See time series feature derivation.

Feature Impact

A measurement that identifies which features in a dataset have the greatest effect on model decisions. In Arkangel AI, the measurement is reported as a visualization available from the Leaderboard.

Feature imputation

(Time series) A mechanism that uses forward filling to enable imputation for all features (target and others) when using the time series data prep tool. This results in a dataset with no missing values (with the possible exception of leading values at the start of each series where there is no value to forward fill).

Feature list

A subset of features from a dataset used to build models. Arkangel AI creates several lists during EDA2 including all informative features, informative features excluding those with a leakage risk, a raw list of all original features, and a reduced list. Uses can create project-specific lists as well.


(Data Prep) A Data Prep column tool that is both a filter to help you transform your data and a histogram to help you visualize your data. Filtergrams allow you to visualize your data before, during, and after every transformation.


See model fitting.

Forecast Distance

(Time-aware) A unique time step—a relative position—within the Forecast Window. A model outputs one row for each Forecast Distance.

Forecast Point

(Time-aware) The point you are making a prediction from; a relative time "if it was now..."; Arkangel AI trains models using all potential forecast points in the training data. In production, it is typically the most recent time.

Forecast vs Actual

A model Leaderboard tab (Evaluate > Forecast vs Actual) commonly used in time series projects that allows you to compare how different predictions behave from different forecast points to different times in the future. Although similar to the Accuracy Over Time chart, which displays a single forecast at a time, the Forecast vs Actual chart shows multiple forecast distances in one view.

Forecast Window

(Time-aware) Also known as FW. Beginning from the Forecast Point, defines the range (the Forecast Distance) of future predictions—"this is the range of time I care about." Arkangel AI then optimizes models for that range and ranks them on the Leaderboard on the average across that range.


(Time-aware) Predictions based on time, into the future; use inputs from recent rows to predict future values. Forecasting is a subset of predictions, using trends in observation to characterize expected outcomes or expected responses.

Frozen run

A process that “freezes” parameter settings from a model’s early, small sample size-based run. Because parameter settings based on smaller samples tend to also perform well on larger samples of the same data.


See Forecast Window.


Governance lens

(MLOps) A filtered view of Arkangel AI's deployment inventory on the Deployments page, summarizing the social and operational aspects of a deployment. These include the deployment owner, how the model was built, the model's age, and the humility monitoring status.

GPU (graphics processing unit)

A mechanism for processing computational tasks. GPUs are GPUs are highly optimized to do mathematical calculations and great at parallelism, but only for less complex tasks. Deep learning specifically benefits from that since it's mainly batches of matrix multiplication, and these can be parallelized very easily.

Grid search

An exhaustive search method used for hyperparameters.



A subset of data that is unavailable to models during the training and validation process. Use the Holdout score for a final estimate of model performance only after you have selected your best model.


(MLOps) A user-defined set of rules for deployments that allow models to be capable of recognizing, in real-time, when they make uncertain predictions or receive data they have not seen before. Unlike data drift, model humility does not deal with broad statistical properties over time—it is instead triggered for individual predictions, allowing you to set desired behaviors with rules that depend on different triggers.


Image data

(Visual AI) A sequence of digital images (e.g., video), a set of digital images, a single digital image, and/or one or more portions of any of the foregoing. A digital image may include an organized set of picture elements (“pixels”) stored in a file. Any suitable format and type of digital image file may be used, including but not limited to raster formats (e.g., TIFF, JPEG, GIF, PNG, BMP, etc.), vector formats (e.g., CGM, SVG, etc.), compound formats (e.g., EPS, PDF, PostScript, etc.), and/or stereo formats (e.g., MPO, PNS, JPS).

Image preprocessing

(Visual AI) A computer vision technique. Some examples include image re-sampling, noise reduction, contrast enhancement, and scaling (e.g., generating a scale space representation). Extracted features may be:

  • Low-level: raw pixels, pixel intensities, pixel colors, gradients, textures, color histograms, motion vectors, edges, lines, corners, ridges, etc.

  • Mid-level: shapes, surfaces, volumes, etc.

  • High-level: objects, scenes, events, etc.

Inference data

(Predictions) Data that is scored by applying an algorithmic model built from a historical dataset in order to uncover practical insights. See also Scoring data.

In-sample predictions

(Predictions) Models trained on data outside of the training set (i.e., Validation and potentially Holdout). Arkangel AI uses 64% of the training set by default. When models are trained with a sample size above 64%, Arkangel AI marks the Validation score with an asterisk to indicate that some in-sample predictions were used for that score. If you train above 80%, the Holdout score is also asterisked. Compare to stacked (out-of-sample) predictions.

Irregular data

(Time-aware) Data in which no consistent spacing and no time step is detected.



See Known in advance features.

Known in advance features

(Time-aware) Also known as KA. A variable for which you know the value in advance and does not need to be lagged, such as holiday dates. Or, for example, you might know that a product will be on sale next week and so you can provide the pricing information in advance.



The list of trained blueprints (models) for a project, ranked according to a project metric.


See target leakage.

Learning Curves

A graph to help determine whether it is worthwhile to increase the size of a dataset. The Learning Curve graph illustrates, for the top-performing models, how model performance varies as the sample size changes.


(Data Prep) A Data Prep construct that lets you generate a snapshot of your dataset at a particular step in a project. You create a lens to identify the project steps to be published to the AnswerSet.

Lift Chart

Depicts how well a model segments the target population and how capable it is of predicting the target to help visualize model effectiveness.

Linkage keys

(Feature Discovery) The features in the primary dataset used as keys to join and create relationships.

Location AI

Arkangel AI's support for geospatial analysis by natively ingesting common geospatial formats and recognizing coordinates, allowing ESDA, and providing spatially-explicit modeling tasks and visualizations.


A model Leaderboard tab (Describe > Log) that displays the status of successful operations with green INFO tags, along with information about errors marked with red ERROR tags.


Machine Learning Operations

See MLOps.

Majority class

If you have a categorical variable (e.g., true/false or cat/mouse ), the value that's more frequent is the majority class. For example, if a dataset has 80 rows of value cat and 20 rows of value mouse, then cat is the majority class.

Make Predictions tab

A model Leaderboard tab (Predict > Make Predictions) that allows you to make predictions before deploying a model to a production environment.

Management agent

(MLOps) A downloadable client included in the MLOps agent tarball (accessed via Developer Tools) that allows you to manage external models (i.e., those running outside of Arkangel AI MLOps). This tool provides a standard mechanism to automate model deployment to any type of infrastructure. The management agent sends periodic updates about deployment health and status via the API and reports them as MLOps events on the Service Health page.


A modeling mode that causes Arkangel AI to complete EDA2 and prepare data for modeling, but does not execute model building. Instead, users select specific models to build from the model Repository.


(AI Catalog) Materialized data is data that Arkangel AI has pulled from the data asset and is currently keeping a copy of in the catalog. See snapshot.


(AI Catalog) Details of the data asset, such as creation and modification dates, number and types of features, snapshot status, and more.


See optimization metric.

Minority class

If you have a categorical variable (e.g., true/false or cat/mouse ), the value that's less frequent is the minority class. For example, if a dataset has 80 rows of value cat and 20 rows of value mouse, then mouse is the minority class.

MLOps (Machine Learning Operations)

(MLOps) A scalable and governed means to rapidly deploy and manage ML applications in production environments.

MLOps agent

(MLOps) One of two downloadable clients included in the MLOps agent tarball (accessed via Developer Tools) that allows you to monitor and manage external models (i.e., those running outside of Arkangel AI MLOps). See monitoring agent and management agent.


A trained ML pipeline, capable of scoring new data. Models—descriptive, predictive, prescriptive—form the basis of data analysis. Modeling extracts insights from data that you can then use to make better business decisions. Algorithmic models tell you which outcome is likely to hold true for your target variable based on your training data. They construct a representation of the relationships and tease out patterns between all the different features in your dataset that you can apply to similar data you collect in the future, allowing you to make decisions based on those patterns and relationships.

Model Comparison

A Leaderboard tab that allows you to compare two models using different evaluation tools, helping identify the model that offers the highest business returns or candidates for blender models.

Model fitting

A measure of how well a model generalizes similar data to the data on which it was trained. A model that is well-fitted produces more accurate outcomes. A model that is overfitted matches the data too closely. A model that is underfitted doesn’t match closely enough.

Model Info

A model Leaderboard tab (Describe > Model Info) that displays an overview for a given model, including model file size, prediction time, and sample size.

Model package

(MLOps) Archived model artifacts with associated metadata stored in the Model Registry. Model packages can be created manually or automatically, for example, through the deployment of a custom model. You can deploy, share, and permanently archive model packages.

Model Registry

(MLOps) An organizational hub for the variety of models used in Arkangel AI. Models are registered as deployment-ready model packages; the registry lists each package available for use. Each package functions the same way, regardless of the origin of its model. The Model Registry also contains the Custom Model Workshop where you can create and deploy custom models. Model packages can be created manually or automatically depending on the type of model.

Model scoring

The process of applying an optimization metric to a partition of the data and assigning a numeric score that can be used to evaluate a model performance.

Modeling dataset

(Time-aware) A transform of the original dataset that pre-shifts data to future values, generates lagged time series features, and computes time-series analysis metadata. Commonly referred to as feature derivation, it is used by time series but not OTV. See the time series feature engineering reference for a list of operators used and feature names created by the feature derivation process.

Modeling mode

A setting that controls the sample percentages of the training set that Arkangel AI uses to build models. Arkangel AI offers four modeling modes: Autopilot, Quick (the default), Manual, and Comprehensive.

Monitoring agent

(MLOps) A downloadable client included in the MLOps agent tarball (accessed via Developer Tools) that allows you to monitor external models (i.e., those running outside of Arkangel AI MLOps). With this functionality, predictions and information from these models can be reported as part of deployments. You can use this tool to monitor accuracy, data drift, prediction distribution, latency, and more, regardless of where the model is running.

Monotonic modeling

A method to force certain XGBoost models to learn only monotonic (always increasing or always decreasing) relationships between specific features and the target.


See classification.


A classification task where each row in a dataset is associated with one, several, or zero labels. Common multilabel classification problems are text categorization (a movie is both "crime" and "drama") and image categorization (an image shows a house and a car).


A model type that supports multiple var types at the same time, in the same model.


(Time-aware) Datasets that contain multiple time series (for example, to forecast the sales of multiple stores) based on a common set of input features.


Naive model

See baseline model.

No-Code AI Apps

A no-code interface to create AI-powered applications that enable core Arkangel AI services without having to build models and evaluate their performance. Applications are easily shared and do not require consumers to own full Arkangel AI licenses in order to use them.


A sequence of words, where N is the number of words. For example, "machine learning" is a 2-gram. Text features are divided into n-grams to prepare for Natural Language Processing (NLP).


An interactive, computational environment that hosts code execution and rich media. Arkangel AI provides its own in-app environment to create, manage, and execute Jupyter-compatible hosted notebooks.


(Time-aware) A method of time series modeling that predicts the current value of a target based on past and present data. Technically, it is a forecast window in which the start and end times are 0 (now).



Feature(s) that should be treated as a fixed component for modeling (coefficient of 1 in generalized linear models or gradient boosting machine models). Offsets are often used to incorporate pricing constraints or to boost existing models.

Optimization metric

An error metric used in Arkangel AI to determine how well a model predicts actual values. After you choose a target feature, Arkangel AI selects an optimization metric based on the modeling task.


(Time-aware) Also known as out-of-time validation. A method for modeling time-relevant data. With OTV you are not forecasting, as with time series. Instead, you are predicting the target value on each individual row.


A situation in which a model fits its training data too well and therefore loses its ability to perform accurately against unseen data. This happens when a model trains too long on the training data and learns (and models on) its "noise," making the model unable to generalize.



Segments of training data, broken down to maximize accuracy. The segments (splits) of the dataset. See also training, validation, cross-validation, and holdout.

Per-Class Bias

A model Leadboard tab (Bias and Fairness > Per-Class Bias) that helps to identify if a model is biased, and if so, how much and who it's biased towards or against. Bias and Fairness settings must be configured.

PID (project identifier)

An internal identifier used for uniquely identifying a project.


Personal identifiable information, including name, pictures, home address, SSN or other identifying numbers, birthdate, and more. Arkangel AI automates the detection of specific types of personal data to provide a layer of protection against the inadvertent inclusion of this information in a dataset.

Portable prediction server (PPS)

(MLOps) A Arkangel AI execution environment for Arkangel AI model packages (.mlpkg files) distributed as a self-contained Docker image. It can be run disconnected from main installation environments.


(Time-aware) For non-time-series modeling. Use information in a row to determine the target for that row. Prediction uses explanatory variables to characterize expected outcomes or expected responses (e.g., a specific event in the future, gender, fraudulent transactions).

Prediction data

(MLOps, Predictions) Data that contains prediction requests and results from the model.

Prediction environment

(MLOps, Predictions) An environment configured to manage deployment predictions on an external system, outside of Arkangel AI. Prediction environments allow you to configure deployment permissions and approval processes. Once configured, you can specify a prediction environment for use by Arkangel AI models running on the Portable Prediction Server and for remote models monitored by the MLOps monitoring agent.

Prediction Explanations

A visualization that helps to illustrate what drives predictions on a row-by-row basis—they provide a quantitative indicator of the effect variables have on a model, answering why a given model made a certain prediction. It helps to understand why a model made a particular prediction so that you can then validate whether the prediction makes sense. See also SHAP, XEMP.

Prediction intervals

Prediction intervals help Arkangel AI assess and describe the uncertainty in a single record prediction by including an upper and lower bound on a point estimate (e.g., a single prediction from a machine learning model). The prediction intervals provide a probable range of values that the target may fall into on future data points.

Prediction point

The point in time when you made or will make a prediction. Plan your prediction point based on the production model (for example, “one month before renewal” or “loan application submission time”). Once defined, create that entry in the training data to help avoid lookahead bias. With Feature Discovery, you define the prediction point to ensure the derived features only use data prior to that point.

Primary dataset

(Feature Discovery) The dataset used to start a project.

Primary features

(Feature Discovery) Features in the project’s primary dataset.


A referenceable item that includes a dataset, which is the source used for training, and any models built from the dataset. Projects can be created and accessed from the home page, the project control center, and the AI Catalog. They can be shared to users, groups, and an organization.

Protected class

(Bias and Fairness) One categorical value of the protected feature.

Protected feature

(Bias and Fairness) The dataset column to measure fairness of model predictions against. Model fairness is calculated against the protected features from the dataset. Also known as “protected attribute.”


Quick (Autopilot)

A shortened version of the full Autopilot modeling mode that runs models starting at 32% and then advances to 64%. With Quick, the 16% sample size is not executed.


Rating Table

A model Leaderboard tab (Describe > Rating Table) where you can export the model's complete, validated parameters.

Real-time predictions

(Predictions) Method of making predictions when low latency is required. Use the Prediction API for real-time deployment predictions on a dedicated and/or a standalone prediction server.

Receiver Operating Characteristic Curve

See ROC Curve.


A type of prediction problem that predicts continuous values (for example, 1.7, 6, 9.8…).

Regular data

(Time-aware) Data is regular if rows in the dataset fall on an evenly spaced time grid (e.g., there’s one row for every hour across the entire dataset).


(Feature Discovery) Relationships between datasets. Each relationship involves a pair of datasets, and a join key from each dataset. A key comprises one or more columns of a dataset. The keys from both datasets are ordered, and must have the same number of columns. The combination of keys is used to determine how two datasets are joined.

Remote models

(MLOps) Models running outside of Arkangel AI in external prediction environments, often monitored by the MLOps monitoring agent to report statistics back to Arkangel AI.


A library of modeling blueprints available for a selected project (based on the problem type). These models may be selected and built by Arkangel AI and also can be user-executed.

ROC Curve

Also known as Receiver Operating Characteristic Curve. A visualization that helps to explore classification, performance, and statistics related to a selected model at any point on the probability scale. In Arkangel AI, the visualization is available from the Leaderboard.


(AI Catalog) Roles—Owner, Consumer, and Editor—describe the capabilities provided to each user for a given dataset. This supports the scenarios when the user creating a data source or data connection and the enduser are not the same, or there are multiple endusers of the asset.


Sample size

The percentage of the total training data used to build models. The percentage is based on the selected modeling mode or can be user-selected.


See Model scoring, Scoring data.

Scoring Code

(MLOps, Predictions) A method for using Arkangel AI models outside of the application. It is available for select models from the Leaderboard as a downloadable JAR file containing Java code that can be used to score data from the command line.

An exportable JAR file, available for select models, that runs in Java. Scoring Code JARs contain prediction calculation logic identical to the Arkangel AI API—the code generation mechanism tests each model for accuracy as a part of the generation process.

Scoring data

(Predictions) Applying an algorithmic model built from a historical dataset to a new dataset in order to uncover practical insights. Common scoring methods are batch and real-time scoring. "Scored data" (also called "inference data") refers to the dataset being scored.


(Time-aware) Repeating highs and lows observed at different times of year, within a week, day, etc. Periodicity. For example, temperature is very seasonal (hot in the summer, cold in the winter, hot during the day, cold at night).

Secondary dataset

(Feature Discovery) A dataset that is added to a project and part of a relationship with the primary dataset.

Secondary features

(Feature Discovery) Features derived from a project’s secondary datasets.

Segmented analysis

(MLOps) A deployment utility that filters data drift and accuracy statistics into unique segment attributes and values. Useful for identifying operational issues with training and prediction request data.

Segmented modeling

(Time-aware) A method of modeling multiseries projects by generating a model for each segment. Arkangel AI selects the best model for each segment (the segment champion) and includes the segment champions in a single Combined Model that you can deploy.

Semi-regular data

(Time-aware) Data is semi-regular if most time steps are regular but there are some small gaps (e.g., business days, but no weekends).

Segment ID

(Time-aware) A column in a dataset used to group series into segments for a multiseries project. A segment ID is required for the segmented modeling workflow, where Arkangel AI builds a separate model for each segment. See also Segmented modeling.

Series ID

(Time-aware) A column in a dataset used to divide a dataset into series for a multiseries project. The column contains labels indicating which series each row belongs to. See also Multiseries modeling.

Service health

(MLOps) A performance monitoring component for deployments that tracks metrics about a deployment’s ability to respond to prediction requests quickly and reliably. Useful for identifying bottlenecks and assessing prediction capacity.

SHAP (Shapley Values)

A fast, open-source methodology for computing Prediction Explanations for tree-based, deep learning, and linear-based models. SHAP estimates how much each feature contributes to a given prediction differing from the average. It is additive, making it easy to see how much top-N features contribute to a prediction. See also Prediction Explanations, XEMP.

Smart downsampling

A technique to reduce total dataset size by reducing the size of the majority class, enabling you to build models faster without sacrificing accuracy. When enabled, all analysis and model building is based on the new dataset size after smart downsampling.


(AI Catalog) A snapshot is an asset created from a data source. For example, with a database it represents either the entire database or a selection of (potentially joined) tables, taken at a particular point in time. It is taken from a live database but creates a static, read-only copy of data. Arkangel AI creates a snapshot of each data asset type, while allowing you to disable the snapshot when importing the data.

Speed vs Accuracy

A Leaderboard tab that generates an analysis plot to show the tradeoff between runtime and predictive accuracy and help you choose the best model with the lowest overhead.


A model Leaderboard tab (Evaluate > Stability) that provides an at-a-glance summary of how well a model performs on different backtests. The backtesting information in this chart is the same as that available from the Model Info tab.

Stacked predictions

(Predictions) A method for building multiple models on different subsets of the data. The prediction for any row is made using a model that excluded that data from training. In this way, each prediction is effectively an “out-of-sample” prediction. See an example in the predictions documentation. Compare to "in-sample" predictions.


(Time-aware) The mean of the series does not change over time. A stationary series does not have a trend or seasonal variation.

Supervised learning

Machine learning using labeled data, meaning that for each record, the dataset contains a known value for the target feature. By knowing the target during training, the model can "learn" how other features relate to the target and make predictions on new data.



The name of the column in the dataset that you would like to predict.

Target leakage

An outcome when using a feature whose value cannot be known at the time of prediction (for example, using the value for “churn reason” from the training dataset to predict whether a customer will churn). Including the feature in the model’s feature list would incorrectly influence the prediction and can lead to overly optimistic models.


An ML method, for example a data transformation such as one-hot encoding, or an estimation such as an XGBoost classifier, which is used to define a blueprint. There are hundreds of built-in tasks you can use, or you can define your own (custom) tasks.

Time series

(Time-aware) A series of data points indexed in time order. Ordinarily a sequence of measurements taken at successive, equally spaced intervals.

Time series analysis

(Time-aware) Methods for analyzing time series data in order to extract meaningful statistics and other characteristics of the data.

Time series forecasting

(Time-aware) The use of a model to predict future values based on previously observed values. In practice, a forecasting model may combine time series features with other data.

Time step

(Time-aware) The detected median time delta between rows in the time series; Arkangel AI determines the time unit. The time step consists of a number and a time-delta unit, for example (15, “minutes”). If a step isn’t detected, the dataset is considered irregular and time series mode may be disabled.

Tracking agent

See MLOps agent.


The process of building models on data in which the target is known.

Training Dashboard

A model Leaderboard tab (Evaluate > Training dashboard) that provides, for each executed iteration, information about a model's training and test loss, accuracy, learning rate, and momentum to help you get a better understanding about what may have happened during model training.

Training data

The portion (partition) of data used to build models. See also validation, cross-validation, and holdout.

Transfer learning

(Visual AI) A project training on one dataset, extracting information that may be useful, and applying that learning to another.


(Time-aware) An increase or decrease over time. Trends can be linear or non-linear and can show fluctuation. A series with a trend is not stationary.


A trial-and-error process by which you change some hyperparameters, run the algorithm on the data again, then compare performance to determine which set of hyperparameters results in the most accurate model. In Arkangel AI, this functionality is available from the Advanced Tuning tab.


Unit of analysis

(Machine learning) The unit of observation at which you are making a prediction.

Unlimited multiclass

See classification.


(AI Catalog) Unmaterialized data is data that Arkangel AI samples for profile statistics, but does not keep. Instead, the catalog stores a pointer to the data and only pulls it upon user request at project start or when running batch predictions.

Unsupervised learning

The ability to infer patterns from a dataset without reference to known (labeled) outcomes and without a specified target. Types of unsupervised learning include anomaly detection, outlier detection, novelty detection, and clustering. With anomaly detection, Arkangel AI applies unsupervised learning to detect abnormalities in a dataset. With clustering, Arkangel AI uses unsupervised learning to discern natural groupings in the data.



The validation (or testing) partition is a subsection of data that is withheld from training and used to evaluate a model’s performance. Since this data was not used to build the model, it can provide an unbiased estimate of a model’s accuracy. You often compare the results of validation when selecting a model. See also cross-validation.


See feature.

Visual AI

Arkangel AI's ability to combine supported image types, either alone or in combination with other supported feature types, to create models that use images as input. The feature also includes specialized insights (e.g., image embeddings, activation maps, neural network visualizer) to help visually assess model performance.


Word Cloud

A model Leaderboard tab (Understand > Word Cloud) that displays the most relevant words and short phrases in word cloud format.


The processing power behind the Arkangel AI platform, used for creating projects, training models, and making predictions. They represent the portion of processing power allocated to a task. Arkangel AI uses different types of workers for different phases of the project workflow, including DSS workers (Dataset Service workers), EDA workers, secure modeling workers, and quick workers.


XEMP (eXemplar-based Explanations of Model Predictions)

A methodology for computing Prediction Explanations that works for all models. See also Prediction Explanations, SHAP.


Z Score

(Bias and Fairness) A metric measuring whether a given class of the protected feature is “statistically significant” across the population.

Last updated