Glossary of Terms

This glossary defines key terms used throughout Atlas documentation.

A

AbstractModel

Base class that all models must inherit from. Defines the interface for model integration including predict() and contributions() methods.

Acquisition Function

In Bayesian optimization, a function that determines which point to evaluate next based on the posterior distribution. Common types include Expected Improvement (EI) and Upper Confidence Bound (UCB).

Adstock

See Carryover Effect.

Allocation

The distribution of budget or resources across different channels, time periods, or categories.

API (Application Programming Interface)

A set of protocols and tools for building software applications. The framework provides REST APIs for model serving and optimization.

B

Baseline

The reference point for comparison, typically the current or historical budget allocation used to measure improvement.

Bayesian Optimization

An optimization technique that builds a probabilistic model of the objective function and uses it to select the most promising points to evaluate.

Black-box Model

A model where the internal workings are unknown or inaccessible, only input-output relationships are observable.

Budget

The total amount of resources (usually monetary) available for allocation across different channels or categories.

Business Constraint

Limitations imposed by business rules, policies, or practical considerations (e.g., minimum spend requirements, market share limits).

C

Carryover Effect

The phenomenon where marketing impact extends beyond the initial period of spend, also known as adstock.

Channel

A distinct marketing or distribution method (e.g., TV, digital, radio) that can receive budget allocation.

Constraint

A limitation or restriction on the optimization problem, such as budget limits, bounds, or business rules.

Contribution Analysis

The process of determining how much each input variable contributes to the model’s output prediction.

Convergence

The state when an optimization algorithm has found a solution that satisfies stopping criteria and further iterations don’t significantly improve the result.

Convex Optimization

A class of optimization problems where the objective function is convex and the feasible region is a convex set, guaranteeing a global optimum.

CVXPY

A Python library for convex optimization that allows natural mathematical notation for problem specification.

D

Data Pipeline

The series of data processing steps from raw input to model-ready format.

Diminishing Returns

The principle that each additional unit of input produces progressively smaller increases in output.

Docker

A platform for developing, shipping, and running applications in containers, used for model isolation and deployment.

Drift

The phenomenon where model performance degrades over time due to changes in the underlying data distribution.

E

Efficient Frontier

See Pareto Frontier.

Elasticity

The percentage change in output (e.g., sales) resulting from a percentage change in input (e.g., marketing spend).

Ensemble Model

A model that combines predictions from multiple individual models to improve overall performance.

Epoch

One complete pass through the entire dataset during model training.

F

Feasible Region

The set of all possible solutions that satisfy all constraints.

Feature

An individual measurable property or characteristic used as input to a model.

Frequency Capping

Limiting the number of times an individual is exposed to an advertisement within a specific time period.

G

Global Optimum

The best possible solution across the entire search space, as opposed to a local optimum.

Gradient

The vector of partial derivatives of a function, indicating the direction of steepest increase.

Grid Search

An exhaustive search method that evaluates all combinations of a discrete set of parameter values.

H

Heuristic

A practical problem-solving approach that uses shortcuts to produce good-enough solutions quickly.

Hill Transformation

A mathematical function used to model saturation effects in marketing response curves.

Hyperparameter

A parameter whose value is set before the learning process begins, controlling the learning process itself.

I

Incrementality

The additional impact generated by a marketing activity beyond what would have occurred naturally.

Integer Programming

Optimization where some or all variables are restricted to integer values.

Iteration

A single step in an optimization algorithm where the solution is updated.

J

JSON (JavaScript Object Notation)

A lightweight data interchange format used for configuration and API communication.

K

KPI (Key Performance Indicator)

A measurable value that demonstrates how effectively objectives are being achieved (e.g., revenue, conversions, awareness).

L

Lever

An optimization variable that can be adjusted to influence outcomes (e.g., budget allocation to a channel).

Linear Programming

A method for optimizing a linear objective function subject to linear constraints.

Local Optimum

A solution that is optimal within a neighboring set of solutions but may not be the global optimum.

Loss Function

A function that measures the difference between predicted and actual values, which optimization seeks to minimize.

M

Marginal ROI

The return on investment for the next incremental dollar spent.

Marketing Mix Modeling (MMM)

Statistical analysis technique that measures the impact of marketing activities on sales.

Metadata

Data that provides information about other data, such as model version, training date, or performance metrics.

Multi-objective Optimization

Optimization involving multiple, often conflicting objectives that must be balanced.

N

Non-convex Optimization

Optimization problems where the objective function or constraints are non-convex, potentially having multiple local optima.

Normalization

The process of scaling data to a standard range, typically [0, 1] or [-1, 1].

O

Objective Function

The function being optimized (maximized or minimized) in an optimization problem.

Optuna

A hyperparameter optimization framework that uses advanced sampling algorithms for black-box optimization.

Overfitting

When a model learns the training data too well, including noise, reducing its ability to generalize.

P

Pareto Frontier

The set of solutions where no objective can be improved without degrading another objective.

Pareto Optimal

A solution is Pareto optimal if no other solution improves at least one objective without worsening another.

Pipeline

A sequence of data processing steps, from raw input to final output.

Prediction Interval

A range of values within which future observations are expected to fall with a certain probability.

Q

Quadratic Programming

Optimization of a quadratic objective function subject to linear constraints.

R

Reach

The number of unique individuals exposed to a marketing message.

Regularization

Technique to prevent overfitting by adding a penalty term to the loss function.

Response Curve

The relationship between marketing input (e.g., spend) and output (e.g., sales).

ROI (Return on Investment)

The ratio of net profit to the cost of investment.

ROAS (Return on Ad Spend)

Total revenue generated per dollar spent on advertising.

S

Saturation

The point at which additional marketing spend yields diminishing or no returns.

Scenario Analysis

Evaluating different possible future situations and their potential impacts.

SciPy

Scientific computing library for Python that includes optimization algorithms.

Sensitivity Analysis

The study of how changes in input variables affect the output.

Simplex Algorithm

An algorithm for solving linear programming problems.

Stochastic Optimization

Optimization methods that incorporate randomness, useful for problems with uncertainty.

Surrogate Model

A simplified approximation of a complex model used to speed up optimization.

T

Time Series

A sequence of data points indexed in time order.

TPE (Tree-structured Parzen Estimator)

A Bayesian optimization algorithm used in Optuna for hyperparameter tuning.

Trade-off

A situation where improving one objective requires degrading another.

U

Uncertainty Quantification

The process of quantifying and characterizing uncertainty in model predictions.

Underfitting

When a model is too simple to capture the underlying patterns in the data.

V

Validation

The process of assessing whether a model’s predictions are accurate on unseen data.

Variable Bounds

Minimum and maximum allowed values for optimization variables.

Vectorization

The process of converting operations to work on entire arrays rather than individual elements.

W

Warm Start

Initializing an optimization algorithm with a good starting point based on prior knowledge.

Weight

The relative importance assigned to different objectives in multi-objective optimization.

What-if Analysis

Exploring how changes in inputs affect outputs without actually implementing the changes.

X

Xarray

Python library for working with labeled multi-dimensional arrays, used for data management in the framework.

Y

Yield

The return or output generated from a given input or investment.

Z

Zero-sum

A situation where total gains equal total losses, common in competitive market share scenarios.


Technical Acronyms

API: Application Programming Interface
CI/CD: Continuous Integration/Continuous Deployment
CPU: Central Processing Unit
CSV: Comma-Separated Values
GPU: Graphics Processing Unit
HTTP: Hypertext Transfer Protocol
JSON: JavaScript Object Notation
KPI: Key Performance Indicator
ML: Machine Learning
MMM: Marketing Mix Modeling
REST: Representational State Transfer
ROI: Return on Investment
ROAS: Return on Ad Spend
SDK: Software Development Kit
URL: Uniform Resource Locator
UUID: Universally Unique Identifier
YAML: Yet Another Markup Language


Mathematical Notation

: Gradient (nabla)
: Partial derivative
Σ: Summation
: Product
: Element of
: Subset of
: For all
: There exists
arg max: Argument of the maximum
arg min: Argument of the minimum
s.t.: Subject to (in optimization)
w.r.t.: With respect to


This glossary is a living document. If you encounter terms that should be added or clarified, please contribute to the documentation.