liveBookmanning.com

What you will learn in Delivering Machine Learning Projects: From design to deployment

From design to deployment

by Simon Thompson

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• nonfunctional requirementA nonfunctional property describes an operational characteristic of data, such as its scale, speed, history, or update frequency.
• data surveyA data survey is a structured examination of available data resources, their characteristics, and their suitability for a project.
• customer churn prediction systemA customer churn prediction system is a machine learning system that estimates which customers are likely to end their relationship with a business.
• demand prediction systemA demand prediction system is a machine learning system that estimates future customer or product demand from historical and contextual data.
• dashboardA dashboard system is software that presents operational data and events through visual interfaces for users.
• pipelineA production data flow is the operational movement and transformation of data through a deployed machine learning system.
• model selectionModel selection is the process of choosing among candidate machine learning models based on evaluation results and project requirements.
• data engineerA data engineer is a specialist who acquires, moves, transforms, manages, and operationalizes data resources and their supporting infrastructure.
• functional requirementA functional requirement specifies a capability or behavior that a system must provide.
• data assetA data asset is a structured collection of data maintained for analysis, model development, or operation of a system.
• data infrastructureData infrastructure is the hardware, software, storage, networking, and operational environment that supports data access and processing.
• Model ServingA model-serving system is the software and infrastructure that loads trained machine learning models and handles prediction requests in an operating environment.
• data protection impact assessmentA data protection impact assessment is a structured assessment of how planned personal-data processing may affect individuals and the measures needed to address those impacts.
• data testingData testing is the automated checking of data values, distributions, relationships, and construction rules against expected conditions.
• test harnessA test harness is software and supporting components that drive tests, provide controlled inputs, and collect outputs from the artifact under test.
• machine learning projectA machine learning project is an organized effort to design, build, evaluate, deploy, and maintain a system containing machine-learned components.
• data storyA data story is a documented account of a data resource's origins, contents, transformations, quality issues, and intended uses.
• ethicsEthical considerations are judgments about whether a system's purposes, effects, and operation respect human welfare, rights, and societal values.
• model architectureA model architecture is the structural arrangement of components and connections that defines how a machine learning model processes inputs and produces outputs.
• project hypothesisA project hypothesis is a provisional statement of the solution and value proposition that a project will investigate.
• machine learning engineerA machine learning engineer is a specialist who develops, integrates, tests, and operates software systems containing machine learning models.
• risk registerA risk register is a maintained record of identified risks, their characteristics, and actions for addressing them.
• documentationDocumentation is recorded information describing a project's work, decisions, requirements, and system state.
• monitoringMonitoring is the systematic inspection and recording of a system's behavior to detect issues and notify responsible parties.
• requirements analysisRequirements analysis is the systematic process of eliciting, clarifying, assessing, and specifying what a system must provide and under what constraints.
• precisionPrecision is the proportion of instances predicted as positive that are actually positive.
• recallRecall is the proportion of relevant positive instances that a predictive classification system correctly identifies.
• machine learningMachine learning is the computational approach of learning models from data through algorithms rather than specifying all model rules manually.
• assistive systemAn assistive application is software that supports users' decisions or actions without fully replacing their judgment or control.
• model developmentMachine learning modeling is the process of selecting, training, debugging, tuning, and refining machine learning models from data.
• foundational modelA foundation model is a broadly trained machine learning model that can be adapted to support multiple downstream tasks.
• business intelligenceBusiness intelligence is the analysis and presentation of business data to support organizational understanding and decision-making.
• customer churnCustomer churn is the ending or loss of a customer's relationship with a business.
• retentionCustomer retention is the continuation of a customer's relationship with a business over time.
• data wranglingData wrangling is the process of restructuring, cleaning, and transforming data into a form suitable for analysis or modeling.
• RFIAn RFI is a request issued to gather information about potential solutions, capabilities, or approaches before a formal procurement decision.
• data explorationExploratory data analysis is the examination of data to characterize its structure, quality, patterns, and anomalies before or during modeling.
• team designTeam design is the structuring of a project team’s roles, responsibilities, and composition to support delivery.
• logLogging is the recording of software or system events, states, and messages for operational analysis and troubleshooting.
• migrationData migration is the transfer of data between storage systems, formats, schemas, or operational environments.
• machine learning systemA machine learning system is a deployed software system whose behavior depends on one or more machine learning models and their surrounding data and infrastructure.
• evaluationPerformance evaluation is the systematic assessment of a model's predictive quality against defined data, metrics, and application requirements.
• reproducibilityReproducibility is the ability to recreate a computational result from the same specified inputs, components, and procedures.
• requirementsA requirement is a stated need, capability, quality attribute, constraint, or operational condition that a system or project must address.
• securityData security is the protection of data from unauthorized access, disclosure, alteration, loss, or destruction.
• agile software developmentAgile development is an iterative software development approach emphasizing early delivery, customer collaboration, and responsiveness to change.
• execution environmentA development environment is a computing environment configured for building and integrating software before formal testing or production deployment.
• unstructured dataAn unstructured field is a data field containing content such as images, text, or sounds without a fixed tabular representation.
• model repositoryA model repository is a managed storage system for machine learning models and their associated versions or metadata.
• impact assessmentAn impact assessment is a structured evaluation of a system's intended and unintended effects, costs, benefits, and risks on affected parties.
• transparencyTransparency is the degree to which a model's reasoning and outputs can be understood and inspected by people.
• business translatorA business translator is a project role that bridges domain stakeholders and machine-learning specialists by interpreting business needs and technical implications.
• cloud engineerA cloud engineer specializes in configuring and managing cloud infrastructure and cloud-managed services.
• data scientistA data scientist is a specialist who conditions data, creates models, and tests and evaluates their effectiveness.
• DevOps engineerA DevOps engineer specializes in setting up and administering software development infrastructure.
• product ownerA product owner is a customer representative who provides feedback and priorities so a project responds to customer needs.
• descriptive statisticsA summary statistic is a numerical measure that concisely describes an important characteristic of a data set.
• training data biasData bias is systematic distortion in data that reflects unequal, incomplete, or agenda-driven representation of the phenomena or people it describes.
• autoencoderAn autoencoder is a neural network that learns to encode inputs into a lower-dimensional representation and reconstruct the original inputs.
• baseline modelA benchmark model is a baseline model whose results provide a reference for comparing alternative models.
• confidence intervalA confidence interval is an interval estimate expressing uncertainty about an unknown quantity at a specified confidence level.
• counterfactual outcomeA counterfactual scenario is a hypothetical situation describing how outcomes might change under an altered condition or input.
• reconstruction lossA reconstruction error is the discrepancy between an input and the output reconstructed from its encoded representation.
• governanceSystem governance is the set of decision rights, controls, responsibilities, and processes that direct and oversee a system throughout its lifecycle.
• data sourceA source system is an operational system that originates or stores data later used by another system or process.
• data leakageData leakage is the unintended use or transfer of information from evaluation data or future observations into model training or selection, causing misleadingly optimistic performance estimates.
• accountabilityAccountability is the condition in which responsible parties can explain, justify, and demonstrate compliance for a system's decisions and effects.
• autonomous systemAn autonomous system is a system that operates independently for extended periods without human intervention and manages its own operational episodes.
• organizational modelAn organizational model is a structured representation of an organization's context, problems, opportunities, stakeholders, and possible solutions.
• business caseA business case is a factual justification for undertaking an initiative that describes its expected value, costs, feasibility, and constraints.
• ETLETL is a data-integration process that extracts data from sources, transforms it, and loads it into a target system or data store.
• evaluation datasetA validation set is a held-out subset of data used during model development to assess or tune model performance.
• feature engineeringFeature engineering is the transformation of raw data into informative input variables that represent useful structure for a machine learning model.
• ethical reviewEthical evaluation is the systematic assessment of a system's effects, risks, and compliance with ethical considerations.
• testTesting is the systematic checking of software, models, or processes against specified conditions to reveal defects or assess suitability.
• system requirementsA system requirement specifies deployment, maintenance, integration, consumption, resilience, or continuity conditions for a system.
• provenanceData provenance is the record of a data item's origin, derivation, transformations, and custody.
• serverless computingServerless computing is a cloud execution model in which a provider manages infrastructure and servers while workloads run on demand.
• DevOpsDevOps is a software delivery practice that integrates development and operations work through collaboration, automation, and shared responsibility.
• privacyData privacy is the protection and governance of information about people, including how it is collected, accessed, used, and disclosed.
• alignmentAlignment is the correspondence between a system's behavior, control mechanisms, and the goals and values of the humans responsible for it.
• GDPRGDPR is the European Union regulation governing the processing and protection of personal data.
• layered safeguardsLayered safeguards are a safety pattern that combines multiple independent checks and controls to reduce the likelihood and impact of system failures.
• robustnessRobustness is the ability of a system to maintain acceptable performance across specified circumstances and perturbations.
• prototypingA prototype is a preliminary implementation created to demonstrate and evaluate a proposed system or interaction.
• test environmentA test environment is a computing environment configured for checking software behavior, quality, and production-readiness.
• backlogA backlog is an ordered or enumerated list of work items and tasks for a project or development cycle.
• technical debtTechnical debt is the future cost created by expedient technical decisions that increase maintenance, correction, or evolution effort.
• feature storeA feature store is a system for storing, managing, and serving machine learning features for model development and inference.
• UI designUI design is the design of a software application's user interface and its interactive presentation.
• inductive biasInductive bias is the set of assumptions built into an algorithm or model architecture that favors some patterns or hypotheses over others.
• data augmentationData augmentation is the creation of additional training examples by applying transformations or alterations to existing data.
• feasibilityFeasibility is the degree to which a proposed project or solution can be successfully carried out within its available constraints and conditions.
• feedbackFeedback is information, opinions, or evaluations provided about work, behavior, or outcomes to support assessment and improvement.
• baselineA baseline performance measure is an initial quantitative assessment against which later model performance can be compared.
• visualizationA visualization is a graphical representation of data intended to reveal structure, patterns, relationships, or anomalies.
• business problemA business problem is an organizational difficulty, unmet need, or opportunity that a solution is intended to address.
• business valueBusiness value is the beneficial organizational impact produced by a product, service, or project outcome.
• story cardA story card is a structured artifact that records an application or model story, its organizational impact, relevant data, constraints, and intended use.
• business application descriptionA business application description is a document specifying an application's intended purpose, behavior, users, and business context.
• user experienceUX design is the design of user interfaces and interactions to support users' needs and the effects of system use.
• Apache AirflowAirflow is an open-source platform for authoring, scheduling, and monitoring programmatic workflows.
• directed acyclic graphA directed acyclic graph is a directed network of nodes and edges with no directed cycle.
• workflow managerA workflow engine executes and manages ordered computational tasks and their dependencies.
• delegative systemA delegative system is a system that makes decisions on a user's behalf while providing the human with timely mechanisms to review, correct, or override those decisions.
• data modelA data model is a formal representation of the entities, attributes, relationships, and constraints represented by data.
• machine learning readinessMachine learning readiness is the degree to which an organization has the data, infrastructure, processes, governance, and capabilities needed to undertake machine learning work.
• representative sampleA representative sample is a subset whose relevant characteristics adequately reflect those of the population from which it is drawn.
• schemaA schema is a formal specification of the structure, fields, types, and relationships of data.
• statistical sampleA data sample is a subset of a larger data set selected for inspection or analysis.
• data discovery ticketA data discovery ticket is a project work item for collecting and documenting information about data resources and their constraints.
• multi-armed banditA multi-armed bandit is an adaptive sequential decision method that allocates trials among alternatives while balancing exploration and reward.
• acceptance criteriaAcceptance criteria are conditions that a delivered system or project result must satisfy to be accepted.
• commercial feasibilityCommercial feasibility is the degree to which funding is available to pay for the work required to deliver a project.
• user storyA user story is a concise description of a system capability from the perspective of a person or group affected by or using the system.
• model governanceModel governance is the framework of ownership, documentation, oversight, lifecycle management, auditing, and operational control for machine learning models and the systems that contain them.
• sprintA sprint is a bounded chunk of work performed within an agile development process.
• integration testIntegration testing evaluates whether interacting software components or systems work together correctly through their interfaces.
• checklistA checklist is a structured list used to verify that required tasks or conditions have been addressed.
• normalizationNormalization is the transformation of data values into a consistent scale, unit, or representation so that they can be compared or combined appropriately.
• demand functionA demand model predicts the quantity or likelihood of demand for a product, service, or activity under specified conditions.
• deployment environmentA deployment environment is an isolated computing context in which software and its supporting services are developed, tested, or operated.
• infrastructure as codeInfrastructure as code is the declarative or programmatic management of computing infrastructure through versioned machine-readable definitions.
• alertingAlerting is the generation and delivery of notifications when monitored conditions indicate an issue or threshold violation.
• sign-offSystem signoff is formal approval that a system's design and implementation satisfy applicable compliance requirements.
• riskProject risk is the possibility that uncertainty in a project will cause adverse effects such as cost escalation, unmet targets, or failure to deliver.
• model designModel design is the specification of a machine learning model's intended structure, inputs, outputs, and behavior.
• probability distributionA distribution is the pattern of values, frequencies, or probabilities across the possible values of a variable.
• deep learningA deep neural network is a neural-network model composed of multiple computational layers that learn hierarchical representations from data.
• datasetA data set is a collection of data instances available for analysis or machine learning.
• AI agentAn AI system is a computational system that performs tasks associated with artificial intelligence, potentially incorporating machine learning models.
• post-hoc explanation techniqueA post hoc explanation is an interpretation generated after a model produces an output, rather than by an intrinsically interpretable model.
• usabilityUsability is the degree to which intended users can achieve goals with a system effectively, efficiently, and satisfactorily.
• categorical dataCategorical data is data whose values belong to a finite set of discrete categories rather than representing continuous quantities.
• offline evaluationOffline testing evaluates a model using previously collected or static data in a non-live environment before deployment.
• pre-project processA pre-project process is the set of activities performed before a project is funded or formally undertaken to assess feasibility, value, scope, and required resources.
• mitigationA mitigation is an action or measure intended to reduce the likelihood or impact of a risk.
• CRISP-DMCRISP-DM is a data-mining methodology that structures analytical work into phases such as business understanding, data understanding, preparation, modeling, evaluation, and deployment.
• machine learning engineeringMachine Learning Engineering is the discipline of applying engineering practices and design patterns to build and operate machine learning systems.
• machine learning system designMachine Learning System Design is a methodology and guide for designing production machine learning systems across problem definition, data, modeling, and serving.
• mlopsMLOps is the application of development, operations, automation, and lifecycle-management practices to machine learning systems.
• waterfall methodologyWaterfall is a sequential software development methodology in which requirements, design, implementation, testing, and acceptance proceed in ordered stages.
• signalA feature is an input variable or measurable attribute supplied to a machine learning model.
• observational dataObservational data is data collected from naturally occurring conditions without deliberately assigning interventions.
• automationAutomation is the execution of processes by software or machines with reduced manual intervention.
• driftModel drift is a change in the data, relationships, or operating conditions underlying a machine learning model that can reduce its suitability or performance over time.
• versioningVersion control is a system for tracking, comparing, and managing successive versions of files and artifacts.
• data qualityData quality is the degree to which data is accurate, complete, consistent, relevant, timely, and fit for its intended use.
• data aggregationAggregation is the combination of individual data values into summarized groups or totals.
• experimentationAn experiment is a structured investigation in which an uncertain proposition is tested through planned observations or actions.
• multi-criteria decision-makingMulti-criteria decision making is a decision-analysis approach for evaluating alternatives against multiple criteria, often with explicit weights or tradeoffs.
• Pareto frontierA Pareto front is the boundary formed by nondominated outcomes that represent the best available tradeoffs among multiple criteria.
• corporate social responsibilityCorporate social responsibility is an organization's responsibility to account for the social, legal, environmental, and ethical consequences of its activities.
• missing valueA null value is a database value representing the absence of a recorded value rather than a measured numeric value.
• quantitative dataNumerical data consists of values represented as numbers that support arithmetic comparison and quantitative analysis.
• select count(*) statementA SELECT COUNT(*) statement is a SQL query that returns the number of rows selected from a relation.
• data ingestionData ingestion is the process of acquiring data from source systems and bringing it into a target system for processing or storage.
• data integrationData integration is the combination of data from multiple sources into a coherent structure or access layer.
• data transformationData transformation is the conversion of data into a different structure, representation, or format.
• cross-validationCross-validation is a resampling technique that repeatedly trains and evaluates a model on different partitions of available data.
• deployment pipelineCI/CD is the practice of automating continuous integration, testing, and delivery or deployment of software changes.
• numeric variableA numerical field is a data field whose values represent measured or otherwise quantitative amounts.
• 5w1h+r frameworkThe 5W1H+R framework is a questioning framework covering why, who, what, when, where, how, and relationships to other data for investigating data assets.
• pre-project data investigationA pre-project data investigation is an inquiry performed before formal project execution to establish what data exists, how it can be used, and what risks it presents.
• performance testingA performance test checks whether a task meets specified resource-use or execution-time conditions.
• roadmapA project roadmap is a high-level, time-oriented plan that sequences project work, timing, dependencies, milestones, and major phases.
• Occam's razorOccam's razor is the principle of preferring the simplest explanation or model when alternatives have otherwise comparable adequacy.
• agencyAgency is the capacity of a person or system to make and carry out decisions that affect an unfolding situation.
• episodic controlEpisodic control is an autonomy-management technique in which a system independently performs bounded episodes of activity subject to human approval or intervention between episodes.
• fake delegative systemA fake delegative system is a system that nominally transfers control to a human but does so too late or with insufficient authority to prevent harm.
• partitioning strategyA partitioning strategy divides a system into independent subsystems so that an individual subsystem can fail without causing systemic or unacceptable damage.
• capture biasCapture bias is systematic distortion caused by the image positions, viewpoints, or presentation conditions represented in visual data.
• dimensionality reductionDimensionality reduction is the transformation of data into fewer dimensions while attempting to preserve informative structure.
• label coverageLabel coverage is the completeness of labels with respect to the items and relevant objects represented in a data set.
• embeddingAn embedding is a numerical vector representation that encodes meaningful properties of an input for comparison or downstream computation.
• review meetingA review meeting is a structured meeting for examining project work or readiness against defined criteria.
• functional problemA functional failure occurs when a system no longer provides an expected capability because its operating conditions, inputs, or underlying relationships have changed.
• study designA/B testing is a controlled experiment that assigns users or observations to alternative versions or conditions to compare their outcomes.
• deduplicationDeduplication is the process of identifying and removing redundant records that represent the same underlying entity or event.
• human agency and oversightHuman agency and oversight is an ethical principle requiring people to retain meaningful control over and supervision of an AI-enabled system.
• IEEE use case matrixThe IEEE use case matrix is a framework for relating artificial-intelligence application contexts to potentially relevant ethical concerns.
• linear regressionA linear model predicts an output as a linear combination of input variables and model parameters.
• normalization errorA normalization error is a mistake caused by inconsistent scales, units, representations, or aggregation conventions across data sources.
• rare categoryA rare entity is an infrequently occurring subject, case, or category represented in a data set.
• epsilon algorithmAn epsilon algorithm is a bandit strategy that selects the currently best alternative most of the time and explores another alternative with a specified probability.
• exploration-exploitation trade-offThe exploration-exploitation tradeoff is the tension between gathering information about uncertain alternatives and exploiting the alternative currently believed to perform best.
• throughputThroughput is the amount of work or data that a system processes within a specified time interval.
• model managementModel management is the discipline of organizing, tracking, evaluating, and governing machine learning models throughout their development and deployment lifecycles.
• documentation planA documentation plan is a documented arrangement specifying what project documentation will be created, maintained, and shared.
• contractAn API is a defined interface that enables software components to communicate and invoke available data or functionality.
• concept driftConcept drift is a change in the underlying relationships between inputs and outcomes that reduces the applicability of a previously trained model.
• costCost is the monetary expense associated with deploying or operating a model and its required components.
• data architectureA corporate data architecture is an organization's established structure for storing, integrating, governing, and processing business data.
• ML serving engineAn ML serving engine is software specialized for executing machine learning models in an operational environment.
• standup meetingA stand-up meeting is a brief recurring team meeting in which members report work progress, plans, and impediments.
• experiment trackingExperiment tracking is the systematic recording of machine learning experiment parameters, inputs, outputs, and performance results.
• estimateEstimation is the process of forecasting the effort, resources, schedule, or cost needed to deliver a project or its results.
• model-based analysisModel-based analysis uses a model to transform or summarize data so that its properties can be interpreted.
• varianceVariance is a statistical measure of the average squared deviation of values from their mean.
• convolutional neural networkA convolutional network is a neural network architecture that applies local filters over structured inputs to capture spatial or local patterns.
• XGBoostXGBoost is an open-source implementation of gradient-boosted decision trees for efficient supervised learning.
• Anscombe’s quartetAnscombe’s Quartet is a set of four paired data sets with nearly identical summary statistics but substantially different distributions and relationships.
• correlationPearson correlation is a standardized measure of the linear association between two numerical variables.
• samplingSampling is the process of selecting a subset of observations from a larger population or data set.
• scatter plotA scatter plot is a graph that represents paired observations as points on coordinate axes.
• f-scoreF1 score is the harmonic mean of precision and recall for a classification model.
• qualitative selectionQualitative selection is the process of judging candidate models for appropriateness using nonnumeric considerations, context, and requirements.
• constraintA system constraint is a limitation imposed by the capabilities, configuration, interfaces, or operating conditions of a customer system.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.