Predictive Analytics and Machine Learning for HR

Introduction and Context

Human resource management has traditionally relied on descriptive questions: How many employees left? Which departments have vacancies? How many training hours were delivered? Predictive analytics extends these questions by asking what is likely to happen and which interventions may be useful. Examples include estimating the probability that an employee will leave, forecasting staffing demand, identifying groups with similar development needs, and analysing themes in employee comments.

Machine learning (ML) is a collection of computational methods that learn relationships from data to produce predictions, classifications, groupings, or representations. In HR, ML does not replace professional judgement. It supports decisions by making patterns more visible, consistent, and scalable. However, workforce data are especially sensitive. Employees may be affected by predictions about promotion, retention, performance, absence, or recruitment. Consequently, technical accuracy must be considered alongside fairness, privacy, transparency, employment law, organizational context, and the possibility that a prediction may change the behaviour it is intended to forecast.

A useful HR analytics workflow is:

1. Define the organizational decision and the prediction target.
2. Identify the population, time period, and unit of analysis.
3. Acquire, document, clean, and secure the data.
4. Engineer features without using information unavailable at the prediction date.
5. Split data appropriately into training, validation, and test sets.
6. Fit baseline and candidate models.
7. Evaluate predictive performance, calibration, fairness, and stability.
8. Interpret results and determine whether an intervention is justified.
9. Monitor outcomes, drift, and unintended consequences after deployment.

Learning Objectives

After studying this chapter, you should be able to:

• distinguish prediction, explanation, correlation, and causal inference;
• prepare workforce data for supervised and unsupervised learning;
• apply and interpret regression, classification, clustering, and text analytics;
• perform feature engineering and avoid target leakage;
• compare models using suitable validation methods and metrics;
• explain overfitting, class imbalance, drift, fairness, and limits to prediction;
• translate model results into responsible HR decisions.

1. From HR Questions to Predictive Problems

A predictive project begins with a precise outcome. The outcome, or target variable, is denoted by y. Predictors or features are denoted by x₁, x₂, ..., x_p, where p is the number of features. A model estimates a function f such that:

Here, ŷ is the predicted value. A model can estimate an outcome without establishing that any feature causes it. For example, the number of late arrivals may predict resignation, but changing lateness may not cause resignation; both may reflect workload, health, scheduling, or management conditions.

Prediction versus explanation and causation

Prediction asks, “How accurately can an outcome be estimated for new cases?” Explanation asks, “Which relationships are consistent with a theory or mechanism?” Causal inference asks, “What would happen if an intervention changed a particular factor?” These are different objectives. A highly predictive variable may be inappropriate as an intervention target, and a variable with a causal effect may have weak predictive power if it varies little in the available data.

The unit of analysis must also be explicit. A dataset may contain one row per employee, employee-month, job application, absence episode, or survey response. Mixing units can create errors. For instance, joining employee-level information to multiple absence records may duplicate employees and give high-absence employees disproportionate influence.

The prediction time, sometimes called the scoring date, is equally important. Features must represent information available at that date, and the target must occur afterward. A retention model scored on 1 January might use tenure and prior engagement surveys, but not a resignation interview completed in February.

2. Preparing Workforce Data

Data quality and governance

Workforce datasets commonly contain missing values, inconsistent job titles, changing organizational structures, duplicate records, and sensitive attributes. Preparation should include:

• a data dictionary defining fields, units, permissible values, and ownership;
• unique identifiers and documented joining rules;
• checks for duplicates, impossible dates, negative working hours, and inconsistent categories;
• treatment of missingness based on its likely meaning;
• access controls, retention limits, and an audit trail;
• separation of directly identifying information from modelling data where possible.

Missingness is not automatically random. Missing training records may mean that an employee had no recorded access, that a system was not used, or that a manager failed to report completion. Replacing all missing values with zero can therefore encode a false assumption. Numerical variables may be imputed using a training-set median, while categorical variables may use an explicit “unknown” category. The imputation rule must be learned from the training data only and then applied unchanged to validation and test data.

Categorical features such as department or contract type may be represented using one-hot encoding. If a variable has categories A, B, and C, a model can use indicator variables for two categories and treat the omitted category as a reference. Including all indicators with an intercept creates perfect linear dependence, known as the dummy-variable trap, unless a suitable regularization or encoding strategy is used.

Numerical scaling is important for distance-based methods and regularized models. Standardization is commonly defined as:

where xᵢ is an observation, x̄ is the training-set mean, and s is the training-set sample standard deviation. The transformed value has no original unit and is approximately centred at zero with unit sample variance. Scaling parameters must not be calculated using the test set.

Feature engineering

Feature engineering converts raw records into variables that are meaningful at the prediction date. Examples include:

• tenure in months, calculated from hire date to scoring date;
• rolling absence count over the preceding 90 days;
• time since last promotion;
• ratio of completed training modules to assigned modules;
• workload change relative to the previous quarter;
• text-derived sentiment or topic indicators.

Features should be interpretable, stable, and operationally available. A feature that is merely a proxy for a protected characteristic may reproduce discrimination even if the protected characteristic is removed. Postal area, language, educational institution, and employment gap may sometimes act as proxies for race, disability, socioeconomic status, or other sensitive attributes.

Target leakage

Target leakage occurs when a feature contains information about the target that would not have been available when the prediction was made. Examples include using an exit-interview rating to predict resignation, using a final disciplinary action to predict termination, or calculating an engagement score from a survey conducted after the outcome period. Leakage often produces impressive validation scores but poor real-world performance.

A practical test is: “Could a decision-maker retrieve this feature at the exact scoring time, without looking into the future?” If not, the feature should be excluded or redefined.

3. Supervised Learning: Regression and Classification

Supervised learning uses labelled examples in which the target is known. Regression predicts a numerical outcome, such as absence days, time to fill a vacancy, or training expenditure. Classification predicts a category or event, such as voluntary turnover within six months.

Linear regression

In multiple linear regression, the conditional mean of a continuous outcome is represented as:

Here, yᵢ is the observed outcome for case i, xᵢⱼ is feature j, β₀ is the intercept, βⱼ is the coefficient associated with feature j, and εᵢ is the error term. The model is most naturally interpreted as an estimate of the conditional mean when errors have mean zero given the predictors. Ordinary least squares estimates coefficients by minimizing the sum of squared residuals:

The coefficient βⱼ is the expected change in the outcome associated with a one-unit increase in xⱼ, holding the other included features constant. This interpretation depends on the model specification and does not itself establish causality.

Important assumptions include a correctly specified conditional mean, independent or appropriately modelled observations, and reasonable treatment of variance and influential observations. Normality of errors is not required for calculating fitted values, but it can matter for small-sample inferential procedures. If the outcome is a count with strong skewness, ordinary linear regression may be unsuitable; transformations, count models, or other approaches may be considered.

Useful regression metrics include mean absolute error:

and root mean squared error:

where n is the number of evaluated cases. MAE is in the same units as the target and weights all absolute errors linearly. RMSE is also in target units but penalizes larger errors more strongly. Neither metric alone establishes that a model is useful for a business decision.

Logistic classification

For a binary outcome, let pᵢ = P(yᵢ = 1 | xᵢ) be the predicted probability of an event. Logistic regression models the log-odds:

The odds ratio for a one-unit increase in xⱼ, holding other variables constant, is e^βⱼ. This interpretation concerns the modelled association, not a causal effect. Predicted probabilities can be converted into classifications using a threshold, but the threshold should reflect decision costs and capacity rather than automatically being set to 0.50.

A confusion matrix records true positives, false positives, true negatives, and false negatives. From it:

Precision answers, “Among cases flagged positive, how many are positive?” Recall answers, “Among actual positive cases, how many were found?” Accuracy can be misleading when the positive class is rare. The F1 score is the harmonic mean of precision and recall:

The area under the receiver operating characteristic curve (ROC AUC) measures ranking ability across thresholds. It does not guarantee good probability calibration. Calibration asks whether, among cases assigned a probability of 0.70, approximately 70% experience the outcome in a comparable population and time period.

Regularization and model complexity

When predictors are numerous or correlated, regularization can reduce variance. Ridge regression minimizes squared error plus a penalty proportional to the sum of squared coefficients:

Lasso uses the absolute values of coefficients instead:

Here, λ ≥ 0 controls penalty strength and the intercept is usually not penalized. Larger λ shrinks coefficients more strongly; lasso may set some coefficients exactly to zero. The value of λ must be selected using training data and validation, not the test set.

Tree-based models, random forests, gradient boosting, and neural networks can represent nonlinear relationships and interactions. Their flexibility may improve predictive accuracy, but it can reduce transparency and increase overfitting risk. A simpler model with slightly lower statistical performance may be preferable if it is easier to validate, explain, govern, and act upon.

4. Validation, Comparison, and Overfitting

A model must be evaluated on data that were not used to fit it. A common design is a training set for estimating parameters, a validation set for selecting models and hyperparameters, and a final test set for one-time performance estimation. In small datasets, cross-validation divides the training data into folds, fits the model on some folds, and evaluates it on the remaining fold. The process is repeated so each observation serves as validation data.

For workforce data observed over time, random splitting can be inappropriate. If the model will predict future employees or future months, temporal validation is more realistic: earlier periods are used for training and later periods for evaluation. If multiple records belong to the same employee, group-based splitting is needed to prevent records from one employee appearing in both training and test sets.

Overfitting occurs when a model captures noise or idiosyncrasies in the training sample rather than generalizable structure. Training error may continue to fall as complexity increases while test error rises. Underfitting occurs when a model is too simple to capture relevant structure. The bias–variance perspective describes the trade-off: high bias reflects systematic simplicity, while high variance reflects sensitivity to the particular sample.

Model comparison should use:

• a meaningful baseline, such as the historical mean or majority class;
• metrics aligned with the decision, including costs of false positives and false negatives;
• confidence intervals or variation across folds where feasible;
• calibration and subgroup performance, not only a single global score;
• operational considerations such as latency, data availability, explainability, and maintenance.

A test set should remain untouched until the modelling decisions are complete. Repeatedly checking test performance turns the test set into an informal validation set and makes the reported performance optimistic.

5. Clustering and Unsupervised Learning

Unsupervised learning works without a labelled target. Clustering may help segment employees according to development needs, work patterns, or career interests. It is exploratory: a cluster is a mathematical grouping, not a naturally existing or psychologically confirmed category.

K-means clustering assigns observations to k clusters by minimizing within-cluster squared Euclidean distance:

Here, Cᵣ is cluster r, μᵣ is its centroid, and the norm measures distance in feature space. K-means assumes that distance is meaningful, so scaling and feature selection are crucial. It is sensitive to outliers and the chosen value of k. Analysts may compare several values using within-cluster variation, silhouette measures, stability, and—most importantly—whether the groups are actionable and ethically acceptable.

Hierarchical clustering creates a nested tree of groupings and can be useful when analysts want to inspect multiple levels of segmentation. Density-based methods can identify irregularly shaped groups and outliers. No clustering algorithm automatically discovers “high-potential employees” in a valid managerial sense; that label requires a defensible construct, relevant evidence, and careful governance.

6. Text Analytics for Employee Voice

HR text sources include engagement comments, exit interviews, help-desk tickets, job applications, and occupational health narratives. Text analytics can identify recurring themes, classify documents, or summarize large volumes of comments. It should augment, not replace, reading and contextual interpretation.

A document-term matrix represents documents as rows and terms as columns. Term frequency–inverse document frequency (TF–IDF) gives greater weight to terms frequent in a document but less common across the collection. One common form is:

where tf(t,d) is the frequency of term t in document d, N is the number of documents, and df(t) is the number of documents containing the term. Different software may use normalized or smoothed variants, so the exact convention should be documented.

Preprocessing can include lowercasing, tokenization, removal of punctuation, and handling of stop words. Stemming or lemmatization may reduce word forms, but can also remove meaningful distinctions. Sentiment models may perform differently across occupations, dialects, languages, and contexts. Sarcasm, coded language, short comments, and domain-specific terms are difficult for automated systems.

Modern language models can generate embeddings that represent semantic relationships and can support search, clustering, or classification. Their use requires attention to confidentiality, vendor data handling, prompt and output logging, bias, hallucination, and the risk that a generated summary obscures minority viewpoints. Text should be de-identified where possible, and access should follow the sensitivity of the source.

7. Interpretation, Fairness, and Responsible Use

Interpretability means being able to communicate how a model produces or supports a result. A global explanation describes overall patterns; a local explanation describes a particular prediction. Coefficients are relatively direct in a well-specified linear or logistic model, but correlated features complicate interpretation. Tree-based feature importance may reflect frequency or predictive usefulness rather than causal importance. Post-hoc explanation methods can be useful, but they are approximations and should not be treated as proof of how a model reasons.

Fairness assessment should be designed around the decision and relevant legal and organizational requirements. Suppose a selection model produces positive decisions for protected group A and reference group B. Selection rate is:

A selection-rate ratio can be written as SR_A / SR_B. Other measures compare false-positive rates, false-negative rates, precision, recall, or calibration. These criteria can conflict, particularly when base rates differ, so no single metric resolves fairness in every context. Fairness analysis should consider data generation, labels, measurement quality, decision thresholds, intersectional groups, and the consequences of errors.

Protected attributes may need to be used for auditing even when excluded from prediction. Removing them does not remove indirect proxies. At the same time, collecting sensitive attributes creates privacy obligations. Governance should define legitimate purpose, access, retention, challenge and appeal procedures, human oversight, documentation, and rules against using predictions as automatic grounds for adverse action.

Privacy principles include data minimization, purpose limitation, security, transparency, and proportionality. Depending on jurisdiction, employment and data-protection law may regulate automated decision-making, profiling, special-category data, employee monitoring, and the right to explanation or human review. Legal compliance is jurisdiction-specific; organizations should obtain qualified legal advice rather than assuming that a technically accurate model is lawful.

8. Drift and Deployment

A model can deteriorate after deployment because the workforce, labour market, policies, technology, or data collection process changes. Data drift means the distribution of predictors changes. Concept drift means the relationship between predictors and the target changes. Label drift means the outcome prevalence changes. For example, a turnover model trained before a major restructuring may perform poorly after new remote-work policies.

Monitoring should include data completeness, feature distributions, prediction distributions, calibration, subgroup metrics, realized outcomes, and intervention effects. Alerts should have predefined review procedures. Retraining is not automatically the solution: if the target is poorly defined, retraining may simply reproduce a problematic process. Versioning, model cards, decision logs, and documentation of training data and limitations support accountability.

A crucial distinction is between predictive performance and intervention value. If a model identifies employees likely to leave, contacting them may reduce turnover, have no effect, or increase it by signalling that the organization expects departure. The intervention must therefore be evaluated, ideally through a carefully governed experiment or quasi-experimental design, with attention to consent, fairness, and unintended effects.

Worked Example: A Voluntary Turnover Classifier

An organization wants to identify employees at risk of voluntary resignation within the next six months. The scoring date is 1 January. Candidate features include tenure, prior-year absence days, number of manager changes in the previous year, engagement score from December, and participation in a January retention meeting.

The January meeting must be excluded because it occurs after the scoring date and may be influenced by the organization’s response to risk. The target is binary: 1 if voluntary resignation occurs between 1 January and 30 June, and 0 otherwise. The data are split temporally: records from 2021–2022 form training data, 2023 is validation data, and 2024 is the final test period.

Suppose the test set contains 1,000 employees, of whom 100 resign. A model flags 80 employees; 50 of those resign, while 30 do not. It misses 50 resignations, and 870 employees are correctly not flagged. Therefore:

• precision = 50 ÷ (50 + 30) = 0.625, or 62.5%;
• recall = 50 ÷ (50 + 50) = 0.50, or 50%;
• specificity = 870 ÷ (870 + 30) ≈ 0.967, or 96.7%;
• F1 = 2 × (0.625 × 0.50) ÷ (0.625 + 0.50) ≈ 0.556.

The model finds half of the resignations, and 62.5% of flagged cases are actual resignations in this test period. These results do not show that the model causes retention or that the flagged employees should be treated differently. HR must determine whether supportive, non-punitive interventions are available, whether employees would reasonably expect this use of data, and whether performance differs across groups.

A baseline that flags no one has zero recall but may avoid unnecessary interventions. A policy that contacts every employee has 100% recall but may be costly and intrusive. The appropriate threshold depends on intervention capacity, error costs, and fairness constraints. The model should be calibrated and monitored after implementation, and outcomes should be assessed against an appropriate comparison group.

Illustrative Case: Northstar Services

Background

Northstar Services, a fictional customer-support organization, has experienced rising first-year turnover. Leaders propose an ML system to identify “flight risks” and automatically send managers a list of employees requiring retention action.

Problem

The initial dataset includes employee demographics, attendance, schedules, performance ratings, learning records, manager changes, and free-text exit interviews. An analyst discovers that the first model uses “exit interview completed” and “counteroffer discussed.” Both variables occur near or after resignation and create leakage. The model also performs better statistically for employees in large offices because small-office records are sparse.

Analysis

A cross-functional team redefines the target as voluntary resignation within six months of a monthly scoring date. It uses only information available at that date and applies employee-level and temporal splitting. The team compares logistic regression, a regularized model, and a gradient-boosted tree model against a majority-class baseline. Evaluation includes recall, precision, calibration, subgroup performance, and stability across months.

The tree model has the highest ROC AUC but is poorly calibrated for smaller offices. The regularized logistic model has slightly lower ranking performance, clearer relationships, and more stable calibration. Its strongest associations involve recent schedule instability and low access to development opportunities. These are treated as signals for reviewing working conditions rather than as proof that an employee is personally responsible for turnover.

Approach and solution

Northstar rejects automatic adverse action. Predictions are used only to offer voluntary conversations, schedule reviews, career support, and workload assessment. Employees are not shown a risk score, but receive information about available support and a channel to challenge inaccurate records. Sensitive attributes are retained in a restricted audit dataset to evaluate disparities. The model is reviewed quarterly, and the retention programme is evaluated using a pre-specified comparison design.

Outcome

After two quarters, the organization finds modest improvement in retention among employees offered schedule support, but no improvement from generic messages. The model’s calibration worsens after a policy change affecting shifts, triggering review and retraining. Northstar concludes that the value came less from ranking individuals than from revealing organizational conditions that could be improved.

Learning points

1. A clear prediction date is essential for preventing leakage.
2. The most accurate model is not necessarily the best governed model.
3. Predictive signals should be translated into supportive, controllable interventions.
4. Fairness, calibration, drift, and operational impact require continuing monitoring.
5. HR analytics should examine systems and working conditions, not only employee risk.

Best-Practice Checklist

Before approving an HR predictive model, ask:

• Is the decision and target clearly defined?
• Is the prediction useful, proportionate, and legally permissible?
• Are the population, unit of analysis, and scoring date documented?
• Could any feature contain future information or a proxy for protected status?
• Were imputation, encoding, scaling, and feature selection performed within the validation design?
• Was a simple baseline compared with more complex models?
• Are test data genuinely held out?
• Have calibration, subgroup performance, and error costs been examined?
• Can affected people obtain meaningful human review or correction?
• Are drift, security, model changes, and intervention outcomes monitored?

Challenges and Limits of Prediction

First, historical labels may reflect past managerial decisions rather than underlying employee potential or effort. A model trained on promotion decisions can reproduce unequal access to promotion. Second, missing and noisy data are common; employee engagement is not fully represented by survey responses, and performance ratings may vary across managers. Third, predictions are probabilistic. A high-risk score does not mean that an event will occur for a particular person. Fourth, feedback loops can arise when predictions influence treatment and later data are used to retrain the model. Fifth, organizational changes can invalidate historical relationships. Finally, prediction cannot determine what the organization ought to do. Normative judgements about dignity, fairness, acceptable surveillance, and worker voice remain human and institutional responsibilities.

Chapter Summary

Predictive analytics for HR transforms workforce data into estimates, classifications, groupings, and text-based insights. Effective practice begins with a precise decision, target, unit of analysis, and prediction date. Data preparation requires careful treatment of missing values, categories, scaling, duplicates, sensitive information, and temporal structure. Feature engineering can improve usefulness, but target leakage and proxy discrimination must be actively controlled.

Regression estimates numerical outcomes, while classification estimates event probabilities or categories. MAE and RMSE support regression evaluation; precision, recall, F1, ranking metrics, and calibration support classification evaluation. Clustering provides exploratory segmentation rather than objectively true employee types. Text analytics can organize employee voice but is sensitive to context, language, privacy, and model bias.

Validation must reflect deployment, especially time and employee grouping. Overfitting, class imbalance, drift, and changing organizational conditions limit generalization. Model interpretation should distinguish association from causation. Responsible HR use requires privacy, fairness assessment, transparency, human oversight, appeal mechanisms, security, documentation, and monitoring. The ultimate value of an HR model depends not only on predictive accuracy but also on whether it supports fair, effective, and humane organizational action.

Key Terms

Target variable: The outcome a model is designed to predict.

Feature: An input variable used by a predictive model.

Supervised learning: Learning from observations with known target values.

Unsupervised learning: Finding structure in data without a labelled target.

Regression: Prediction of a numerical or continuous outcome.

Classification: Prediction of a category or event, often represented by a probability.

Log-odds: The natural logarithm of the odds of an event.

Regularization: A method that penalizes model complexity to reduce overfitting.

Feature engineering: Creating useful model inputs from raw or joined data.

Target leakage: Use of information unavailable at the prediction time.

Overfitting: Fitting noise or sample-specific patterns that do not generalize.

Cross-validation: Repeated training and validation across data partitions.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Clustering: Grouping observations according to similarity.

TF–IDF: A text representation weighting terms by document frequency and collection rarity.

Class imbalance: A situation in which outcome classes have substantially different frequencies.

Data drift: Change in the distribution of model inputs or outputs over time.

Concept drift: Change in the relationship between predictors and the outcome.

Proxy variable: A feature that indirectly represents a sensitive or protected attribute.

Human oversight: Meaningful review and responsibility by authorized people rather than unexamined automation.

Review Questions

Short Answer Questions

1. Why must the prediction date be defined before selecting features?
2. Distinguish precision from recall.
3. What is target leakage? Give one HR example.
4. Why is scaling important for k-means clustering?
5. Why does a high ROC AUC not guarantee good calibration?

Long Answer Questions

1. Explain how an HR analyst should prepare workforce data for a voluntary-turnover classifier, including governance, missing data, feature engineering, and leakage prevention.
2. Compare linear regression, logistic regression, tree-based models, and clustering for HR applications.
3. Explain how training, validation, and test data should be designed for employee records observed over time.
4. Discuss why predictive accuracy alone is insufficient for evaluating an HR model.
5. Explain how privacy, fairness, transparency, and human oversight should influence deployment of predictive HR analytics.

Analytical / Application Questions

1. An organization has a 5% turnover rate. A model achieves 95% accuracy by predicting “no turnover” for everyone. Evaluate this result and propose more appropriate metrics.
2. A promotion model uses manager rating, overtime, office location, and participation in an optional leadership programme. Identify possible validity, fairness, and proxy concerns, and propose an evaluation plan.
3. A turnover model’s recall falls sharply after a new remote-work policy. Diagnose possible forms of drift and design a monitoring and response process.

Multiple Choice Questions with Answers

Question 1

Which feature is most clearly an example of target leakage when predicting resignation on 1 January?

A. Employee tenure on 1 January

B. Absence days during the preceding 90 days

C. Participation in a retention meeting held in February

D. Engagement score recorded in December

Correct Answer: C. Participation in a retention meeting held in February

Question 2

What does recall measure in binary classification?

A. The proportion of flagged cases that are actual positives

B. The proportion of actual positives that are correctly identified

C. The proportion of all cases that are correctly classified

D. The proportion of actual negatives that are correctly identified

Correct Answer: B. The proportion of actual positives that are correctly identified

Question 3

Which metric is expressed in the same units as a continuous target and weights all absolute errors linearly?

A. Mean absolute error

B. ROC AUC

C. Precision

D. F1 score

Correct Answer: A. Mean absolute error

Question 4

In logistic regression, exp(βⱼ) is interpreted as the:

A. Change in the outcome measured in original units

B. Predicted probability when all features equal zero

C. Odds ratio for a one-unit increase in feature j, holding other included features constant

D. Percentage of correctly classified observations

Correct Answer: C. Odds ratio for a one-unit increase in feature j, holding other included features constant

Question 5

Why should a temporal split often be used when predicting future employee turnover?

A. It guarantees that the model will be unbiased

B. It reflects the time order in which the model will be deployed

C. It eliminates the need for a test set

D. It ensures every class has equal frequency

Correct Answer: B. It reflects the time order in which the model will be deployed

Question 6

What is the primary objective of k-means clustering?

A. Maximize logistic-regression likelihood

B. Minimize within-cluster squared distance to cluster centroids

C. Estimate a causal effect of training

D. Maximize classification recall at a fixed threshold

Correct Answer: B. Minimize within-cluster squared distance to cluster centroids

Question 7

In the TF–IDF expression, a term appearing in many documents generally receives a lower inverse-document-frequency weight because it is:

A. Rare across the collection

B. Less distinctive across the collection

C. Always longer than other terms

D. Automatically a protected attribute

Correct Answer: B. Less distinctive across the collection

Question 8

Which statement best describes calibration?

A. It measures whether predicted probabilities correspond to observed frequencies

B. It measures only the number of features in a model

C. It guarantees that a model is causally valid

D. It is identical to recall at a threshold of 0.50

Correct Answer: A. It measures whether predicted probabilities correspond to observed frequencies

Question 9

A model has TP = 50 and FP = 30. What is its precision?

A. 37.5%

B. 50.0%

C. 62.5%

D. 80.0%

Correct Answer: C. 62.5%

Question 10

What is concept drift?

A. A change in the relationship between predictors and the target

B. A method for scaling numerical variables

C. A technique for removing duplicate employees

D. A guarantee that model performance improves over time

Correct Answer: A. A change in the relationship between predictors and the target