Dell Data Science Associate key terms

A modeling approach where an algorithm learns a mapping from input features to a known target label using labeled training examples, then predicts that target for new, unseen data.

A modeling approach that finds structure in data without a target label, such as grouping similar records or reducing dimensionality, used for exploration and segmentation rather than prediction of a known outcome.

A supervised task that assigns each record to one of a fixed set of discrete categories, such as churn versus no churn, evaluated with metrics like accuracy, precision, recall, and AUC.

A supervised task that predicts a continuous numeric value, such as revenue, latency, or demand, typically evaluated with error measures like mean absolute error or root mean squared error.

An unsupervised technique that partitions records into groups so members of a group are more similar to each other than to records in other groups, useful for segmentation and anomaly exploration.

The subset of labeled data used to fit a model's parameters. It must be kept separate from validation and test data so performance estimates reflect genuinely unseen records.

A held-out subset used to tune hyperparameters and compare candidate models during development. It influences model choices, so it is not a clean estimate of final generalization performance.

A held-out subset touched only once at the end to estimate how a finalized model will perform on new data. Leakage into training or tuning makes this estimate optimistically biased.

A resampling procedure that splits data into folds, trains on all but one fold, and validates on the remainder repeatedly, producing a more stable performance estimate than a single split.

When a model captures noise specific to the training data and performs well there but poorly on new data. It is detected by a widening gap between training and validation performance.

When a model is too simple or poorly specified to capture the underlying pattern, producing weak performance on both training and validation data. Remedies include richer features and more flexible models.

The tension between error from overly simple assumptions (bias) and error from sensitivity to small fluctuations in training data (variance). Model complexity and regularization shift the balance between them.

A penalty added to a model's objective, such as L1 or L2 weight penalties, that discourages complexity and reduces variance. It trades a little training fit for better generalization.

A setting chosen before training rather than learned from data, such as tree depth, learning rate, or regularization strength. It is tuned using validation data or cross-validation.

A table of predicted versus actual classes showing true positives, false positives, true negatives, and false negatives, the basis for precision, recall, and related classification metrics.

Precision is the share of predicted positives that are truly positive; recall is the share of actual positives that were found. They trade off against each other as the decision threshold moves.

A condition where one outcome is far rarer than another, so accuracy becomes misleading and models tend to ignore the minority class. Remedies include resampling, class weights, and threshold tuning.

A model that predicts a continuous outcome as a weighted sum of input features plus an intercept, fitted by minimizing squared error. Coefficients are interpretable as marginal effects under its assumptions.

A model that recursively splits data on feature thresholds to maximize purity in the resulting groups. Single trees are interpretable but prone to high variance, which ensembles address.

An ensemble of decision trees trained on bootstrapped samples with random feature subsets, then averaged. This reduces variance relative to a single tree while retaining useful feature-importance signals.

An ensemble method that builds trees sequentially, each fitting the residual errors of the prior ensemble. It often achieves strong accuracy but requires careful tuning of learning rate and depth.

The process of choosing a subset of input variables that carry the most predictive signal, reducing noise, computation, and overfitting risk while improving model interpretability.

Creating, transforming, or combining raw variables into inputs that better expose patterns to a model, such as ratios, aggregations, time-since-event measures, or encoded categorical values.

Representing a categorical variable as binary indicator columns, one per category, so models that require numeric input do not impose a false ordering on nominal categories.

Rescaling numeric inputs to comparable ranges, through standardization or min-max normalization, so distance-based and gradient-based algorithms are not dominated by large-magnitude variables.

Filling or otherwise handling absent values, for example with mean, median, or model-based estimates, or by adding a missingness indicator, so records remain usable without distorting distributions.

An observation far from the bulk of the data, which may be a data error, a rare valid case, or a signal of interest. Handling depends on cause and on the model's sensitivity.

Detecting and correcting errors such as duplicates, inconsistent formats, impossible values, and mislabeled categories so downstream analysis and models rest on trustworthy records.

Rules and checks that confirm incoming data meet expected type, range, format, and referential constraints, catching problems at ingestion before they propagate into reports and models.

When information from the target or from the future leaks into training features, producing impressive validation scores that collapse in production. Common causes include pre-split scaling and time-unaware joins.

Patterns for moving data: extract, transform, load transforms before loading into storage; extract, load, transform loads raw data first and transforms it in the warehouse, leveraging warehouse compute.

An automated sequence of steps that ingests, transforms, validates, and delivers data to storage or consumers, ideally with scheduling, monitoring, and failure handling built in.

A centralized, structured repository optimized for analytical queries, holding integrated data from many sources organized for reporting and business intelligence rather than transactional processing.

A storage layer holding raw and semi-structured data in native formats at scale, with schema applied on read. Governance and cataloging are essential to prevent it becoming an unusable swamp.

Analytical processing optimized for aggregating large volumes across many dimensions, typically using columnar storage and cube-style queries, in contrast to transaction-focused OLTP workloads.

A visual interface presenting key metrics and trends, usually refreshed on a schedule, designed to let business users monitor performance and drill into detail without writing queries.

A key performance indicator: a small set of metrics tied directly to business objectives, defined with agreed formulas and owners so progress can be tracked consistently over time.

Presenting analysis as a clear narrative that connects findings to a decision, pairing the right chart with context, a recommendation, and appropriate caveats for a non-technical audience.

Encoding data as visual marks such as bars, lines, or scatter points so patterns and comparisons are perceived quickly. Chart choice should match the question and data type.

Summary measures that characterize a dataset, including central tendency like mean and median, spread like standard deviation and range, and distribution shape, used to profile data before modeling.

A mathematical description of how likely each possible value of a variable is. Common forms include normal, binomial, and Poisson, each suited to different data-generating processes.

A framework for judging whether an observed effect is likely due to chance, comparing a test statistic to a null distribution and reporting a p-value and effect size.

The probability of observing a result at least as extreme as the one obtained, assuming the null hypothesis is true. It is not the probability that the null hypothesis is true.

A range of plausible values for a population parameter, constructed so that a stated percentage of such intervals would contain the true value across repeated sampling.

A controlled experiment comparing two variants on a metric, with random assignment and pre-set sample size and decision rules, to estimate causal impact of a change.

A measure of the strength and direction of linear association between two variables, ranging from minus one to one. It does not by itself establish causation.

The distinction between variables moving together and one variable actually producing change in another. Confounders and reverse causality can create correlation without a causal link.

Techniques such as principal component analysis that compress many variables into fewer informative dimensions, reducing noise and computation while often aiding visualization.