IBM Data Science Professional Certificate key terms

A six-phase process model for data science projects: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. It is iterative, so later findings often send teams back to earlier phases.

The CRISP-DM phase where analysts collect initial data, describe its structure and quality, and explore it to spot patterns or problems before any preparation or modeling work begins.

An open-ended investigation of a dataset using summary statistics and plots to reveal distributions, relationships, outliers, and missing values. Findings guide later cleaning, feature engineering, and model choice.

Examination of one variable at a time, typically through histograms, box plots, or frequency tables, to describe its center, spread, shape, and unusual values before comparing it with other variables.

A formal procedure that compares observed sample data against a null hypothesis, producing a test statistic and p-value used to decide whether an apparent effect is likely due to chance.

The probability of observing a test statistic at least as extreme as the one computed, assuming the null hypothesis is true. It is not the probability that the null hypothesis itself is correct.

A decision that an observed result is unlikely under the null hypothesis, conventionally when the p-value falls below a chosen threshold such as 0.05. It says nothing about practical importance or effect size.

A range of plausible values for a population parameter, built from sample data at a stated confidence level. Wider intervals reflect greater uncertainty, often caused by small samples or high variability.

A distortion that occurs when the data collected is not representative of the population it is meant to describe, so estimates and models systematically miss the truth regardless of sample size.

A modeling mistake where information unavailable at prediction time, such as target values or statistics computed on the full dataset, contaminates training and inflates measured performance unrealistically.

The work of transforming raw data into an analysis-ready form: handling missing values, correcting types, renaming columns, reshaping tables, merging sources, and removing duplicates or inconsistent records.

A tabular layout where each row is one observation, each column is one variable, and each cell holds a single value. This structure makes filtering, grouping, and plotting straightforward in pandas.

Replacing absent data with estimated values, such as the mean, median, or a model-based prediction, so that algorithms requiring complete records can run without discarding rows.

An observation that lies far from the rest of the distribution. Outliers can signal data-entry errors, rare but real events, or heavy-tailed behavior, and may distort means and distance-based models.

Creating or transforming input variables so that patterns become easier for a model to learn, for example deriving ratios, extracting dates, binning values, or encoding categories as numbers.

A technique that converts a categorical variable into several binary columns, one per category, so algorithms that expect numeric input do not treat category codes as ordered magnitudes.

Rescaling numeric inputs so they share a comparable range, through normalization to 0-1 or standardization to zero mean and unit variance. It matters for distance-based and gradient-based algorithms.

A modeling approach that learns a mapping from input features to a known target label, using labeled examples for training and then predicting that target for new, unseen records.

A supervised task predicting a discrete class label, such as churn or not churn. Models are judged with metrics like accuracy, precision, recall, and F1 score rather than error distance.

A supervised task predicting a continuous numeric value, such as price or demand. Performance is commonly summarized with mean absolute error, root mean squared error, or R-squared.

Dividing available labeled data into a portion used to fit a model and a separate portion used to estimate how well it generalizes to records it has never seen.

A resampling procedure, commonly k-fold, that repeatedly trains and validates a model on different data partitions, giving a more stable estimate of generalization performance than a single split.

A model fitting noise and idiosyncrasies of the training data so closely that it performs well there but poorly on new data. It is typically revealed by a widening gap between training and validation scores.

A model too simple or too constrained to capture the underlying pattern, producing weak results on both training and validation data. Adding relevant features or model complexity usually helps.

Adding a penalty on model complexity, such as L1 or L2 terms in regression, to discourage large coefficients and reduce overfitting while keeping the model's predictive structure intact.

Searching for settings that control how a model learns, such as tree depth or regularization strength, by evaluating candidate values with validation data rather than fitting them directly from training data.

A table crossing predicted and actual classes, showing true positives, true negatives, false positives, and false negatives. Every common classification metric can be derived from its four counts.

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 changes.

An unsupervised algorithm that partitions records into k groups by repeatedly assigning points to the nearest centroid and recomputing centroids. The analyst chooses k in advance.

Modeling that works without labeled targets, seeking structure such as clusters, low-dimensional representations, or anomalous records within the feature data itself.

A fixed-type, multidimensional container that stores data in contiguous memory and supports fast vectorized arithmetic, indexing, slicing, and broadcasting far beyond what plain Python lists offer.

NumPy's rule for combining arrays of different shapes by stretching size-one dimensions, letting code add a column vector to a matrix without writing explicit loops.

Expressing operations on whole arrays instead of element-by-element Python loops, which yields shorter code and much faster execution because work happens in compiled code.

A two-dimensional labeled table with row index and named columns, supporting mixed types, alignment, filtering, grouping, and joins. It is the standard structure for tabular analysis in Python.

A one-dimensional labeled array holding values of one type plus an index. Selecting a single column from a DataFrame returns a Series, and most column operations act on it.

A split-apply-combine operation that partitions rows by one or more keys, computes an aggregation on each group, and returns the combined result, enabling comparisons across categories.

A clause combining rows from two tables based on a matching condition. Inner joins keep matches only; left and right joins keep unmatched rows from one side with nulls filled in.

A column or set of columns whose values uniquely identify each row in a table. Keys enforce entity integrity and provide reliable join paths between related tables.

A column in one table whose values must match a primary key in another table, expressing a relationship between them and preventing references to records that do not exist.

A SQL function such as COUNT, SUM, AVG, MIN, or MAX that collapses many rows into one summary value, often combined with GROUP BY to summarize each category separately.

Python's foundational plotting library, offering fine-grained control over figures, axes, and marks. It underlies many higher-level tools and is used for precise, publication-style static charts.

A chart that bins a numeric variable into intervals and draws bars whose heights count observations per bin, revealing the distribution's shape, spread, skew, and possible multiple peaks.

The proportion of a chart's ink devoted to actual data rather than decoration. Maximizing it means removing gridlines, borders, and 3D effects that do not convey information.

A single screen combining several linked charts and key metrics, designed so a viewer can monitor performance or explore a question quickly without reading a long report.

Structuring an analysis presentation around a clear audience, a central message, and supporting evidence, choosing chart types and annotation so the key finding is understood at a glance.

An interactive document mixing executable code cells, their outputs, and narrative text, widely used for exploration and sharing analysis. Notebooks can be exported to HTML or slides.

A distributed system that records snapshots of files over time, letting collaborators branch, merge, and review changes. Commits, branches, and pull requests are core to sharing data science work.

A hosted project space containing code, notebooks, and documentation plus full change history. It supports issues, pull requests, and README files that make work reproducible and reviewable.

IBM's cloud environment for data science work, providing hosted notebooks, data assets, model training, and deployment tools within a managed project workspace.

Moving a trained model from an analysis environment into a service or application where it scores new data, including packaging, monitoring, and refreshing it as conditions change.

The property that an analysis can be rerun by someone else and yield the same results, achieved through versioned code and data, fixed random seeds, and documented environments.