Data Science, Machine Learning, and Ethics Review

Flashcards covering core principles of data science lifecycle, exploratory data analysis, sampling bias, machine learning algorithms, evaluation metrics, data privacy frameworks, and professional ethics.

Variable

A characteristic or feature.

Value

The state of a variable.

Examples

Data points or observations represented as a tuple of (variable, value).

Label

A dependent variable.

Volume (4 V's)

The amount of data generated every second.

Velocity (4 V's)

The speed at which data is generated and processed.

Variety (4 V's)

The different types and sources of data.

Veracity (4 V's)

The trustworthiness, completeness, consistency, and accuracy of data.

Data Tidying

The process of organising the structure of data.

Data Cleaning

The process that fixes missing, invalid, inconsistent, or duplicated data.

Series (Pandas)

Columns in Pandas, including the header and values.

Tidy Data

A standard data format where each variable is a column, each row is an observation, and each cell is a single value.

Density Plot

A continuous histogram where the total area under the curve equals $1$.

Smoothing Curve

A scatter plot that uses a moving average to remove noise.

Hexbin Plot

A 2D density visualization using a normal scatter plot grouped into hexagon bins, where colour indicates the density or number of points in a bin.

Faceting

Putting multiple graphs side by side in a table, where each subplot represents a different category.

Sampling Frame

The list of subjects in the population from which a sample is drawn (for example, a list of voters).

Attrition Bias

A sampling bias caused by dropouts when they systematically differ from non-dropouts in some way.

Survivorship Bias

A sampling bias that occurs when only individuals who have finished, continued, or survived are included in the sample.

Model

A trained algorithm that has learned patterns and can predict future outcomes.

Generative AI

Artificial intelligence that creates new content (such as text or images) and creates data from data.

Adjunctive AI

Artificial intelligence used as an extra or supporting tool alongside a human expert rather than replacing them.

Supervised Learning

A machine learning paradigm that relies on labelled training data and known answers.

Unsupervised Learning

A machine learning paradigm that works with unlabelled data to discover patterns and generate answers itself without a simple end goal.

$k$-means Clustering

An algorithm that splits data into $k$ distinct clusters to reduce intra-cluster variation, typically quantised using squared Euclidean distance.

$k$-means++

A variant of $k$-means that selects initial centroids that are far away from each other.

Inertia

The sum of squared distances between data points and their closest centroids.

Silhouette Score

The mean of silhouette coefficients for every point, where a score near $1$ means points are well inside their own cluster, $0$ is near the boundary, and $-1$ is in the wrong cluster.

Agglomerative Clustering

A hierarchical clustering approach that starts with every point in its own group and iteratively groups similar ones together.

Divisive Clustering

A hierarchical clustering approach that starts with all data points in one big group and iteratively splits dissimilar ones.

Confounding Variable

A third variable that influences the association between two other variables.

Lurking Variable

A variable that was forgotten to be taken into account in the analysis.

Classification

A supervised learning method where a model learns from data to assign categorical labels or discrete categories to an object.

Naïve Bayes

A classification method that applies conditional probability across independent variables to assign instances to the class with the highest probability.

Overfitting

A condition occurring when a model has too many parameters, fits too closely to the data, and loses the overall structure.

Underfitting

A condition occurring when a model has too few parameters and is too simple to capture the underlying structure.

Support Vector Machine (SVM)

A supervised machine learning algorithm that finds an optimal line or plane in $n$ dimensions to separate classes by maximising the boundary gap between the closest points of different classes.

$k$-Nearest Neighbour

A classification algorithm that assigns a data point to the majority class among its $k$ nearest neighbours.

Data Leakage

When information from the test set leaks into the training set, causing the model to overperform artificially on testing.

Null Model

A baseline model that classifies every instance as the exact same class.

Single Variable Model

A model that relies solely on a single, best feature to make predictions.

Precision

The fraction of items flagged as being in a class that actually belong to that class, calculated as $\frac{\text{TP}}{\text{TP} + \text{FP}}$.

Specificity

The true negative rate, calculated as $\frac{\text{TN}}{\text{TN} + \text{FP}}$.

Sensitivity / Recall

The fraction of actual items in a class that are correctly detected by the classifier, calculated as $\frac{\text{TP}}{\text{TP} + \text{FN}}$.

F1 Score

A metric combining precision and recall, calculated as $\frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$, where $0$ is worst and $1$ is best.

Area Under Curve (AUC)

A metric that measures the two-dimensional area underneath the Receiver Operating Characteristic (ROC) curve, ranging from $0.5$ for random guessing to $1.0$.

Batch Gradient Descent

A gradient descent variant that calculates the loss gradient using all points in the dataset at every iteration.

Stochastic Gradient Descent

A gradient descent variant that calculates the gradient using only a single random data point at each step.

Mini-batch Gradient Descent

A gradient descent variant that calculates the gradient using a small random subset of points at each step.

Gradient Descent Update Rule

The parameter update formula $\theta_{\text{new}} = \theta_{\text{old}} - \eta \nabla L(\theta)$, where $\theta$ represents model parameters, $\eta$ is the learning rate, and $\nabla L$ is the loss gradient.

Findable (FAIR Principles)

The principle stating that metadata and data should be easy to locate for humans and computers via machine-readable metadata.

Accessible (FAIR Principles)

The principle stating that once required data is identified, users need to know how it can be accessed, including necessary authentication and authorisation.

Interoperable (FAIR Principles)

The principle that data must be capable of integrating with other data and operating across different workflows, applications, and storage environments.

Reusable (FAIR Principles)

The ultimate goal of FAIR, requiring well-described metadata and data so they can be replicated or combined in diverse settings.

APP 6

An Australian Privacy Principle stating that personal data can only be disclosed for the specific purpose for which it was originally collected.

APP 11

An Australian Privacy Principle requiring organisations to take reasonable steps to protect personal data from misuse, interference, loss, or unauthorised disclosure.

4T's of Risk Management

A risk framework consisting of Tolerate (accept and monitor), Terminate (avoid by stopping activity), Treat (mitigate impact or probability), and Transfer (hand off to a third party).

Homomorphic Encryption

An encryption method that allows computations to be performed directly on encrypted data without decrypting it first.

Role-Based Access

A security control mechanism where users are assigned roles (e.g., admin, user) to restrict database permissions such as read, write, and insert.

Pseudoanonymisation

The processing of personal data such that it cannot be attributed to a specific data subject without separate, securely kept additional information.

Prosecuter Risk

The risk that individuals in an anonymised dataset can be re-identified by an adversary using pre-existing background knowledge about a target individual.

Journalist Risk

The risk that individuals can be re-identified by combining an anonymised dataset with external datasets and side channels (e.g., social media posts).

Data Synthesis

The generation of a synthetic dataset that replicates the properties, associations, and distributions of the original data without leaking real records.

Data Obfuscation

The process of adding noise to raw data to enhance privacy at the expense of analytical utility.

$k$-Anonymity

A privacy property ensuring that an individual cannot be distinguished from at least $k - 1$ other individuals in the dataset.

Homogeneity Attack

A vulnerability in $k$-anonymity where all individuals in an equivalence group share the exact same sensitive trait.

Background Knowledge Attack

An attack that combines external information with an anonymised dataset to discover sensitive information.

Unicity

A measure of how much information is required to uniquely re-identify an individual within a dataset.

Differential Privacy

A technique that adds random noise to query outputs so that an observer cannot tell whether any single individual's data was included in the dataset.

Epsilon ($\epsilon$)

The privacy parameter or privacy budget representing the maximum impact an individual's data can have on an output, where a smaller $\epsilon$ denotes stronger privacy, more noise, and less accuracy.

Mind Bugs

Common cognitive pitfalls that affect how data scientists interpret and present data, including patternicity, story telling, and confirmation bias.

Patternicity

A cognitive pitfall characterized by finding patterns where none exist.

Story Telling (Cognitive Bias)

A cognitive pitfall where a data scientist rationalises non-existent patterns.