Glossary¶

Use this page to record terms and ideas that help you understand professional analytics projects.

This project covers applied machine learning: calling deployed models, investigating their behavior, and working with text and image data.

Pro-tip: Expand the VS Code Outline view (below the navigator on the right) to see this file organization at-a-glance.

Deployed Models and APIs¶

API (Application Programming Interface)¶

An API is a defined way for one program to send requests to another. A prediction API accepts input data and returns a model's prediction as a response.

endpoint¶

An endpoint is a specific URL that an API exposes for a particular action. The /predict endpoint in this project receives feature values and returns a predicted species.

payload¶

A payload is the data sent in a request to an API. In this project, the payload is a JSON object containing the feature values for one penguin.

HTTP POST request¶

An HTTP POST request sends data to a server and receives a response. The requests.post() function in Python sends a payload to an API endpoint and returns the server's response, including the prediction.

cold start¶

A cold start happens when a server that has been sleeping wakes up to handle a request. Free hosting tiers often sleep after inactivity, causing the first request to be slow.

edge case¶

An edge case is an input that is unusual, extreme, or outside the expected range. Testing edge cases reveals whether an API handles bad input gracefully or crashes.

feature sensitivity¶

Feature sensitivity describes how much a model's prediction changes when one input feature is varied while others are held fixed. A feature with high sensitivity has strong influence over the model's output.

Text as Data¶

corpus¶

A corpus is a collection of text documents used for analysis or training. Each document in a corpus may be a sentence, paragraph, article, or other text unit.

TF-IDF (Term Frequency-Inverse Document Frequency)¶

TF-IDF is a method for representing text as numbers. Each word in a document gets a weight that is high when the word appears frequently in that document but rarely across the full corpus. TF-IDF captures which words appear but discards word order and meaning.

vocabulary¶

The vocabulary is the set of unique words the TF-IDF vectorizer learns from the training corpus. New documents are represented using only the words already in the vocabulary.

word order¶

Word order is the sequence in which words appear in a sentence. TF-IDF ignores word order, so "dog bites man" and "man bites dog" produce identical vectors. This is one of TF-IDF's known limitations.

n-gram¶

An n-gram is a sequence of n consecutive words treated as a single feature. Using bigrams (n=2) gives TF-IDF some ability to capture word pairs, partially recovering word order information.

Images as Data¶

pixel¶

A pixel is one point of color or brightness in an image. A grayscale 8x8 image contains 64 pixels, each with a value from 0 (black) to 16 (white).

flattening¶

Flattening converts a 2D image grid into a 1D array of pixel values. An 8x8 image becomes a vector of 64 numbers that a classifier can use as features. Flattening discards spatial structure - which pixels are neighbors - which limits a model's ability to handle rotated or shifted images.

spatial structure¶

Spatial structure refers to the arrangement and relationships between pixels in an image. Flattening an image into a feature vector removes spatial structure. More advanced representations such as convolutional neural networks preserve it.

confusion matrix¶

A confusion matrix shows how often a classifier predicted each class compared to the true class. Off-diagonal entries are misclassifications. For digit images, a confusion matrix reveals which digit pairs the model confuses most often.