It depends on which LDA you mean. The abbreviation “LDA” refers to two entirely different algorithms in machine learning, and they fall on opposite sides of this question. Linear Discriminant Analysis is a supervised method that requires labeled data. Latent Dirichlet Allocation is an unsupervised method that works without any labels at all.
Linear Discriminant Analysis Is Supervised
Linear Discriminant Analysis (the LDA most common in statistics and classification tasks) is a supervised learning algorithm. It needs labeled training data to work. You give it examples that are already tagged with their correct category, and it learns a rule for separating those categories as cleanly as possible.
The core math behind it aims to maximize a ratio: the variance between classes divided by the variance within classes. In plain terms, the algorithm looks for a way to project your data so that points in the same group cluster tightly together while the groups themselves spread far apart. If you have two categories, it finds a single line that best separates them. With more categories, it finds multiple axes of separation.
This is why class labels are essential. Without knowing which data points belong to which group, the algorithm has no way to calculate within-group or between-group spread. It’s solving a fundamentally different problem than an unsupervised technique like Principal Component Analysis (PCA), which ignores labels entirely and simply finds the directions of maximum overall variance in your data.
Key Assumptions
Linear Discriminant Analysis works best when two conditions hold: the data within each class follows a roughly normal (bell-curve) distribution, and the spread of data in each class is similar. When the classes have very different spreads (unequal covariance matrices), performance drops, and Quadratic Discriminant Analysis often works better. Research comparing the two methods found that LDA is the stronger choice specifically when the covariance structures across classes don’t vary much.
Common Applications
Linear Discriminant Analysis shows up wherever you need to classify labeled data into discrete groups, especially when you also want to reduce the number of features you’re working with. It’s used in activity recognition (distinguishing sitting from standing using sensor data), facial recognition, medical diagnosis, and financial classification like credit scoring. Because it doubles as a dimensionality reduction tool, it’s often a preprocessing step before feeding data into another classifier.
Latent Dirichlet Allocation Is Unsupervised
Latent Dirichlet Allocation (the LDA common in natural language processing) is an unsupervised algorithm. It takes a collection of documents and discovers hidden topics within them, with no labels or predefined categories required. The only observed data is the words in the documents themselves.
The model assumes each document is a mixture of topics, and each topic is a mixture of words. It uses a probability distribution called the Dirichlet distribution to model both of these mixtures. A parameter called alpha controls how many topics each document tends to cover: a low alpha means documents focus on just one or two topics, while a higher alpha produces documents that blend many topics together. A similar parameter called eta controls how focused each topic is on a small set of words versus a broad vocabulary.
Because it works without labels, Latent Dirichlet Allocation is a density estimation technique. It tries to learn the underlying structure of the text so that its model assigns high probability to documents it hasn’t seen before. The topics it discovers are not categories you define in advance. They emerge from patterns of word co-occurrence across the entire collection.
Common Applications
Latent Dirichlet Allocation is widely used in text mining and information retrieval. Researchers have applied it to analyze open-ended survey responses, identifying recurring themes like emotional well-being, stress, and social isolation across thousands of text entries. It’s also used to organize large document collections (think tagging thousands of research papers by topic), power recommendation systems, and summarize the themes in customer reviews or social media posts. Once the topics are discovered, they can be used as features for downstream classification, but the topic discovery step itself is fully unsupervised.
How to Tell Which LDA Someone Means
Context usually makes it obvious. If the discussion involves classification, labeled data, or dimensionality reduction for structured datasets, it’s Linear Discriminant Analysis (supervised). If the discussion involves text, documents, topics, or NLP, it’s Latent Dirichlet Allocation (unsupervised).
A quick reference:
- Linear Discriminant Analysis: Supervised. Requires class labels. Used for classification and dimensionality reduction on structured data. Rooted in Fisher’s criterion for maximizing class separation.
- Latent Dirichlet Allocation: Unsupervised. No labels needed. Used for topic modeling on text data. Rooted in Bayesian probability and the Dirichlet distribution.
The two algorithms share nothing beyond the abbreviation. They operate on different types of data, use completely different mathematics, and answer fundamentally different questions. One asks “which known group does this belong to?” The other asks “what hidden themes run through this collection?”