Online machine learning is a method of machine learning where the model continuously updates and evolves as it is exposed to new data. Unlike traditional machine learning methods that require retraining on the entire dataset, online machine learning models adapt and learn from each new data point they receive. This makes them particularly useful in situations where data is continuously generated and the model needs to adapt to changing patterns in real-time.
Online machine learning offers several practical advantages, particularly for data that arrives continuously or changes over time. Firstly, it can adapt to concept drift — shifts in the underlying data distribution — by updating incrementally rather than waiting for a full retrain. Secondly, it is memory-efficient, since it processes one example (or a small batch) at a time instead of holding the entire dataset in memory, which makes it well-suited to streaming or resource-constrained settings. Lastly, it can start making predictions immediately and keep improving as more data arrives, rather than requiring a complete dataset up front. Whether an online model is more accurate than an offline one trained on the same data depends on the problem; offline batch training with multiple passes over the data will often reach a lower error on a fixed, unchanging distribution.
True online learning algorithms are designed to update their parameters incrementally from a stream of examples, rather than being retrained on a full batch. Commonly used examples include:
Stochastic Gradient Descent (SGD) — Updates model weights using one example (or a small mini-batch) at a time, and underlies online variants of linear and logistic regression.
Perceptron — A simple linear classifier that adjusts its weights after each misclassified example, making it naturally incremental.
Passive-Aggressive Algorithms — A family of online learners for classification and regression that make a minimal update on each new example: no change if the current model already predicts it correctly, and a larger correction if it doesn't.
Online/Incremental Naive Bayes — Updates class and feature probability estimates as each new example arrives, without recomputing statistics over the full dataset.
Hoeffding Trees (Very Fast Decision Trees) — Decision trees built incrementally from a data stream, using statistical bounds to decide when there is enough evidence to split a node.
Note that batch algorithms such as standard support vector machines, random forests, and most deep neural network training are typically trained offline on a fixed dataset; some can be adapted to incremental or mini-batch updates, but that is an adaptation rather than their default mode.
Choosing an online learning algorithm involves considering several factors:
Type of Data — Linear methods like SGD-based regression or the Perceptron work well when the relationship between inputs and outputs is roughly linear. Hoeffding Trees can capture more complex, non-linear patterns while still updating incrementally.
Update Speed — If predictions and updates need to happen with minimal latency per example, simpler models like the Perceptron or Passive-Aggressive algorithms are typically cheaper to update than tree-based methods.
Available Resources — Streaming settings often have tight memory and compute budgets. Algorithms that maintain a compact, fixed-size state (such as SGD-based linear models) are generally easier to run continuously than approaches that grow their state over time (such as trees).
As with any modeling choice, these are general guidelines. Evaluating a small set of candidate algorithms on your actual data stream is the most reliable way to pick one.
Evaluating the performance of an online machine learning algorithm involves several key metrics.
Accuracy — This is measured by comparing the predicted values to the actual values in the data set. The closer the predicted values are to the actual values, the higher the accuracy of the algorithm.
Convergence Speed — This refers to the rate at which the algorithm learns and makes accurate predictions. A faster convergence speed indicates a more efficient algorithm.
Scalability — This is the algorithm's ability to handle larger data sets and more complex problems without a significant decrease in performance. An algorithm with high scalability is more versatile and robust.
These metrics provide a comprehensive evaluation of the algorithm's performance, helping you choose the most suitable one for your specific needs.
More terms
Glossary term
Glossary term
Collaborate with your team on reliable Generative AI features.
Want expert guidance? Book a 1:1 onboarding session from your dashboard.