I've read in some literature,that SVMs are characterized by their adaptivity. Does that mean they can learn while in use?


1 Answer 1


Support vector machines are a supervised machine learning method that can be used for both classification and regression problems. They are based on finding a Maximum Margin hyperplane that best separates the data into different classes or predicts the output value.

Support vector machines are not designed to learn online while in use. They are trained on a fixed set of data points and then used to classify new data points.

However, they can learn online, if they are implemented with online learning algorithms. Online learning algorithms can update the SVM model with new data that arrives in batches or streams, without retraining the model from scratch every time. This can improve the performance and efficiency of the SVM model in dynamic or large-scale applications.

There are some extensions of support vector machines that can enable online learning, such as incremental SVM, kernel recursive least squares, and budgeted SVM.

Incremental SVM helps in online learning by updating the SVM model with new data that arrives in batches or streams. It does not need to retrain the model from scratch every time new data comes in, which saves time and resources. It also can handle concept drift, which means that the data distribution may change over time. For example, if you are using SVM to classify spam emails, the features and patterns of spam emails may change over time as spammers use different techniques to avoid detection. Incremental SVM can adapt to these changes by adding or removing support vectors as needed.

Kernel recursive least squares helps the SVM to learn online by performing online regression with nonlinear kernel functions. It updates the model parameters and the kernel matrix with new data, using a recursive algorithm that is similar to the Kalman filter. It does not need to store all the data or solve a large optimization problem, which reduces the memory and computational requirements. It can also handle noisy data and outliers, by using a forgetting factor or a sparsity criterion. For example, if you are using SVM to predict the stock price of a company, the data may be noisy and contain outliers due to market fluctuations or external events. Kernel recursive least squares can filter out the noise and outliers and update the model with the most relevant data.

Budgeted SVM helps in online learning by performing online classification or regression with a fixed number of support vectors. It updates the model by adding new support vectors and pruning the least relevant ones when the budget is exceeded. It uses a strategy that balances the trade-off between accuracy and sparsity. It can also handle nonlinear kernel functions and different loss functions. For example, if you are using SVM to classify images of animals, the data may be large and diverse, and you may have a limited memory or computational capacity. Budgeted SVM can reduce the size of the model and the complexity of the calculations by keeping only the most informative support vectors and discarding the redundant or noisy ones.

Does "kernel trick" help with online learning?

The kernel trick does not directly help the SVM to learn online, but it helps the SVM to solve nonlinear problems by mapping the data into a higher dimensional space where it may be easier to find a linear hyperplane. The kernel trick also avoids the explicit computation of the mapping function and the coordinates of the data in the higher dimensional space, which can be costly and inefficient. Instead, the kernel trick uses a kernel function that can replace the inner product of the mapping function. This makes the SVM more flexible and scalable.

SVMs are characterized by their adaptivity, Does that mean they can learn while in use? $\rightarrow$ I think you may have misunderstood the meaning of adaptivity in the context of SVMs. Adaptivity does not mean that they can learn while in use, but rather that they can handle different types of data and problems, such as linear or nonlinear, classification or regression, and various kernel functions. SVMs are characterized by their adaptivity because they can find the optimal hyperplane that separates the data with the maximum margin, regardless of the shape or dimensionality of the data. However, once the SVM is trained, it does not change its parameters based on new inputs.


  • 2
    $\begingroup$ How exactly does the "kernel trick" help with online learning? $\endgroup$
    – NikoNyrh
    Nov 2, 2022 at 21:06
  • 1
    $\begingroup$ @Faizy imprecise. The ability to perform online learning is a property of SVM itself; the kernel trick has nothing to do with it. In fact, due to the computational expense of kernel, it is more difficult to learn online with a kernel than without. $\endgroup$
    – lpounng
    Nov 3, 2022 at 3:10
  • $\begingroup$ Check out this answer. $\endgroup$
    – lpounng
    Nov 3, 2022 at 3:10
  • 1
    $\begingroup$ The maximum margin property does not imply online trainability, as this answer claims. What matters is that the maximum margin property is maintained as new data arrives. This depends entirely on the employed optimization algorithm and the data generating process’ properties. (Aside: the kernel trick is only relevant if it affects the interaction of those things.) $\endgroup$
    – kdbanman
    Nov 9, 2022 at 14:11

You must log in to answer this question.