I am classifying into two classes and data points in one class are almost 3 times more than data points in the second class. My question is that is fitcsvm sensitive to unbalanced datapoints? oe should I do some tricks to overcome this?
For example fitcsvm uses prior probabilities (give some probabilities to each datapoint based on their frequencies) in the algorithm and it might help to eliminate the effect of unbalanced data points on classification outcome.
There are some algorithms for generating under-sampling data points for example, SMOTE is an effective algorithm that is a KNN based method.
Have a look at
https://www.researchgate.net/post/How_to_solve_classification_tasks_with_unbalanced_data_sets_on_Matlab_Neural_Network_Toolbox
It may give you some clue about what you're looking for!
I think your question is very specific, which is good but in a bigger picture of training classifiers; this question could be answered by laying down some other facts.
Classifiers such as SVM, neural networks or random forest, etc. are sensitive, unbalanced data. You will face the problem of unbalanced data again and again, from training a classifier to validating the classifier's performance.
So instead of thinking of a classifier that handles the unbalanced data, you can make your data balanced. This solution is a common practice.
This is achieved by undersampling the larger class. To use all your data, you can run this subsampling many times, which also enables you to measure the confidence intervals of your classifier performance.
Also, think about using stratified testing-i.e., the cross-validation process generates the test sets with identical sample distribution as your training.
However, when your classification task handles classes with different levels of importance/sensitively, or if one class misclassification is very costly. (one example is when classifying a cancer patient as healthy has a more devastating consequence than classifying as a healthy subject as cancer)You can impose weight on the loss function.
Hello, this is an interesting question. I had a similar problem when I classifying my volcano-seismic data of 5 classes to classify unbalanced data, I advise 3 algorithms.
The first and the one I use the most is to balance the data to a fixed number of samples, for example, to the number of samples in the class that contains the least amount of data. Let's say we have 5 classes with a number of samples 800, 3000, 3200, 2800, 1900, so the training data could be set to 70% of 800 for each class, and leave the rest of each class as a test stage (Hold Out validation), in this sense, SVM is a great classifier, as it works with just a subsamples from the training database, called support vectors.
The second method is to create synthetic data of the different classes to the class that contains the largest number of samples, and match all the samples to the number of samples of this last class.
As a last algorithm, I would use some technique to give weights to the unbalanced data, for example, fuzzy logic, and after that use artificial intelligence models.
I hope these ideas help you, many successes.
You can use data balancing techniques such as over-sampling and under-sampling. in over-sampling ADASYN is the best and for under-sampling, you can use CC(centroid cluster)
You should use input argument ('Weights' — Observation weights numeric vector | name of variable in Tbl) of the fitcsvm. It is argument, directly used for your situation, unbalanced data set.
Dear Zeynab,
Details for Classification with Imbalanced Data can be found as follow:
https://www.mathworks.com/help/stats/classification-with-imbalanced-data.html
Dear Zeynab,
Details for Classification with Imbalanced Data can be found as follow:
https://www.mathworks.com/help/stats/classification-with-imbalanced-data.html
- Badges
- Science topic
- Similar topics
- Vectorization