Hello, I just implemented a SOM algorithm in MATLAB that outputs component planes and U matrix....but i want to be able to calculate sensitivity, accuracy and specificity....how do i go about doing this in MATLAB??....any ideas or useful links would be highly appreciated??
You should mention that this question is related to another one.
You already get the formulas for sensitivity, accuracy and specificity. There should not be a problem to code them in MATLAB.
Remember these terms are for binary classification and you have to know true classes in your test set to estimate sensitivity, accuracy and specificity for each class.
Let's say, in your dataset you have classes A, B and C. After SOM classification you get also 3 classes, but not all samples were classified correctly. Then you can build 2x2 confusion matrix for each class and estimate sensitivity, specificity and accuracy.
Please read the responses provided for your last question by BioStar members. You should have a dataset, which contains both positive and negative data class labels. You can divide the dataset as I explained in the response to your last question. You don't need a tool box for this. Once you generate your SOM, use this datasets to get the performance assessments.
Hey yuk.....my problem is.....i'm using the SOM toolbox and it does not perform classification....its using the unsupervised methodology....which I have already implemented....but how do I get the SOM toolbox to divide a dataset into training and test sets....it is not implemented with the SOM toolbox....nd i'm having a hard time doing dat.....do u know if the Neural Network Toolbox does classification with SOM???
good point! I have a question here. We all know that the SOM maps high-D data to 2D (or 3D), and this mapping is not bi-directional. For example, SOM has a 20x20 grid, and I send in one 1x10 vector into SOM each time step for 1000 steps. After learning, the neurons are clustered as 3 classes which is expected as my data is composed of 3 classes. BUT, how do we define the success rate here? Because we only know which neuron is classified into which class, we do not know where each data vector is classified!
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
It would have been much better if you had rephrased your first question in the light of the answers you already got and responding to the comments you got. Something like: "Methods to turn a SOM into a supervised classifier/predictor" Because that is the first step and requires a bit more than just implementing functions to calculate these values. I assume that in the end you just want a good classifier. Why exactly use a SOM? Even with lots of tuning and reasoning to turn the SOM into a supervised classifier it will still be inferior to any out-of-the-box support vector machine (SVM).