Suppose that three subjects lets say A1, B1, and C1 belong to class 1 and A2 B2, and C2 belongs to class 2.
Then, feature extraction is applied on all A1 A2 B1 B2 C1 C2 to extract their features.
Now, I would like to know that should I model all features of class 1 together as a whole model or should I make a model for A1, make a model for B1, and make model for C1 and so on?
If a new subjects lets say D1 is missed in previous step, which approaches are better?
How about modelling time?
How about testing time?
Hi Mohammad,
In general, the model has to be formed from the whole features of all your instances. Precisely, in your case, one model has to be built from the features of your instances (A1, B1, and C1) to predict A2, B2, and C2, respectively. Then, the model that is generated from the features of A1, B1, and C1 is going to be applied on the new instance ( D1).
The performance of the model is the most important factor to worry about. What's the benefit of having fast prediction model with low performance?
HTH.
Samer, PhD.
Dear Samer thanks for your respond. I got your point. However, the problem is arising when you need to update your model again with new features. For this purpose, it needs to rebuilt previous features as well as new features then model it again.
Dear Mohammad,
In this case, you can perform an online learning, where the updateable algorithms can be applied effectively.
Cheers,
Samer
I guess you are working on classification models and you have two classes. So you need to model based on all samples from the two classes over all your extracted features to start with. You can now test your model on unknown data. There are various classification algorithms to try out. All the best.
I am agree with Adeyemi Adegbenjo
+++
it is important to involve the priority aspect of the feature
Well, are you doing a biometric based user authentication or user identification. In the case of the authentication, you know the user's identity and you are simply authenticating the user into a known, as you called it, class -- for this case you need a single classifier for each class. In the class of identification, user will present their credentials and the biometric passcode -- this is the case pointed out by Adeyami and your concern is not only checking the self-similarity of the biometric signatures, but also cross-similarity of the biometric signatures. All this ties into feature extraction and model selection... ..the rest of the answers given the above caveat are sound.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling