I use PSO to find the best feature subset for classification. Population is analyzed every iteration in PSO in order to achieve the particle (best combination of features) that has the least cost (best fitness).
Every Particle is a vector with the length equal to the number of features. For example: We assumed that particle [1 0 1 1 0 1] is the best solution. This vector demonstrates that there are 6 features. F1, F3, F4, and F6 will be used and F2 and F5 will be ignored in the classification step.
How can I calculate the cost function of this population when I don't know the result of classification in advance?
Is your feature selection algorithm wrapper based or filter based?
Assuming that it's filter based, you need to use a certain measurement to evaluate the quality or goodness of the feature subset. This measurement is the cost function. It can be a function that measures the correlation of features. It can be the case that a higher correlation value indicates a poor combination of features. In this case, the goal is the minimise the correlation value in order for you to get a good subset of features.
Assuming that it's wrapper based, PSO will choose the features for you. For every set of feature subset that is proposed, the classifier will perform the classification tasks (k-fold cross validation can be used) and the goal now is the maximise the classification accuracy. In this case, the cost function used is the classification accuracy of the classifier used.
Is your feature selection algorithm wrapper based or filter based?
Assuming that it's filter based, you need to use a certain measurement to evaluate the quality or goodness of the feature subset. This measurement is the cost function. It can be a function that measures the correlation of features. It can be the case that a higher correlation value indicates a poor combination of features. In this case, the goal is the minimise the correlation value in order for you to get a good subset of features.
Assuming that it's wrapper based, PSO will choose the features for you. For every set of feature subset that is proposed, the classifier will perform the classification tasks (k-fold cross validation can be used) and the goal now is the maximise the classification accuracy. In this case, the cost function used is the classification accuracy of the classifier used.
Kee Huong Lee
Thank you. Your answer really helped me. Actually, the cost function was SVM classifier that I hadn't noticed.
If you want to use wrapper-based or embedded-based (based on feature subset evaluation methods), a two objective function based on TPR and FPR (that will be achieved from SVM model by creating and evaluating it on the each feature subset) is a good candidate for fitness function. For more details refers to the following paper:
Article Hybrid of binary gravitational search algorithm and mutual i...
Actually, fitness function is determined depend on what is your problem ?
I agree with Aguilar-Ruiz response. in case of clustering , you must test different methods as fitness function, and at the end of these examinations, you should selet the best function as your fitness func. so you notice that there are a big different between "fitness' & "cost" .
Because we use from "cost" as fitness only when we wanna reach to a solution with minimize cost, while in some cases, you have to find a solution with maximized fitness value. depend on your defined fitness function , you can have "fitness" or "cost" as respectively namely by "maximization" and "minimization " sub-problem.
Best.
You can use any of the following:
https://brenocon.com/blog/2012/03/cosine-similarity-pearson-correlation-and-ols-coefficients/
They are good to calculate the "cost" in a filter based system.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining