There is no unique answer for this. A threshold-based classifier may work well in many applications, but it may be the case in which a more complicated system (maybe a neural network?) will perform better. It depends on the problem you are dealing with. Also, remember that "real time" requirements change for different time scales and hardwares.

There is no unique answer for this. A threshold-based classifier may work well in many applications, but it may be the case in which a more complicated system (maybe a neural network?) will perform better. It depends on the problem you are dealing with. Also, remember that "real time" requirements change for different time scales and hardwares.

please cite some reference in support of your answers

I don't think there is just an answer to your question. But Cascade Classifiers showed promising results in the last decade. For example Viola-Jones approach for object detection uses cascade classifier.

http://en.wikipedia.org/wiki/Viola%E2%80%93Jones_object_detection_framework

SVM can be really fast too. Of course it's speed depends on the number of support vectors.

http://dl.acm.org/citation.cfm?id=958479

http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=4624313&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D4624313

I thank you all for your answers but I didn't get the acceptable answer till now, the question is still open for the acceptable answer

Extreme Learning Machine: https://www.researchgate.net/publication/262334663_Extreme_learning_machine_a_robust_modeling_technique_yes%21?ev=prf_pub

Conference Paper Extreme Learning Machine: A Robust Modeling Technique? Yes!

The classifiers depend on the measurement scales of the data, depend on the structure of the problem (the main objective and the appropriate subobjectives), depend on the presence or not of the uncertainty (stochastic noise) of the initial data and so on. There is not unique and definitive answer.

Dear Arslan Ahmad,

you are absolutely correct.

A: Your question is still open

which translates in this very case to:

the problem is between not solved and not solvable.

B:  You did not got an "acceptable" answer,

because there is no unique answer.

Unfortunately your question is too general to be answered by a simple, easy to describe way. 

What do you mean by "real time applications?"

hard real-time, with a guaranteed response time?

soft-real time, with a response time "fast enough"?

What order of magnitude are we talking?

seconds, minutes, milliseconds, micro seconds?

Or are you seeking for something with a guaranteed response time in nano-second scale?

Or do you consider the steering of an oil-tanker ? Where real-time is more on the hours time scale?

With a a more specific question it would be possible to get an answer mor e to the point.

Now back to your question:

Talking "real-time" you mean probably something like soft-real time, beeing fast enough for your application.

A very good example is (mentioned above by Mohsen Biglari) the alogorithm of Viola and Jones for pattern, or especially for face detection.

The smart pre-processing of the picture (here filtering and the idea of integral images) is a second key to performance.

A cascade of classifiers, reducing the number of false positives is one of the successful paradigms today in machine learning and classification.

The sequence of classifiers within the cascade is build in such a way, that a lot of false positives are sorted out already in the first stage, an large amount of the remaining is sorted out by the second stage, and so on.

This approach makes the algorithm be quick in finding false positives, but takes longer for the pattern you are looking for.  Thus, it will not have a pre-definable response time, although it might be fast enough, for a "soft real-time" application, it won't meet the "hard real-time"  requirements.

A lot of other successful "real-time" applications show the same paradigm:

They are only successful because several application dependent aspects are combined in a smart way.

I hope the example was helpful for you.

What exact real-time application did you have in mind?

I am sure with more information on that there will be a lot of persons in the community willing to help.

regards

Nils Goerke

I agree with the above. Also it depends to a great deal on how you encode your input  features (can speed up the classifier) and the computational complexity of the coding process (can slow down overall classification process). 

Decision trees can be coded in Assembly language very easily allowing you to make a classification in around 2 microseconds for a depth of 10 nodes. 

My interest in this field is in rapid classification of dialogue acts required for human dialogue with web-based agents, which naturally have issues of scalability. You can see some papers on this at www.semanticsimilarity.net - look for dialogue act classification using function word features.

Good luck with your work, this is an interesting area.

I would highly recommned Random Forests, given the great success of Kinect, which is a real application. I am working in this field and they are working well in practice. 

As written above I strongly underline that there exists not a unique answer! First you have to define clearly what do you mean with the wording "Real-time application". Usually we call it "Process-Real-time" because the process determines the time range which is interesting for your application. Real-time does mean that data are processed before new data are coming in. In this sense you have many control parameter which you can teak. Many of them have only a little to do with classification itself, like pipelining, parallel processing (if possible), etc. Furthermore, you are able to optimize you classifier design hardware-wise by applying graph-theoretical aspects like Systolisation oder fast mathematical modeling. Al these above mentioned methodologies (if applicable) generate a faster classification, independent from the classifier itself!

Having said this, it is of course possible to apply classifier approaches which guaranty fast computing, but this is secondary. Why? You are mixing two approaches i) the classification and ii) the process-real-time concepts.

From my experience the fist step should be to design a "good" classifier based on your constraints or the application constraints and after that you make it fast by appropriate methodologies.

I think that LDA and QDA classifiers are the best methods to classify data into different classes for real time application. I used them for classifying data into five classes.

I would like to add another aspect which directly relates to the answer from Sebastian Raschka. Real-time applications correlate with real-world scenarios, that is, usually you are confronted with a low amount of available training data. Therefore, it is necessary to apply a classifier concept which is able to handle this aspect also. Please refer to https://www.researchgate.net/publication/221453394_Fuzzy-Pattern-Classifier_Training_with_Small_Data_Sets. There we have discussed this topic.

Conference Paper Fuzzy-Pattern-Classifier Training with Small Data Sets

It is well established in the literature and according to the ‘no free lunch’ theorem, there is no optimal classifier that works perfectly for every class of problems, as the performance of the classifier depends mainly on the problem and the data used.

I just have published a related article: 

https://www.researchgate.net/publication/265168466_Solving_the_Problem_of_the_K_Parameter_in_the_KNN_Classifier_Using_an_Ensemble_Learning_Approach?ev=prf_pub

Article Solving the Problem of the K Parameter in the KNN Classifier...

interesting!