How to choose a function selection algorithm? - advice

Question

Is there a research article / book that I can read that can tell me about the problem, which function selection algorithm will work best.

I'm trying to simply identify twitter posts as pos / neg (for starters). I started by choosing a function based on frequency (starting with the NLTK book), but soon realized that different people chose different algorithms for a similar problem

Although I can try the frequency, mutual information, information, and various other algorithms, the list seems endless .. and wondered if there is an effective way, and then trial and error.

any advice

Answer 1

Have you tried the book that I recommended on your last question? It is freely available online and is completely about the task you are dealing with: "Mood Analysis and Intelligent Assessment" by Pang and Lee. Chapter 4 (“Extraction and Classification”) is exactly what you need!

Answer 2

I finished the last course of NLP, and it became clear that the analysis of moods is that no one really knows how to do it (for now). Of course, doing this through uncontrolled learning is even more difficult.

There are quite a lot of research on this subject, some of them are commercial and, therefore, are not open to the public. I cannot point you to any research work, but the book we used for the course was this ( Google Book Preview ). However, the book covers a lot of materials and may not be the fastest way to find a solution to this particular problem.

The only thing I can draw your attention to is try to use googling, perhaps on academar.google.com, for "mood analysis" or "mining."

Take a look at the NLTK movie_reviews . Grades are already in the pos / neg category and can help you learn your classifier. Although the language you find on Twitter is probably very different from what it is.

As a final note, please post any success (or failure, for that matter) here. At some point, this problem will arise for certain.

Answer 3

Unfortunately, there is no silver bullet for anything when working with machine learning. This is usually called a "No Free Lunch . " In principle, a number of algorithms work for a problem, and some are better at handling some problems and worse than others. In the end, they all do roughly the same thing. The same set of functions can lead to the fact that one algorithm will work better, and another - worse for a given data set. For another data set, the situation can be completely reversed.

Usually I select several function selection algorithms that worked for others on similar tasks, and then start with them. If the performance I get using my favorite classifiers is acceptable, scrounging another half percentage point is probably not worth my time. But if this is unacceptable, then it's time to re-evaluate my approach or look for additional methods for selecting objects.