Which machine learning algorithm to use?!

I have a training set which is set of essays written by students for a question. These essays are all scored by human evaluators with labels such as 1, 2 , 3 which is actually marks allotted for those essays. I want know whether to use regression or classification algorithm for machine learning purposes! My readings on Machine learning suggests me to go with classification algorithm but again should I go with numeric classification or nominal classification. I am thinking numeric - Am I correct?

• Occam's Razor. Try something simple to code first like a regression/perceptron. If that doesn't work, try something more complicated like a logistic regression. If that doesn't work try an support vector machine. – Alex R. Oct 21 '13 at 15:21
• Thanks Alex. So the only way to decide on the best fit algorithm is trial and error? I am surprised. I thought there will be clear cut definitions on when to use a particular algorithm. – SUNIL KUMAR C Oct 21 '13 at 15:26
• It all depends on what you are trying to measure and the kind of error you are willing to allow for. Some data is clustered nicely allowing for a multi-class perception or a linear regression. Other times, it's really mixed together so you need to unravel it something akin to a support vector machine. For something like this, perhaps a mixture of gaussians would be appropriate if you have strong but spread out clustering for each label. – Alex R. Oct 21 '13 at 15:36

You would need to construct $A_{V\times D}$ term-document matrix where $V$ is the vocabulary size and $D$ is the number of documents. Thus, each column of $A$ is a document or a collection of word-counts (here we are assuming that words are exchangeable and only their counts matter). I recommend using standard text-preprocessing such as tokenization, stop-word removal and tf-idf smoothing. Gensim library implements many of these functionalities.
The input to your Lostigic Regression (LR) becomes an $(X,y)$ pair where $X=A^{T}$ (each row of $X$ is a document with $V$ features) and $y \in \{1,...,K\}$ is the label. Once trained, you can use LR to predict the label of a new document (pre-processed in the same way as the training data).