I am working with a Support Vector Machine to predict class prevalence in a binary classification problem.
The model will take a sparse representation of an instance as input, where the number of the feature is mapped to the feature value. Only the feature numbers with a non-zero value are included.
I am planning to use the model for text classification of short documents, and therefore aim to use a BoW representation of documents. The values for my features will be integers representing the frequency of a word occurring in the document. I am expecting that most BoW frequency features will fall in the interval
[1, 3]. Besides the BoW features, I want to extend the input vector with other numerical features. These numerical features will likely be much larger than the BoW frequency values.
My question is, whether it would still be a good idea to scale the data with for example normalization in the case of combining BoW features with other numerical features. I have looked into literature and previous questions but have not managed to find any good information from experience.