I am using RM for the sentiment analysis of the movie review dataset. i have tokenised the sentiments and have calculated the term frequency and TF-IDF for the words. for classification want to use 10-fold cross validated SVM-PSO but after 11th execution the tool returns the error "incompatible number of attributes (821! = 4260) !".
Solved! Go to Solution.
the problem is that the dataset you apply it on has not the same (or not enough) attributes. Are you sure that you passed over the wordlist from one process documents to the other?
You might have a look at this Knowledgebase post i just created for this: http://community.rapidminer.com/t5/RapidMiner-Studio-Knowledge-Base/Text-Mining-and-the-Word-List/ta...
Thanks for your reply.
sorry for replying late. Does the execution of PSO takes a long time ?? And in case of n-fold cross-validation, it is defined as (n-1) iterations are for training and nth iteration is for testing. In RM, the validation occurs for (n+1) times. Why is it so??
Let me refer you this discussion on X-validation: http://community.rapidminer.com/t5/RapidMiner-Studio/What-about-n-models-generated-in-cross-validati...
There is another link in solution that has an even more detailed thread about X-validation.
i have come to end of my work. after preprocessing, transforming tokens into TF and TF-IDF and classifying the generated unigram, bigram and trigram terms using SVM and SVM-PSO on five kernel types - dot, radial, polynomial, epachenikov and anova, i have found that for bigram and trigrams on all kernel types it gives same accuracy.
why is it so???
i would argue that the statistical significance of bigrams and trigrams are so small, that they are negleted by the SVM. Simply not enough occuruces of the bi/tri grams.
this is I think a difficult question and belongs definitly into the area of computer science. There is a overview paper by Kaestner etl al where also a linear kernel works quite well. I read some time ago, that the RBF kernel is worse because it introduces smearing which is bad in text mining - i forgot to save the bookmark.
If i remember correctly @land told me once that kernels are not good because they add additional degrees of freedom, which result usually in overfitting because d>>n.
thank you for ur reply.
Sir it is said and observed that SVM-PSO performs better than SVM reason being that PSO optimises and overcomes the drawbacks of SVM. what are these drawbacks of SVM overcomed by the PSO? What parameter are optimised by PSO in SVM ??? What actually makes SVM-PSO better than SVM??? i have implemented SVM and SVM-PSO for analysing the sentiments.
Eagerly waiting for ur early reply.